MEG - An Introduction To Methods
MEG - An Introduction To Methods
MEG - An Introduction To Methods
Edited by
Peter C. Hansen
Morten L. Kringelbach
Riitta Salmelin
1
2010
1
Oxford University Press, Inc., publishes works that further
Oxford University’s objective of excellence
in research, scholarship, and education.
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
www.oup.com
MEG : an introduction to methods / [edited by] Peter C. Hansen, Morten L. Kringelbach, and Riitta Salmelin.
p. ; cm.
Includes bibliographical references and index.
ISBN 978-0-19-530723-8
1. Magnetoencephalography. I. Hansen, Peter C., 1963- II. Kringelbach, Morten L. III. Salmelin, Riitta.
[DNLM: 1. Magnetoencephalography—methods. 2. Nervous System Diseases—diagnosis.
WL 141 M496 2010]
RC386.6.M36M44 2010
616.8’047547—dc22
2010002216
9 8 7 6 5 4 3 2 1
Introduction vii
Contributors xi
SECTION 1 BASICS
v
vi Contents
SECTION 3 APPLICATIONS
to be aware of the different types of detection coils that can be used for record-
ing the MEG data, as this influences the appearance of the signals at the sensor
level. If conclusions are drawn solely based on the sensor signals, without
proceeding to some form of source reconstruction, it is crucial to know the
structure of the detector coils. Chapter 2 (Parkkonen) describes the principles
of MEG hardware, and Chapter 3 (Parkkonen & Salmelin) provides guide-
lines for performing successful measurements. These chapters also discuss the
important steps involved in preprocessing MEG data, including evaluation of
possible artifacts and the effects of filtering. Chapter 4 (Salmelin & Parkkonen)
focuses on the design of MEG experiments and considers the selection of
parameters also with respect to that in behavioral and fMRI neuroimaging
studies.
The unique power of MEG is that it combines the localization of active
brain areas with reasonable spatial accuracy, together with the extraction of
the time courses of activation in those areas with excellent temporal accuracy.
Source analysis and reconstruction is now usually an integral part of the anal-
ysis of MEG data. Indeed, decomposition of the MEG sensor signals into
activity of specific neural sources enhances and clarifies stimulus and task
effects. Sources of early cortical somatosensory responses, for example, may
be localized very precisely because of the lack of simultaneous activation from
elsewhere in the cortex. Relative locations of sources within subjects can also
be determined fairly accurately, such as the distance between areas in the
somatosensory cortex that are activated by electrical stimulation of the thumb
versus little finger. It is important to keep in mind that MEG (or EEG) data
allow estimation of the centre of an active brain area but do not provide
detailed information about its shape. The precise appearance of the resulting
activity map is determined by the specific analysis method employed.
Chapter 5 (Baillet) outlines the general concepts and assumptions for the
various approaches that are currently used in MEG source analysis. Many of
those methods are applicable to EEG data as well or were initially developed
in that domain. Because of the inverse problem, some constraints are needed
in order to proceed from the distribution of magnetic field to configuration
of neural sources. A widely used and robust approach is to represent the neu-
ral sources by a physiologically plausible and mathematically simple model,
an equivalent current dipole. Chapter 6 (Salmelin) discusses the practical
application of this method to analysis of both evoked responses and oscilla-
tory background activity. Chapter 7 (Jensen & Hesse) focuses on another
approach in which constraints are set to the solution of the inverse problem
by accepting the result that accounts for the measured signals with the small-
est overall amount of electric current. Chapter 8 (Hämäläinen et al.) illus-
trates a conceptually similar approach, but with minimization of the overall
power, in which the activity is additionally constrained to the cortex based on
individual structural magnetic resonance images (MRIs) and the process may
be further guided by functional MRI (fMRI) activation maps. Chapter 9
(Gross et al.) details the use of spatial filters (‘beamformers’) in the detection
of interconnected neural networks. Chapter 10 (Pantazis & Leahy) outlines
x Introduction
xi
xii Contributors
1
2 MEG: An Introduction to Methods
EPSP IPSP
+
−
−
−
+
+
Closed fields
B
C
+
−
Open
D and mixed fields
Figure 1–2. Examples of closed (A, B), open (C) and open–closed (D) fields
according to Lorente de Nó. The isopotential lines resulting from the acti-
vation of the neuronal population are shown on the right side.
(Adapted from Niedermeyer & Lopes da Silva, 2005).
dendrites that are perpendicular to the overlying skull, i.e., those that are at
the top of a gyrus; whereas others are parallel to the skull, i.e., those that are
on the wall of a sulcus. The point to note in this respect is that the orientation
of the neurons with respect to the skull influences the resulting MEG signal
recorded outside the skull. In fact the MEG “sees” only those magnetic fields
that have a component perpendicular to the skull. These magnetic fields are
generated by neuronal currents that have a component oriented tangentially
to the skull. In contrast, those currents that are oriented radially to the skull
do not generate a magnetic field outside the head (Figure 1–3).
propagation
Direction of
Qd
B B
Qr
I
VI
A B B
Qr Qd
Direction of
propagation
0
0
that the total current density field J t consists of J i and a passive part (second-
ary current) that obeys Ohm’s law:
Jt = Ji + s E (1–1)
where σ is the medium conductivity and E is the electric field.
Considering that the div J t must be zero, since there is no accumulation
of charge anywhere in the medium, we may note that
div J I = −s div E (1–2)
Electrophysiological Basis of MEG Signals 7
()
V r0 = (−1/4ps )∫ div J i (1/ R ) d 3 r (1–4)
where R represents the distance between the measuring point r0 and the
source location r .
()
With respect to the magnetic field H r we should note that it is related
()
to the magnetic flux density field, or induction field, B r by the following
expression:
()
B r = mH r () (1–5)
∇ × H = Jt (1–6)
div B = 0 (1–7)
and
∇ × B = m( J i + s E ) (1–8)
()
∇×E r =0 (1–9)
B = ∇× A (1–10)
8 MEG: An Introduction to Methods
Magnetoencephalography versus
Electroencephalography
MEG EEG
Field patterns
Figure 1–4. Field pattern of MEG, on the left, and EEG, on the right, caused
by a current dipole model source in a concentric 4-layer spherical model
of a head. The shaded areas indicate the magnetic flux out of the head
(MEG) and the positive potential (EEG).
(Adapted from Hari, in Niedermeyer & Lopes da Silva, 2005).
Electrophysiological Basis of MEG Signals 9
are dipolar but are rotated by 90° with respect to each other. The isocontour
lines are tighter in the case of the MEG compared with the EEG. This is mainly
due to the influence of the inhomogeneities that smear out the electric potential
much more than the magnetic field.
s.o.
s.p.
~2 mm high
s.r.
A - Intracellular current
CA3
directed from base to
y apex
CA 1
CA 3
x
SQ-2 SQ-4
1 mm
C Magnetic field D
detecting coils to SQUID
SQ-3
Rec elec Stim elec B - Intracellular current
CA3 slice
directed from apex to
SQ-2 +
− SQ-4
base
CA3 slice
Cell layer
Stim. elec. SQ-1 SQ-2 SQ-4
Ringer in Rec. elec.
SQ-2
Data
SQ-4 +
8 pt
a 50 ms
b
SQ-2
Model
SQ-4
Figure 1–6. Magnetic fields recorded from the CA3 hippocampal slice as
indicated in Figure 1–5, and results obtained with a computer model of
a population of pyramidal cells. The magnetic fields recorded from SQ2
and SQ4 are shown for two conditions: on the left, after stimulation at
the level of the soma; on the right, after stimulation at the level of the
apical dendrites. Positivity (deflection upwards) indicates magnetic field
directed out of the bath. Note the general similarity of the waveforms of
the experimental and the theoretical results. Note also the difference in
polarity between early and late components.
(Adapted from Murakami et al. 2002).
The simulations showed, among other effects, that the magnetic field corre-
sponding to the evoked initial spike was directed into the slice at SQ-2 and
out of the slice at SQ-4 (Figure 1–5), which indicates that the intracellular
current is directed from the basal to the apical dendrites of the pyramidal
neurons. The slow wave that follows the spike shows the opposite polarity,
indicating that it is generated by an intracellular current in the reversed direc-
tion (Figure 1–6). This model allows one to analyze different patterns of den-
dritic branching, and how the distribution of the density of ion channels
along the soma-dendritic membrane may be reflected in the magnetic fields.
Interestingly, this experimental and modeling study showed that sodium
spikes can be detected magnetically and electrically in this preparation.
Murakami et al. (2002) suggest even that sodium spikes would be also detect-
able in EEG and MEG recorded from the scalp, “if the geometry of the distri-
bution of neurons is favorable”. However, this condition is most likely not
sufficient, because a high degree of synchrony is also necessary for spikes of
this kind to be recordable at the scalp.
12 MEG: An Introduction to Methods
order of 0.29 – 0.90 pAm. Apparently this value deviates appreciably from
that estimated previously for a cell by Hämäläinen et al. (1993), which was
just 0.02pAm. That larger value, however, is of the same order of magnitude
as values estimated for hippocampal pyramidal neurons (0.2 pAm per cell in
CA1 and 0.17 pAm per cell in CA3; see Kyuhou and Okada, 1993 and Okada
et al., 1997). Murakami et al. (2006) point out that assuming a Q of 0.2 pAm
per cortical pyramidal neuron, a population of 50,000 synchronously active
cells would generate a field with a magnitude of 10nAm, which corresponds
precisely to the value measurable from the human cortex according to
Hämäläinen et al. (1993). Assuming that a cortical minicolumn with a diam-
eter of 40μm contains 100 cells (Mountcastle, 1997), the cortical surface that
would correspond to 50.000 cells should form a patch of about 0.63 mm2 in
area. If this cortical patch had a circular form, its diameter would be about
0.9 mm (for more details see FAQ, Q5).
In order to translate a cortical surface to a number of cortical cells, in
which the synchronous activity is responsible for the signal measured, one
must have a good estimate of the density of pyramidal cells in the cortex.
These kinds of estimates, however, are rather imprecise, and those based on
the quantitative computational models described above are to be preferred.
We have already discussed the possibility of spikes contributing to the
magnetic field measured at a distance. This computer model offers more
insight into this issue. According to the calculations performed in the model,
a sodium spike corresponds to a Q with the magnitude of 1pAm. Therefore,
to reach the measurable value of 10nAm, it would be sufficient that 10,000
neurons generate such spikes synchronously. Although this is not impossible,
as these authors note, it is unlikely to occur outside situations of forced
synchronization, since there is considerable jitter in the ongoing cortical neu-
ronal activity (de Munck et al., 1992). This model study revealed, in addition,
two particularly interesting and unexpected results: first, the activity of basal
dendrites may contribute significantly to the MEG and EEG signals; second,
the magnitude of Q was unexpectedly large for a spiny cell, since it was esti-
mated to be 0.27pAm, whereas it was only 0.06 for an aspiny cell. Nevertheless,
we cannot conclude that spiny cells will make an important contribution to
MEG signals, since the dendrites of these neurons are not oriented parallel
to each other (as are those of pyramidal neurons) but have quite variable
orientations.
microscopic level the organization of the tissue is far from homogeneous. The
extracellular space appears to be constituted by a complex mesh that has been
called by Nicholson (1980) the “extracellular jungle.” Furthermore, extracel-
lular shifts of ionic concentrations, the role of glial cells and gap junctions, all
must be taken into account. Nonetheless, to model the electric and magnetic
phenomena as measured from the scalp, a macroscopic description is, in gen-
eral, adequate. It is not within the scope of this chapter to deal with the math-
ematics of how electric and magnetic fields are influenced by the boundaries
between media with different conductivities. We will consider only whether,
and if so, how, the magnetic phenomena are affected by the existence of layers
with different conductivities.
It is sometimes argued that these layers affect the EEG but not the MEG
signal; in other words, that these inhomogeneous media would be transpar-
ent to MEG signals. This is, however, not entirely accurate. In general,
the magnetic induction field B only depends on the rotation of the current
field, as shown above, and that the passive current field s E , being conserva-
tive, has no influence on the vector magnetic potential A . In a volume
conductor with varying conductivity, however, the passive current field can
also have rotational components at the boundaries of media with different
conductivities. A rigorous mathematical treatment of this matter is given by
van Rotterdam (1987).
oriented toward the cortical surface, and the soma and basal dendrites are in
deeper layers. A positive current would flow from the pial surface to deep
layers, as the consequence of an excitatory synaptic stimulation at the level of
the apical dendrites. This is not, however, the only possible interpretation, as
will be discussed further in the next question.
Q 2. On the basis of the direction of fields in MEG, or the polarity of scalp
potentials in EEG, can one say something about whether we record excitatory or
inhibitory PSPs?
In general, this question cannot be answered in a simple way. Different con-
figurations of EPSPs and IPSPs can give rise to magnetic fields with the same
direction, or to electric potentials with the same polarity at the cortical sur-
face. The reason is that an EPSP arising in the apical dendrites, or an IPSP
arising in the soma or basal dendrites, may result in an intracellular current
with the same orientation. In the former case there is an intracellular current
carried by positive ions from the apex to the base of a pyramidal cell; in the
latter, there is an intracellular current carried by negative ions from the base
to the apex, which is equivalent to a positive current in the opposite direction.
In both cases one would have a sink at the apex and a source at the base. Thus,
in order to be able to give an unambiguous reply to this question one would
need to have additional information about the site along the pyramidal cell
where the current is initiated.
Q 3. Can we state as a rule that the direction of current flow is from the cortical
surface to the depth of the cortex?
In general, we cannot. As described in the answer to Q 2. above, one can have
situations where the direction of the intracellular currents reverses. For exam-
ple, in the case of the experimental studies in hippocampal slices (Murakami
et al., 2002) we may observe, in a response to electrical stimulation of pyrami-
dal cells, an initial component of the magnetic field due to a current directed
from the basal to the apical sites, followed by a second component due to
reversed currents (Figure 1–6). It is likely that the same process may occur in
the neocortex. For example, the MEG evoked responses, elicited by stimula-
tion of the tibial nerve in man, show, as a function of time, a rotation of the
magnetic field patterns (Hari et al., 1996). This rotation may depend on
changes of the cortical areas being activated, but also on changes in the direc-
tion of intracellular currents of the pyramidal cells in one particular cortical
area that is engaged in these responses. It should be noted,
however, that the major part of the electric currents that cause a field that is
measurable at a distance flows perpendicular to the cortical surface.
Q 4. How many neurons, synchronously active, are necessary to generate a
recordable signal in the human MEG?
A precise answer to this quantitative question cannot be given. Nonetheless,
recent model studies suggest the likely order of magnitude. Murakami and
16 MEG: An Introduction to Methods
Okada (2006) computed that the current dipole of cortical pyramidal cell is in
the order of 0.2 pAm/cell. Considering that, according to Hämäläinen et al.
(1993), the weakest measurable cortical signals are on the order of 10nAm, we
may assume that such magnetic fields can be produced by about 50,000 cells
synchronously active.
Q 5. How large is the cortical area within which neurons must be synchronously
active to produce a measurable MEG signal at the scalp?
To answer this, we must take into consideration the estimates of the quantita-
tive distribution of neurons in the cortex. First, it is important to understand
some classic concepts about cortical organization. The cortex is organized
according to the columnar principle as proposed by Mountcastle in the 1970s
(see review, Mountcastle, 1997), which means that the basic unit of the mature
neocortex is the minicolumn: “a narrow chain of neurons extending vertically
across the cellular layers II/VI, perpendicular to the pial surface,” having a
cross-section with a diameter of about 50μm. A minicolumn in primates con-
tains about 80 to 100 neurons, although this number may vary between areas;
in the striate cortex the cell density appears to be 2.5 times larger. Many mini-
columns are bound together by short-range horizontal connections, sharing
“static and physiological dynamic properties” and thus forming what has
been denominated cortical columns or cortical modules (Mountcastle, 1997).
These columns, in the somatic sensory cortex, contain about 80 minicolumns
and are roughly hexagonal with a width of about 300–400μm (Favorov &
Diamond, 1990). Of course, these numbers are just estimates. Nonetheless,
based on these estimates we can attempt to give a rough answer to the ques-
tion formulated here. Assuming that a column with a diameter of 40μm con-
tains 100 cells, the cortical surface that would correspond to 50,000 cells
should form a patch with about 0.63 mm2 in area. If this cortical patch had a
circular form, then its diameter would be about 0.9 mm. It should be noted
that this columnar cortical organization is encountered throughout the sen-
sory and motor cortex, albeit with some peculiarities depending on cortical
area and sensory/motor modality as reviewed by Buxhoeveden and Casanova
(2002). Interestingly, these authors note that from a functional perspective,
cortical columns may exist in different dynamic states. They coined the term
“physiological macrocolumn” to indicate a set of cortical columns that coop-
erate in a given functional state or process. These physiological macrocol-
umns must be considered as dynamic ensembles consisting of a number of
columns that may vary as a function of time. In the context of the present
question, it is important to note that neurons in separate columns can present
oscillatory synchronous activities, mediated by tangential and recurrent con-
nections between different columns (Freiwald et al., 1995; Gray et al., 1989).
Q 6. Can action potentials make a significant contribution to MEG signals?
The most common answer to this question is that the contribution of action
potentials is minimal compared to that of synaptically meditated activity and
Electrophysiological Basis of MEG Signals 17
other slow waves. There are two reasons for this. First, the influence of action
potentials on recordings at a distance attenuates much more strongly than
postsynaptic potentials. Second, the probability that action potentials of dif-
ferent cells synchronize precisely is rather low, since action potentials are very
short and there is always a considerable jitter between the discharges of differ-
ent cells. Nevertheless, Murakami and Okada (2006) in their computer study
showed that sodium spikes may have dipolar moments considerably stronger
than previously thought. Assuming that the dipolar moment of a spike of a
cell is about 1pAm, 10,000 perfectly synchronized neurons could produce a
field on the order of 10nAm, which corresponds to the strength of a measur-
able cortical generator. This would, however, imply that these action poten-
tials should be perfectly synchronized—which may be the case during epilep-
tiform spike discharges, but is not likely to occur under normal conditions.
Nevertheless, synchronized population action potentials may contribute to
very-high-frequency MEG signals on the order of 600 Hz (Curio et al., 1994
and Hashimoto et al., 1996). Ikeda et al. (2002) detected such MEG
signals outside the pig brain, elicited by peripheral electrical stimulation;
these authors showed that the high-frequency oscillations consisted of a
presynaptic component generated by the specific thalamocortical axonal
terminals in layer IV, and another postsynaptic component due to activation
of cortical neurons.
Q 7. Is it possible to record, outside the head, MEG signals generated by deep
subcortical sources?
In considering this question as applied to human MEG recording, we should
note that not all types of magnetic sensors have the same sensitivity to distant
sources. In descending order of sensitivity to the depth of sources, magne-
tometers are most sensitive, followed by first-order axial gradiometers, sec-
ond-order gradiometers and, finally, planar gradiometers (Figure 1–7). Planar
gradiometers have maximum sensitivity to sources directly under them, i.e.,
superficial cortical sources (Hämäläinen, 1995), which makes them less sensi-
tive to artifacts and distant disturbances.
Experimental evidence supporting a direct answer to this question has
been provided by Hashimoto et al. (1996), who recorded, in pigs, somatosensory
evoked magnetic fields (SEFs) that were generated by neuronal populations at
the level of the thalamus. It should be noted that the latter were elicited by
electrical stimulation of peripheral structures, causing a high degree of
synchronous neuronal activity. This indicates that deep-lying structures in
the brain can generate sufficiently strong MEG signals to be detectable at a
distance from the brain surface, at least in the pig head.
In addition, studies using MEG arrays with partial head coverage in
humans have suggested the presence of generators of the P3 component of
auditory evoked responses in deep-lying source areas, including the hip-
pocampus (Okada et al., 1983), and thalamus (Rogers et al., 1991). However,
in these studies the evidence is indirect, since the localization of the sources
18 MEG: An Introduction to Methods
100
(ii)
(i)
10
(iii)
1
(iv)
0.1
1 2 3 4 5 6 7 8 9 10
depth (cm)
EEG by Vanhatalo et al. (2005) and for MEG by Burghoff et al. (2004).
Phenomena such as the Contingent Negative Variation (CNV) first described
by Walter et al. (1964), and the Bereitschaftspotential (readiness potential)
first described by Kornhuber and Deecke (1965), are typical examples of very
slow shifts of electric potential or of magnetic fields that can be recorded using
appropriate recording and analysis techniques. During slow-wave sleep, EEG
ultraslow frequency components of around 0.5 Hz have been recorded
(Achermann and Borbely, 1997; Amzica and Steriade, 1997; Massimini and
Amzica, 2001; Mölle et al., 2002) and the same is the case with MEG (Simon
et al., 2000). These components correspond to the ultraslow oscillations
found in cat neocortex during sleep, or anaesthesia, recorded intracellularly
from cortical neurons in layers II to VI. These oscillations consist of pro-
longed depolarizing and hyperpolarizing components, and have been ana-
lyzed in detail by Steriade and collaborators (2006). The study of ultraslow
oscillations in MEG and EEG is an active area of research, in conditions rang-
ing from studies of peri-infarct depolarisations in stroke patients, and cortical
spreading depression in migraine patients (Leistner et al., 2006), to record-
ings of preterm neonates (Vanhatalo et al., 2005). The underlying neurophys-
iological mechanisms of ultraslow shifts of cortical activity are well known in
respect to the changes in neuronal membrane potentials caused by increases
of CO2 tension and hypoxia in cerebral tissue (Speckmann and Elger, 2005),
and basic mechanisms of spreading depression in neurons and glial cells have
been put in evidence (Kager et al., 2002; Somjen, 2001).
General Conclusions
In this overview we have briefly covered some of the basic concepts of neuro-
physiology and biophysics that are important in understanding how magnetic
and electric fields are generated in the brain, and give rise to measurable sig-
nals at the scalp. A number of recent experimental and modeling studies have
provided new insights in this respect. The approach outlined here is that of
quantitative neurophysiology, in which the forward problem of MEG/EEG is
modeled. We currently have a comprehensive set of methodologies that allow
estimation of MEG/EEG signals measured at a distance, namely at the level of
the scalp, given a number of sources of electrical/magnetic activity at the
cellular level. In this context, the combination of neurophysiology and com-
putational models appears to be essential. The forward problem, however,
becomes more complex when we take into consideration the 3-dimensional
geometry of neuronal sources within the cortex and its convoluted structure.
For instance, when we state that to reach a MEG signal with the magnitude of
10nAm it is necessary that 50,000 pyramidal neurons are synchronously active
(since each cortical pyramidal neuron has a Q of 0.2 pAm), the forward prob-
lem is not completely solved. In fact, the estimate will be different depending
on (i) how these 50,000 neurons are distributed within the cortex, (ii) how
20 MEG: An Introduction to Methods
the cortex is oriented in relation to the pial surface, and on (iii) the degree of
synchrony. Such data are not readily available for most phenomena of inter-
est, and investigations at the level of the neocortex, combining basic neuro-
physiology with computer models of the forward problem with simultaneous
MEG/EEG recordings and structural/functional (fMRI) data, are therefore
necessary for a fuller understanding. Ongoing work using these methodologies
remains an active esearch area.
Acknowledgments
I thank Bob van Dijk and Stiliyan Kalitzin for their comments and suggestions
on the first draft of this paper.
Notes
1 For a more extensive mathematical treatment of this subject the reader may
consult Lopes da Silva, & van Rotterdam (2005) and Plonsey (1969).
2 The dot product can be defined for two vectors X and Y by
X ⋅ Y = X ⋅ Y cos θ
where θ is the angle between the vectors and X is the norm. It follows
immediately that X ⋅ Y = 0 if is X perpendicular to Y . The dot product
therefore has the geometric interpretation as the length of the projection of
X onto the unit vector Y when the two vectors are placed so that their tails
coincide.
References
Achermann, P., & Borbely, A. A. (1997). Low-frequency (< 1 Hz) oscillations in the
human sleep electroencephalogram. Neuroscience, 81, 213–222.
Amzica, F., & Steriade, M. (1997). The K-complex: its slow (<1-Hz) rhythmicity and
relation to delta waves. Neurology, 49, 952–959.
Burghoff, M., Sander, T. H., Schnabel, A., Drung, D., Trahms, L., & Curio, G., et al.
(2004). DC-Magnetoencephalography: direct measurements in a magnetically
extremely-well-shielded room. Appl Phys Lett, 85, 6278–6280.
Buxhoeveden, D. P., & Casanova, M. F. (2002). The minicolumn hypothesis in
neuroscience. Brain, 125, 935–951.
Cohen, D. (1968). Magnetoencephalography: evidence of magnetic fields produced
by alpha-rhythm currents. Science, 161, 784–786.
Curio, G., Mackert, B. M., Burghoff, M., Koetitz, R., Abraham-Fuchs, K., & Harer,
W. (1994). Localization of evoked neuromagnetic 600 Hz activity in the cerebral
somatosensory system. Electroencephalogr Clin Neurophysiol, 91, 483–487.
de Munck, J., Vijn, P., & Lopes da Silva, F. (1992). A random dipole model for spon-
taneous brain activity. IEEE Trans Biomed Eng, 39, 791–804.
Electrophysiological Basis of MEG Signals 21
Tesche, C. D., & Karhu, J. (2000). Theta oscillations index human hippocampal acti-
vation during a working memory task. Proc Natl Acad Sci U S A, 97, 919–924.
Traub, R. D., Jefferys, J. G., Miles, R., Whittington, M. A., & Toth, K. (1994). A bran-
ching dendritic model of a rodent CA3 pyramidal neurone. J Physiol, 481(1), 79–95.
Traub, R. D. and Miles, R. (1991) Neuronal networks of the hippocampus. New York:
Cambridge University Press.
Van ‘t Ent, D., Manshanden, I., Ossenblok, P., Velis, D. N., de Munck, J. C.,
Verbunt, J. P., & Lopes da Silva, F. H. (2003). Spike cluster analysis in neocortical
localization related epilepsy yields clinically significant equivalent source localiza-
tion results in magnetoencephalogram (MEG). Clin Neurophysiol, 114, 1948–1962.
van Rotterdam, A. (1987). Electric and magnetic fields of the brain computed by way
of a discrete systems analytical approach: theory and validation. Biol Cybern 57,
301–311.
Vanhatalo, S., Voipio, J., & Kaila, K. (2005) Full-band EEG (fbEEG): a new standard
for clinical electroencephalography. Clin EEG Neurosci, 36, 311–317.
Walter, W. G., Cooper, R., Aldridge, V. J., McCallum, W. C., & Winter, A. L. (1964).
Contingent Negative Variation: an Electric Sign of Sensorimotor Association and
Expectancy in the Human Brain. Nature, 203, 380–384.
2
Instrumentation and Data Preprocessing
Lauri Parkkonen
Introduction
This chapter reviews the methods and technology required for magnetoen-
cephalographic (MEG) measurements.
The magnetic fields due to neural activity are extremely weak, thus the
task of measuring them is challenging both in terms of required sensitivity,
and also in the ability to suppress interference several orders of magnitude
stronger than the signals of interest. Therefore, an appropriate combination
of multiple techniques, both physical and computational, is required to make
MEG measurements feasible.
24
Instrumentation and Data Preprocessing 25
Instrumentation
Review of Relevant Electromagnetic Concepts
• Electric current is the flow of charge carriers. The strength of the
current I is measured in Amperes (A).
• All currents generate a magnetic field around them. A magnetic field
is a vector field; it has a direction at each point in space. The field vec-
tor can be expressed as components, e.g., as the field strength to the x,
y and z-directions of the 3-dimensional space. Two distinct quantities
are referred to as magnetic fields: the magnetizing field or auxiliary
magnetic field H (measured in Amperes/meter), and the magnetic
flux density or magnetic induction B (measured in Teslas), commonly
called the magnetic field.
• Magnetic flux is the net magnetic field through a given surface (taking
into account only the field component perpendicular to the surface).
Magnetic flux f is a scalar quantity and its unit is Weber (Wb).
• Spatial derivative of the magnetic field is the rate of change of the field
along a certain direction. It is measured as Teslas per meter (T/m).
If all spatial derivatives are zero, the field is said to be homogeneous.
Field derivative is a tensor quantity; there are 9 first-order derivatives
(3 field components × 3 directions of the derivatives) at each point
in space. For example, the derivative of the z-component along the
x-direction at location r is denoted as ∂Bz(r)/∂x.
• A time-varying magnetic flux induces an electromotive force (a
voltage) in the circuit. Similarly, a circuit moving in a magnetic field
experiences an electromotive force.
• Magnetic permeability indicates the extent to which a material
magnetizes when subjected to an external magnetic field. Permeability
μ is the ratio of the internal and applied fields. Permeability of vacuum
relates the magnetization M and the magnetizing field H to the
magnetic field: B = μ0 (H + M).
• Ferromagnetic materials remain magnetic after being exposed to
a magnetic field (“permanent magnets”). They usually have a high
permeability.
26 MEG: An Introduction to Methods
To motivate the development and use of the relatively complex and costly
MEG instrumentation, let us compare the strength of the MEG signals to
some ambient magnetic fields. As shown in Figure 2–1, magnetic fields due to
brain activity are 8–9 orders of magnitude (about one billion times) smaller
than the Earth’s static magnetic field, which orients the compass needle. Even
the magnetic noise, (mostly generated by various electric devices and moving
magnetic objects) encountered in a typical laboratory environment is often
more than a thousand times stronger than any magnetic signal due to the
activity of neurons in the brain.
Sensors
Superconductivity
In order to understand the physical principles behind the sensors employed in
all modern MEG systems, we should familiarize ourselves with a remarkable
property exhibited by certain materials: when cooled down to a sufficiently low
1012
Earth’s
steady field
1010
Spectral Density (fT/Hz1/2)
108
Laboratory Lung
106 noise particles
Geo-
magnetic Heart (QRS)
104 noise
alpha rhythm
Evoked
fields
102 DC-SQUID “brain noise”
Dewar noise
100
Thermal noise of the body
10−2
0.01 0.1 1 10 100 1000
Frequency (Hz)
SQUID
Output
Φa voltage
V mm
Base ~1
Ba
electrode
IB Josephson junctions
Josephson
Counter electrode
junctions
Figure 2–2. A dc-SQUID with two Josepson junctions. (a) A schematic illus-
tration of the SQUID loop in the magnetic field Ba which causes magnetic
flux Φa in the SQUID loop. (b) A modern thin film dc-SQUID and the signal
coil on top of the SQUID loop.
Adapted from Hämäläinen et al., (1993).
The first SQUIDs were designed and made by James Zimmerman (1970)
in the late 1960s. At that time, most SQUIDs comprised only a single Josephson
junction because fabricating a properly working junction took a great deal of
care and effort. This kind of SQUID is read via an inductive coupling to an
external circuit operating at radio frequencies. Hence, these single-junction
SQUIDs are called rf-SQUIDs. Later, when manufacturing processes devel-
oped in the semiconductor industry allowed a relatively cheap “mass produc-
tion” of Josephson junctions, the two-junction dc-SQUID became more
popular, as it outperforms the rf-SQUID in many respects—most impor-
tantly, by having a lower noise level and by allowing much simpler electronics
and minimal crosstalk between channels. All these factors are indispensable
for MEG, and thus all modern MEG systems are equipped solely with
dc-SQUIDs.
Negative feedback
The periodic response of the SQUID to the variation of the applied flux, as
shown in Figure 2–3, prevents us from using the SQUID output directly as a
measure of the magnetic signal of interest; changes larger than half a flux
quantum would lead to ambiguous results. Therefore, in MEG systems, the
SQUIDs are operated in a flux-locked loop by providing them a feedback sig-
nal that cancels the effect of the signal we are measuring. In other words, the
output of the SQUID is held constant by artificially generating an additional
magnetic flux that “undoes” the effect of the actual flux to be measured. It is
easy to see that this additional artificial signal, or feedback signal, depends
directly on the signal to be measured; it only has the opposite sign and hence
the term negative feedback.
By using negative feedback, the SQUID is locked to a certain operating
point as illustrated in Figure 2–4b; changes in the measured signal only
30 MEG: An Introduction to Methods
(a) (b)
SQUID output voltage V SQUID output voltage V
Φa = n Φ0
IB > 2IC
ΔV
IB = 2IC
Φa = (n + 1/2) Φ0
ΔV
IB/IC
−2 2
Bias current
IB = 1.5IC
Φa/Φ0
–1 –0.5 0.5 1
Magnetic
flux
deviate the SQUID output slightly from the operating point until the feedback
controller reacts and adjusts the negative feedback signal accordingly to again
perfectly cancel the measured signal. Thus, the signal that the MEG system
outputs and stores into a computer is not the SQUID output as such, but the
inverted negative feedback signal, which linearly tracks the true measured
signal, even if it undergoes variations larger than a flux quantum.
Sensor Coils
To optimize the sensitivity, SQUIDs are made rather small, typically less than
1 mm in outer diameter. Due to the small surface area, SQUIDs have rather
poor coupling to the magnetic field. In MEG applications, SQUIDs cannot be
used as they are; the coupling must be enhanced with flux transformers that
“squeeze” more magnetic flux into the SQUID loop by collecting it from a
much larger area. This is not a drawback, since the use of flux transformers
allows us to measure different components of the magnetic field without
changing the SQUID geometry.
Like SQUIDs, the flux transformers are made of superconducting mate-
rial. They comprise a pick-up coil closest to the brain, an optional, more
distant compensation coil, and a signal coil on top of the SQUID loop
(Figure 2–4a,c). These coils are typically connected in series so that the net
shielding current induced by the magnetic fields at the pick-up and compen-
sation coils circulates through the signal coil, and thus generates a magnetic
Instrumentation and Data Preprocessing 31
(a) IB (b)
SQUID output voltage V
V
Operating
point
Voffset
−1 −0.5 0.51 Magnetic
flux Φa/Φ0
dV
gain =
d Φa
(c)
PI Output
open closed
loop loop
Rf
Lp Ms Mf
Ls Φa Lf
Lc If
Tuning signal
in liquid helium, T = 4.2 K in room temperature input
Figure 2–4. (a) The principle of a flux transformer, which picks up the field
due to a neural current (green arrow) and couples it to a SQUID (blue
ring). (b) Illustration of the operating points at which the SQUIDs locks
when used in a flux-locked loop, and (c) a simplified block diagram of the
sensor electronics (utilizing direct readout) of one MEG channel.
field and flux to the SQUID loop. Because of superconductivity, flux trans-
formers work also at DC, whereas normal transformers do not, since a static
magnetic field does not induce currents in normal, stationary conductors.
The simplest configuration is a magnetometer: a single pick-up coil and
no compensation coil (Figure 2–5a). This setup measures the magnetic field
component along the direction perpendicular to the surface of the pick-up
coil, usually denoted as Bz. While being very sensitive to nearby sources, such
as neural currents in the brain, a magnetometer is sensitive also to sources far
away. To decrease the sensitivity to distant sources, one may add a compensa-
tion coil that measures mostly the interfering signal; this configuration is a
gradiometer (see Figure 2–5b,c), which measures a spatial gradient of a mag-
netic field component rather than the field component itself. The underlying idea
is that the far-away interference sources manifest themselves as homogeneous
magnetic fields, i.e., the field is about the same at close-by spatial locations.
32 MEG: An Introduction to Methods
(a) (c)
(d)
z dA
(b) y dA
Thus, the interference field at the pick-up coil and compensation coil is about
the same. When those coils are wound in opposite directions, the interference
fields produce no net shielding current, and thus do not couple to the SQUID,
making the sensor blind to sources distant enough to be seen as homoge-
neous fields.
The two coils of a gradiometer can be arranged in several ways to sensi-
tize the sensor for different spatial derivatives of the field. Perhaps the most
intuitive configuration is shown in Figure 2–5c, where the coils are along the
same radial axis, but the pick-up coil is typically 5 cm closer to the head. This
is called an axial gradiometer and it measures the change of the radial field
component along the radius.
Another option is to place the coils side-by-side in the same plane to
form a planar gradiometer (Figure 2–5b). Both of these arrangements are
insensitive to homogeneous fields but their responses to nearby sources are
very different; the signal from an axial gradiometer peaks for sources around
the rim of the sensor while planar gradiometers give the maximum signal for
sources right beneath them. Formally, these spatial sensitivity patterns can be
described with the concept of lead field. It is a fictitious vector field whose
value at a spatial location gives the direction of the current that yields the
maximal output at that location, and the gain with which the source current
affects the output of the sensor. Thus, each sensor type, or pick-up coil geom-
etry, has a specific lead field, some of which are illustrated in Figure 2–6.
Knowing the lead field Li of the i’th sensor, the output of that sensor, bi, can
be expressed mathematically as
bi = ∫ L i (r ) jp dG. (2–1)
G
where the integration is carried out throughout the volume G where currents
can flow, and r points to the center of the integration element dG.
Instrumentation and Data Preprocessing 33
Planar gradiometer
Magnetometer Bz x
z
Axial gradiometer dBz /dz dBz /dx dBz /dy
The lead field can be estimated by virtually scanning the volume G with a
set of three orthogonal unit-strength test current dipoles. At each spatial loca-
tion, the signals elicited by the x, y and z-directed dipoles are the correspond-
ing components of the lead field vector at that location. Assuming that we can
calculate the total magnetic field B(r) at location r due to a current dipole in
a conducting volume, the output of the i’th channel is
N
bi = ∫ B (r )⋅ dA ≈ ∑ w k B (rk )⋅ n k (2–2)
A
k =1
where the integral is calculated over the surface A of the pick-up and compen-
sation coils, taking the winding direction of the coil into account, and with
r pointing to the center of the surface patch dA (see Figure 2–5d); the approx-
imation by N points involves the unit normal vectors nk and surface areas wk
associated with each point k.
Lead fields have practical relevance also when interpreting MEG data
visually. Figure 2–7 shows auditory evoked responses measured with an array
comprising only magnetometer sensors, and with a similar array of planar
gradiometers. In this experiment, there was significant activity only in the left
and right auditory cortices at about 100 ms after the presentation of a short
tone. For the source in either auditory cortex, there is a group of magnetom-
eters showing a positive deflection (field coming out of the head), another
group with a negative deflection (field going into the head). The magnetom-
eters right above the underlying source current do not show any signal. The
situation is quite the opposite for the planar gradiometers shown in the same
34 MEG: An Introduction to Methods
100 ms
Right
Left Right Left
Back Back
Figure; they peak just on top of the neural current. Thus, knowing the sensi-
tivity pattern of the sensors is essential for the correct interpretation of the
data. Proper source modeling algorithms, to be discussed in the subsequent
chapters, take the different lead fields automatically into account, so that the
source currents can be localized correctly. Nevertheless, for the human
observer who wants to get a rough idea of the source locations, or needs to
disentangle the contributions of different neural sources at the sensor level,
understanding the sensitivity pattern is essential.
Sensor Arrays
The first MEG measurements with SQUIDs (Cohen, 1972) were done with a
single-channel instrument. While this is sufficient for the mere detection of
the signal, a proper mapping requires multiple measurements at distinct spa-
tial locations. Such mapping can be done even with a single-channel system
by repeating the measurement multiple times with the measurement probe at
different locations. This is not only cumbersome and time-consuming, but
also prevents studying, e.g., induced synchronization between two brain
areas. During the 1980s several MEG systems with the number of channels
ranging up to 37, covering approximately one lobe of the brain at a time, were
developed. In 1992, the first whole-head MEG system, comprising 122 planar
Instrumentation and Data Preprocessing 35
gradiometer channels, was introduced (Ahonen et al., 1993b) and since then,
the MEG market has been dominated by whole-head systems.
Several factors and trade-offs are to be considered in the design of the
sensor array. For example, larger pick-up coils yield higher signal-to-noise
ratios (considering only the intrinsic system noise), but the larger the loops
the fewer of them fit on the head surface. In addition, the distance from the
most superficial neural sources sets a limit on the highest spatial frequencies
present at the pick-up coils; thus, packing the sensors more densely increases
the spatial resolving power of the array only up to a limit. Instead of increasing
36 MEG: An Introduction to Methods
the density of sensors that measure one particular field component, one can
measure multiple independent components of the field at each sensor loca-
tion to increase the total information captured by the array (Ahonen et al.,
1993a).
The optimal size of the pick-up coil depends also on the noise level of the
SQUID: lower SQUID noise allows a reduction in the size of the pick-up coil
without an adverse effect on the signal-to-noise ratio for a given source.
Strictly speaking, this is true only if we assume the instrument to be the sole
source of noise in the measurement. In practice, the “brain noise,” i.e., the
brain activity not of interest, dominates among the noise sources, which
complicates the pick-up loop optimization further.
The coverage of the sensor array is another difficult optimization task.
A sensor helmet as extended as possible would be optimal in capturing the
neuromagnetic fields; however, practical limitations exist—for example, the
MEG device should not severely limit the visual field of the subject. The overall
size of the array should also be optimized to fit a high percentage of the popu-
lation while minimizing the average distance from the sensors to the scalp.
The distance between the sensor coils and the head surface should be minimized
to maximize the neuromagnetic field at the pick-up coils. Further, the coils
must be superconducting, i.e., their temperature should remain below the
critical value while the head surface is at body temperature. Maintaining this
high temperature difference (about 300 K or °C) across a relatively small
distance of 2–3 cm without excessive use of the coolant requires elaborate
thermal isolation. A special container called Dewar, after the inventor James
Dewar, comprises two concentric vessels with a vacuum jacket and radiation
shields in between. The vacuum prevents heat conduction from the outside to
the inside vessel, and the shields block thermal radiation. The sensors reside
in the inner vessel2 immersed in the coolant (liquid helium, T = 4.2 K). An
MEG Dewar has to be strictly nonmagnetic not to distort the fields being
measured.3 The Dewars are usually built of glass-fibre composites which are
magnetically transparent. Unfortunately, helium atoms slowly diffuse to the
vacuum through the glass-fibre wall of the inner vessel. Therefore, the outer
surface of the inner vessel is usually covered with a medium that absorbs the
helium atoms.
Despite the extreme thermal isolation, there is still heat leakage, albeit
small, into the inner vessel, causing the liquid helium to slowly evaporate. The
gaseous helium exits the Dewar along an exhaust line which guides the gas out
of the system and the magnetically shielded room. The helium gas is either
collected into pressurized containers for reliquification, or just let out into the
open air outside of the building. The feasibility of re-collection depends on
the local price of liquid helium, and is seldom profitable for just one MEG
system alone.
Instrumentation and Data Preprocessing 37
Connection to
electronics
Neck plug
(conventional thermal insulator)
Outer vessel
Vacuum jacket and
radiation shields
Inner vessel
Liquid helium
Sensor array
A typical whole-head MEG system boils 10–20 liters of liquid helium per day.
The helium reservoir of the Dewar is usually 70–90 liters, thus a refill is
required 1–3 times a week to keep the system operational. Liquid helium is
transferred from a storage Dewar by means of a vacuum-isolated siphon. The
storage Dewar is pressurized by gaseous helium to “push” the liquid along the
siphon into the MEG Dewar. Some liquid evaporates when cooling the siphon
from room temperature down to 4.2 K. Typically, up to 10 liters of liquid per
transfer should be budgeted for such losses.
Care should be taken to do the refills in time. If all liquid helium has
evaporated and the temperature of the inner vessel thus starts to increase, the
helium atoms trapped in the absorbant are released in the vacuum due to
their increased thermal energy, and they start contributing to the heat
conduction from the outer to the inner vessel. In such an event, the Dewar
usually cannot be cooled down simply by transferring liquid helium; instead,
the vacuum jacket has to be re-evacuated first.
Gantry
The mechanical system supporting the Dewar is called the gantry. It often
allows adjusting the elevation and angle of the Dewar to accomodate subjects
of different heights and in different measurement positions, such as seated or
supine. Alternatively, the height adjustment can be addressed by moving the
seat up/down. The gantry should be very rigid, since even minute movements
of the sensors in the remanent field inside the shielded room gives rise to
artifacts.
38 MEG: An Introduction to Methods
Electronics
through the EEG electrodes and amplifiers, should any electric system connected
to the subject break and supply dangerously high voltages. The preamplifier
ground is, therefore, often referred to as the isolated ground, or iso-ground. The
required galvanic isolation is provided either by a special isolation unit, or by
converting the EEG signal to a digital form already in the preamplifier and using
an optic fiber to convey the samples to subsequent processing stages.
In addition to the MEG and EEG signal processing, complete MEG systems
also feature electronics for driving the head-position tracking coils and the
artificial sources in a phantom, and for monitoring the level of liquid helium
in the MEG Dewar.
Data Acquisition
With the advent of digital SQUID electronics, the boundary between the
electronics and data acquisition became fuzzy; traditionally, the acquisition
system sampled and stored the amplified and filtered analog signals from the
electronics, whereas nowadays the conversion to the digital domain happens
much earlier on the signal path, and many of the acquisition system tasks are
handled by the main electronics. The theoretical background remains the
same: sampling theory and relevant performance criteria are reviewed briefly
in the following paragraphs.
The bulk of the MEG activity occurs in the conventional Berger bands, defined
by Hans Berger in 1920s after his first EEG measurements (Berger, 1929).
MEG responses typically contain frequencies up to about 100 Hz with a grad-
ual fall-off towards higher frequencies as shown in Figure 2–10. More recently,
the higher frequency bands, up to 700 Hz, have also received attention follow-
ing the discovery of certain fast oscillatory responses measured directly on the
cortex, and also noninvasively by EEG and MEG. The 600-Hz burst response
to electric nerve stimulation (Curio, 2000; Hashimoto, 2000; Okada et al.,
2005) contains probably the highest-frequency oscillatory components so far
detected by MEG. Responses from peripheral nerves, generated by compound
action potentials and not much studied by magnetic measurements, extend
above 1 kHz in frequency content. Generally, for MEG and EEG, as for most
physical systems, it appears that the higher the frequency, the weaker the
signal. The 600-Hz response, for example, is hidden in the noise due to the
background brain activity and instrumentation, and can be recovered only by
averaging hundreds of responses to a specific stimulus type.
The ratio of the maximum to the minimum signal amplitude, or the
dynamic range, is rather low in the MEG signals that really originate in
the brain; however, the residuals of the environmental interference, as well as
the artifactual signals from the subject, may exceed the brain signals by orders
of magnitude.
40 MEG: An Introduction to Methods
100
Spectral density (fT/√Hz)
10
1
1 10 100 1000
Frequency (Hz)
Most physical quantities, including magnetic field and its derivatives, are
represented by a temporally continuous signal, that is, the signal has a value
at every point in time. If there is a limit to the rate at which the signal can vary,
all available information is retained by considering the amplitude of the con-
tinuous signal only at certain intervals. Indeed, it can be shown that it is suf-
ficient to sample the continuous signal at a rate that is twice the frequency of
any component of the signal, and yet to perfectly reconstruct the original sig-
nal from the discrete samples. This important result is known as the sampling
theorem. It should be stressed here that all signal components, whether of
Instrumentation and Data Preprocessing 41
interest or noise, must be below half the sampling rate; as a result of the
sampling process, those signals that are above will fold along the frequency
axis to appear as lower frequencies. This undesirable phenomenon, called
aliasing, can be avoided by low-pass filtering the signal before sampling to
ensure that there is no signal above half the sampling rate, or the Nyqvist
frequency. This low-pass filtering just prior to sampling to avoid aliasing is
often referred to as anti-alias filtering. Figure 2–11 illustrates sampling of
continuous signals.
Given the MEG signal frequencies and the Nyqvist condition described
above, the sampling rates range between 300 Hz and 4 kHz. It is often desir-
able to temporally oversample the signal of interest to avoid the non-idealities
of the anti-alias filters such as phase distortion and the finite fall-off rate, and
also to allow for an easier reconstruction of the original signal by linearly
interpolating the values between the samples instead of using the optimal, but
computationally expensive, sinc interpolation.
Converting a signal from the analog to digital domain involves the
discretization of the signal amplitude as well. This process is often referred to
as quantization, since the graded analog amplitude values are represented
with a finite number of amplitude bins. The height of each amplitude bin, or
the quantization step size, determines how faithfully the amplitude of the
original signal can be represented in the digital domain. A smaller step size
allows more precise reconstruction of the signal, but also requires more bits
and thus more storage space to encode the steps. Assuming a certain step size,
the signal dynamics determines how many such steps are required in the
quantization.
Practically all quantizers utilized in MEG/EEG data acquisition systems
are linear and memoryless, i.e., the quantization step size does not depend on
the signal amplitude (see Figure 2–12) nor does the output depend on its
previous values. Analyzing the dynamics of such a quantizer is straightfor-
ward if we assume that the signal amplitude has a uniform probability
distribution when considering the amplitudes that would fall within a single
quantization step. Let L be the number of steps of size Δ and q(t) the quanti-
zation error, i.e., the difference between the original and quantized signals.
This error can be considered as an uncorrelated noise source even though
q(t) is a deterministic function of the input signal. The variance of the quan-
tization error
∞
s q2 = ∫ q 2 pq (q ) dq (2–4)
−∞
where pq(q) is the probability distribution of the error term, which can be
assumed to be uniform provided that the input signal does not exceed the
normal range of the quantizer; the probability density function is thus
⎧1/ Δ q ≤ Δ /2,
pq (q ) = ⎨ (2–5)
⎩ 0 otherwise.
(a) fS
(b) xc (t)
amplitude
t
⏐Xc (f )⏐ time
(c)
f / fs
–1/2 1/2 frequency
xs (t)
(d)
t / Ts
0 1 2 ... time
⏐Xs (f )⏐
(e)
f / fs
−3/2 −1/2 1/2 3/2
frequency
⏐H(f )⏐
(f)
f / fs
−1/2 1/2 frequency
42
Instrumentation and Data Preprocessing 43
(a) y
3
1
Δ
x
−4 −3 −2 −1 1 2 3 4
−1
−2
−3
q
(b)
+1/2
x
−1/2
traded for a larger dynamic range provided that there is some uncorrelated
noise on top of the signal. This perhaps counterintuitive effect is best explained
with an example; consider a slowly varying signal with an amplitude 1.7 at
some point in time. When this signal is fed to a quantizer whose step size
Δ = 1, the quantizer output will be 2 since the value 1.7 falls into the bin
[1.5, 2.5] centered around 2. On the other hand, if broadband noise (whose
variance is larger than the step size) is added to the input signal, and this
signal is sampled at a rate which is several times higher than what is required
by the sampling theorem, the quantizer output is 2 for most of the samples,
1 almost as often, and may, less frequently, also assume other values if
the noise variance is large. The average of multiple such samples approaches
1.7 without a limit when the number of samples tends to infinity. Thus, tem-
poral oversampling is able to extend the dynamic range at the small-signal
end. When the sampling rate is dropped, i.e., the signal is downsampled,
additional accuracy is gained; it can be shown that downsampling by a factor
F yields
N ds = log 2 F (2-9)
bits in the small-signal end, provided that the original signal contains enough
wideband noise.
Averaging trials (see previous section) has the very same effect; the average
signal can show brain responses whose amplitude is smaller than the quantiza-
tion step size of the raw signal, provided that the implementation of the averager
is such that it preserves the emerging “sub-bits.” In other words: if the quantiza-
tion step is much smaller than the system noise amplitude in the frequency band
of interest, the step size does not limit the smallest discernible brain response,
since averaging is required anyway to recover the response amid noise.
Acquired Signals
MEG systems typically include 8–32 independent trigger lines, or bits, which
are encoded on one or more trigger channels.
MEG experiments may also involve recording additional signals from the
subject such as the gaze direction (provided by an eye tracker), position of
fingers or limbs, or speech. These can be acquired using the auxiliary analog
inputs provided by most MEG systems.
All the above input signals are usually sampled synchronously at the same
sampling rate and stored into a single data file. Common sampling rate usu-
ally implies equivalent low-pass or anti-alias filtering of all signal; however,
the trigger signals are sampled without filtering. Since filtering involves a
delay, the trigger signals must be shifted accordingly in time, or the trigger
event timing has to be compensated mathematically off-line to ensure perfect
synchronization of all acquired signals.
Acquisition Modes
Interference Suppression
As discussed earlier in this chapter, MEG signals are several orders of magni-
tude weaker than the ambient magnetic noise due to sources like powerlines,
electric appliances, and traffic. Detecting MEG signals in a magnetically silent
environment is already challenging, and doing it in normal surroundings is
even more so. Sufficient suppression of environmental magnetic interference
involves a combination of multiple methods. The mostly widely applied tech-
niques are briefly reviewed in this section.
Time (s)
(a)
0 1 2 3 4 5 6 7 8 9
MEG
channels
trigger A
trigger B
Σ Σ
Figure 2–13. Data collection modes provided by MEG systems. The raw
MEG/EEG signals can be stored either (a) continuously, (b) in short
windows about a trigger event (epoch mode), or (c) as on-line aver-
ages. Often the continuous or epoch mode and on-line averaging can be
engaged simultaneously. Here, the simulated experiment involves two
different stimuli whose onsets are marked by triggers A and B, and
accordingly, two sets of average responses are computed.
120
Mu 2 mm
and
100 Al 12 mm
80
Shielding factor (dB)
Only Mu,
60 2 mm
Only Al,
40
12 mm
20
0
0.001 0.01 0.1 1 10 100
Frequency (Hz)
Passive magnetic shields can be enhanced by active systems that measure the
interference field and generate a compensating field to cancel the interference
at the location of the MEG system. The typical active compensation system
comprises a flux-gate sensor, driver electronics, and pairs of Helmholtz coils
outside of the room to supply the cancellation fields. Such a setup can provide
10–30 dB of additional shielding if the interference sources are far away (tens
of meters or more), so that their fields are approximately homogeneous at the
location of the room. Unfortunately, nearby sources may turn problematic,
since proper compensation would require the spatial derivatives of the field to
be taken into account. Tuning the setup against a particular nearby source
may still give satisfactory results.
Despite all the effort of suppressing ambient magnetic fields, some residual inter-
ference is typically still present within the shielded room. Thus, further noise
reduction techniques have to be applied within the MEG systems themselves.
Gradiometrization. As discussed earlier in this chapter, employing gra-
diometers instead of magnetometers is a straightforward method to protect
the MEG sensors from far-away interference sources; a gradiometer’s response
to a source falls off much faster with distance than that of a magnetometer.
A carefully manufactured (well-balanced) gradiometer can attenuate homo-
geneous fields by as much as 60 dB (factor 1,000). On the other hand, fields
from the most distant brain regions are picked up better by magnetometers
than gradiometers.
Reference sensor array. Interference can also be measured explicitly and
then subtracted from the signals. Reference sensors located some tens of centi-
meters away from the MEG helmet do not measure brain signals, but capture
mainly the interference. By optimally coupling the output of the reference-
sensor array to the MEG channels proper, the interfering signal can be removed.
This arrangement works well with homogeneous interference field; however,
the presence of gradients may degrade the performance, as the interference at
the helmet must be extrapolated from the measurements at the reference
sensors. For this purpose, the reference sensor arrays usually include both
magnetometers and gradiometers. The reference-sensor approach can also be
considered as a higher-order gradiometrization (Vrba & Robinson, 2001).
The optimal couplings (“weights”) from the reference sensors to the
MEG sensors can be determined either by direct calculation, if the geometry
is known to a high precision, or adaptively from real measurements of
external interference by the same array.
When performing correlation or coherence analysis of MEG data, it
should be noted that the couplings to the reference sensors may introduce
spurious correlations between the MEG channels, if not explicitly addressed.
Instrumentation and Data Preprocessing 49
The SSS subspaces are derived from series of spherical harmonic functions.
In the SSS framwork, the data are first expressed as two multipole expansions,
one for the inside and the other for the outside contribution, in the spherical
harmonic spaces. Subsequently, the sensor-level data are reconstructed using
only the inside expansion. Both series are truncated to stay within the limits
imposed by the number of channels in the system; the inside expansion typically
corresponds to about 100 degrees of freedom.
The SSS method is data-independent and time-invariant, however, it
does require precise information on the geometry of the sensor array; with
0.1 percent calibration accuracy, the shielding provided by SSS is roughly 40 dB.
Co-registration
The MEG source estimate, i.e., the estimated spatial distribution of the neural
(primary) currents given the MEG measurements, is usually visualized super-
imposed on the anatomical MR-image of the subject. In addition, the esti-
mate can be mapped into a normalized space, such as Talairach, the Montreal
Neurological Institute (MNI) standard brain, or other atlas brains; see, e.g.,
Fischl et al. (1999); Mazziotta et al. (2001); Van Essen (2005); Toga et al.
(2006). On the other hand, the MEG measurements are taken at locations
known only with respect to the MEG device itself, instead of the anatomy of
the subject. For subjects with smaller heads, the MEG helmet typically allows
for head movements as large as a few centimeters; not knowing the head posi-
tion within the sensor array would lead to drastic errors when superposing
the MEG sources onto an individual anatomical MRI. Therefore, MEG devices
include a subsystem to determine the position of the head with respect to the
MEG sensors.
Since MEG — unlike MRI — cannot directly measure the position of
the head, small coils generating magnetic fields at known locations on the
scalp of the subject are employed in the head position measurement. When
the coils are energized, the MEG sensor array can be used to localize the
coils, just like it is used to localize neural currents in the brain. If we could
place the coils at anatomical locations that are accurately identifiable on ana-
tomical MRIs, this step would be sufficient to provide us with a coordinate
transformation between the MEG device coordinate system and the MRI
device coordinate system. Unfortunately, such anatomical locations are either
not covered by the MEG sensor array (nasion and the tip of the nose) or they
are inconvenient as coil locations (preauricular points). This necessitates the
definition and use of a head coordinate system, which is based on landmarks
identifiable accurately both in the MRIs and on the real head. By measuring
the coil locations in the head coordinate system with a 3D digitizer, and
combining that information with the locations of the coils in the MEG
device coordinate system, we are able to bridge the gap between the MEG
and MRI device coordinate systems and thus to translate MEG results onto
the MRI space.
Instrumentation and Data Preprocessing 51
z
(a) (b)
LPA
Nasion y
RPA
- visualization of
MEG source
estimate
Safety
MEG is inherently very safe; it measures magnetic fields that are always
present outside of the head, and does not employ high magnetic fields, as does
(f)MRI, or radioactive tracers, as does PET. The only safety concern is liquid
helium, utilized as a coolant for the superconducting parts of the MEG sys-
tem. Helium is nontoxic and nonflammable but may replace oxygen when
present in the air in large quantities. Helium gas is lighter than air, so it con-
centrates at the ceiling level. Liquid helium may also cause severe frostbites
because of its very low temperature.
During normal operation, the helium gas evaporating inside the MEG
Dewar is fed outside through an exhaust line. Should this line be blocked, the
pressure inside the Dewar starts to rise gradually. To prevent the system from
eventually exploding, the Dewars are fitted with safety valves that let out the
excess gas by releasing it to the air inside the shielded room. Therefore,
the shielded room must have proper ventilation, arranged in such a way that
the air outlet is near the ceiling.
A sudden loss of the vacuum isolation of the MEG Dewar leads to a rapid
boil-off of liquid helium. Since one liter of liquid helium expands to about
750 liters of gas at room temperature and normal atmospheric pressure, many
of the MEG Dewars are equipped with a high-capacity safety exhaust line to
provide a low backpressure path to the outside air.
Stimulators
Most MEG measurements involve stimulation of the subject by, e.g., sounds,
touch, images, video clips, or their combinations. Delivering stimuli without
interferring with the MEG signals is often challenging, as many conventional
devices that could be used for generating the required sensory input also pro-
duce unwanted magnetic signals that are picked up by the MEG. In addition,
stimulus delivery should be temporally precise; sloppy timing yields smeared
responses, particularly in the primary sensory areas. The following sections
describe the principles of MEG compatibility and highlight the most com-
monly employed stimulation setups for MEG. Developing, selecting and
applying stimulators is a large undertaking; this section is merely an brief
introduction.
Electromagnetic Compatibility
spurious RF noise. The regulatory directives that govern the level of these
emissions are not strict enough for MEG compatibility, and thus a fully com-
pliant device may still be unusable with MEG. Even if the generator of RF
signals, e.g., a cell phone, is outside of the shielded room, directly or indirectly
(capacitively or inductively) coupled wires may act as antennas and carry the
unwanted interference into the room. Therefore, all cables entering the
shielded room should be properly low-pass-filtered at the feedthrough to
remove any RF contamination. Omitting feedthrough filtering may result in
a setup that works most of the time if the environment is relatively RF-free,
but exhibits spurious artifacts when RF sources happen to be in the vicinity.
Timing
Auditory
Proper
timing
+/− 10 ms
jitter
Visual
per second. When driving a projector at a higher rate, it may adapt by simply
dropping out a frame every now and then. This coarse downsampling gives
rise to jittering on top of the constant delay. This jitter cannot be compen-
sated for, except by triggering each trial by measuring the stimulus onset
directly on the screen with a photodetector. Note that this elaborate arrange-
ment is not required to remove the effect of a constant delay, yet it is always
worthwhile to verify with a photodetector that a new setup (including the
stimulation software, graphics hardware and driver, video projector and the
employed resolution and refresh rate) works flawlessly.
Somatosensory
Other Modalities
Preprocessing
Signal-to Noise Ratio and Averaging
The MEG signal amplitude is affected by several factors: the extend of the
activated area, the level of neuronal synchrony, the anatomical location and
orientation of the source, and cancellation effects due to opposing coincident
nearby activations. Signal amplitude may also change due to medication and
pathologies. Therefore, MEG response amplitudes span a wide range, from
few femtotesla to a picotesla.
Noise from several sources hampers MEG. Brain activity not of interest
(“brain noise”), biological noise from sources other than the brain, instru-
mentation and ambient magnetic interference, all contribute to the noise seen
in MEG recordings. The relative strengths of these sources depend on the
frequency: in rough terms, at the lowest frequencies (below 1 Hz) the ambient
and biological noise are usually most prominent; the mid-frequency band
(1–100 Hz) is dominated by brain noise (except at the line frequency of
50/60 Hz), and at higher frequencies most of the noise originates in the MEG
instrument itself.
Instrumentation and Data Preprocessing 59
Filtering
Signal-to-noise ratio of MEG data can be improved also by limiting the window
of frequencies so that only the band where the response’s energy lies is
retained. This operation is filtering, also referred to as time-domain or temporal
filtering in order not to confuse it with spatial filtering.
To understand how filtering may reduce the noise level, a simple example
comes in handy: If the noise spectrum can be assumed white,5 the RMS (root-
mean-square) amplitude of noise n = Bn Δf where Bn is the spectral density
of noise and Δf is the bandwidth. For example, if the white noise level
Bn = 5fT / cm / Hz and the pass-band is 0–100 Hz thus Δf = 100 Hz, the
observed noise amplitude n = 50 fT/cmRMS. If we know that the responses of
interest are confined to 0–25 Hz and we filter the data accordingly, the noise
N20m
100 fT/cm
100 ms
amplitude drops to 25 fT/cmRMS, i.e., to one half. Here, the gain in SNR
achieved by filtering is equivalent to prolonging the experiment fourfold.
Some responses have lower- and higher-frequency components which
can be separated by filtering. The somatosensory response of the previous
example (Figure 2–17), if measured with a wide enough pass band, comprises
the traditional low-frequency responses and a high-frequency burst-like
response around 600 Hz, which react to experimental manipulations differ-
ently and likely reflect partially different neural events (Curio, 2000; Hashimoto,
2000; Okada et al., 2005). Figure 2–18 shows such an average response and its
low- and high-frequency components.
Low-pass filters limit the frequency band at its upper end, i.e., attenuate
all frequencies above their corner frequency, usually denoted as the frequency
at which the filter drops the signal amplitude by 3 dB (to 71 percent), assuming
unity gain at DC. Similarly, high-pass filters remove the frequency compo-
nents below their corner frequency. These two types can be combined into a
bandpass filter. Conversely, a contiguous range of frequencies can be removed
by a band-stop filter, and when this range is very narrow the filter is often called
a notch filter as it has a “notch” in the frequency response. Regularly-spaced
notch filters form a comb filter, which is useful in removing a signal with a
harmonic structure, e.g., the line frequency with the fundamental at 50/60-Hz
and its harmonics (100/120 Hz, 150/180 Hz, etc.).
No filter has an infinitely steep transition from the pass-band to the stop-
band but filters rather exhibit a certain roll-off rate, often expressed as dB per
octave (doubling of the frequency). This value is influenced mostly by the
order of the filter; the attenuation of a 2nd-order low-pass filter typically
Low-pass
filtering at 300 Hz Bandwidth
0.1 – 300 Hz
100 fT
Bandwidth
0.1 – 1200 Hz
5 fT
100 fT Bandwidth
300 – 1200 Hz
10 ms
High-pass filtering
at 300 Hz
activity. If this residual interference occupies the same frequency range as the
brain responses of interest, the simple time-domain filtering cannot be used
to remove the interference. However, exploiting the statistical and spatial
properties of the interference can still provide a way to clean the data.
Independent component analysis (ICA) is a blind source separation
method that seeks for directions in the signal space that are maximally inde-
pendent in the statistical sense, and expresses the data along these directions;
see, e.g., Hyvärinen and Oja (2000) for a thorough description of ICA and its
implementations. Since artifactual signals are most often independent of
brain activity, ICA suits the task very well (Vigário et al., 2000): however, ICA
relies on the user to classify the obtained independent components to reflect
either artifacts or brain activity.
If there is a priori information about the artifact, the suppression method
can take advantage of it. Knowing the precise timing (or determining it from
the data if possible) of spatially similar, repeating events—e.g., the magnetic
artefact due to a heartbeat—allows us to average several such events for a bet-
ter signal-to-noise ratio of the artifact itself, and thus obtain its spatial pattern
on the sensor array. Once the pattern is known, it can be removed from the
data by signal-space projection (see earlier discussion). If the pattern is not
stable but undergoes a stereotypical sequence, principal component analysis
(PCA) can be employed to extract a small set of patterns from the “artifact
average” to be removed from the data.
Within the signal-space separation framework, many artifactual signals
can be detected as temporal correlations between the inside and outside
spaces. If high correlations are found, the corresponding interference can be
removed by projection (Taulu and Simola, 2006); this method is known as
temporal signal-space separation, or tSSS.
Notes
1 Wavefunction in quantum mechanics gives the probability of finding the
particle at a certain location in space.
2 In some MEG systems the sensor coils are located in the vacuum space and
they are thermally linked to the inner vessel.
3 An experimental MEG system by Volegov and colleagues (2004) employs a
superconducting magnetic shield inside the Dewar to provide high attenu-
ation towards external interference fields, and to act as a “magnetic mirror”
which turns magnetometers into axial gradiometers.
4 Sensation level: decibels with respect to the individually-determined hearing
threshold.
5 In a white signal, the spectral density is the same at all frequencies, i.e., the
spectrum is “flat.” The term is analoguous to white light, which contains all
visible wavelengths.
Instrumentation and Data Preprocessing 63
References
Ahonen, A. I., Hämäläinen, M. S., Ilmoniemi, R. J., Kajola, M. J., Knuutila, J. E.,
Simola, J. T., et al. (1993a). Sampling theory for neuromagnetic detector arrays.
IEEE Trans Biomed Eng, 40(9), 859–869.
Ahonen, A. I., Hämäläinen, M. S., Kajola, M. J., Knuutila, J. E., Laine, P. P., Lounasmaa,
O. V., et al. (1993b). 122-Channel SQUID Instrument for Investigating the Magnetic
Signals from the Human Brain. Physica Scripta T49, 198–205.
Bardeen, J., Cooper, L. N., & Schrieffer, J. R. (1957). Theory of Superconductivity.
Phys Rev, 108, 1175–1204.
Berger, H. (1929). Uber das Elektroenkephalogramm des Menschen. Archiv für
Psychiatrie und Nervenkrankheiten, 87, 527–570.
Clarke, J. & Braginski, A. I. (2004). The SQUID Handbook: Fundamentals and
Technology of SQUIDs and SQUID Systems, (1st ed.) Berlin: Wiley-VCH.
Cohen, D. (1972). Magnetoencephalography: Detection of the Brain’s Electrical
Activity with a Superconducting Magnetometer. Science, 175, 664–666.
Curio, G. (2000). Linking 600-Hz “spikelike” EEG/MEG wavelets (“sigma-bursts”) to
cellular substrates: concepts and caveats. J Clin Neurophysiol, 17(4), 377–396.
De Tiège, X., Op de Beeck, M., Funke, M., Legros, B., Parkkonen, L., Goldman, S.,
et al. (2008). Recording epileptic activity with MEG in a light-weight magnetic
shield. Epilepsy Res, 82(2-3), 227–231.
Fischl, B., Sereno, M. I., Tootell, R. B., & Dale, A. M. (1999). High-resolution
intersubject averaging and a coordinate system for the cortical surface. Hum Brain
Mapp, 8(4), 272–284.
Hashimoto, I. (2000). High-frequency oscillations of somatosensory evoked potenti-
als and fields. J Clin Neurophysiol, 17(3), 309–320.
Hyvärinen, A. & Oja, E. (2000). Independent component analysis: algorithms and
applications. Neural Netw, 13(4-5), 411–430.
Hämäläinen, M., Hari, R., Ilmoniemi, R. J., Knuutila, J. & Lounasmaa, O. V. (1993).
Magnetoencephalography: theory, instrumentation, and applications to noninva-
sive studies of the working human brain. Rev Mod Phys, 65, 413–497.
Josephson, B. D. (1962). Possible New Effects in Superconductive Tunnelling. Phys
Lett, 1(7), 251–253.
Jousmäki, V., Nishitani, N., & Hari, R. (2007). A brush stimulator for functional brain
imaging. Clin Neurophysiol, 118(12), 2620–2624.
Kelhä, V. O., Pukki, J. M., Peltonen, R. S., Penttinen, A.J., Ilmoniemi, R. J., & Heino, J. J.
(1982). Design, Construction, and Performance of a Large-Volume Magnetic
Shield. IEEE Trans Magn, MAG-18(1), 260–270.
Mazziotta, J., Toga, A., Evans, A., Fox, P., Lancaster, J., Zilles, K., et al. (2001).
A probabilistic atlas and reference system for the human brain: International Con-
sortium for Brain Mapping (ICBM). Philos Trans R Soc Lond B Biol Sci, 356(1412),
1293–1322.
Okada, Y., Ikeda, I., Zhang, T., & Wang, Y. (2005). High-frequency signals (> 400 Hz):
a new window in electrophysiological analysis of the somatosensory system. Clin
EEG Neurosci, 36(4), 285–292.
Parkkonen, L., Simola, J. T, Tuoriniemi, J. T., & Ahonen, A. I. (1999). An Interference
Suppression System for Multichannel Magnetic Field Detector Arrays. In: Proc. 11th
Intl. Conf. on Biomagnetism, pp. 13–16. Sendai, Japan: Tohoku University Press.
64 MEG: An Introduction to Methods
Saunders, G. H., & Morgan, D. E. (2003). Impact on hearing aid targets of measuring
thresholds in dB HL versus dB SPL. Int J Audiol, 42(6), 319–326.
Seppä, H., Ahonen, A., Knuutila, J., Simola, J., & Vilkman, V. (1991). DC-SQUID
Electronics Based on Adaptive Positive Feedback: Experiments. IEEE Trans Magn,
27(2), 2488–2490.
Taulu, S., Kajola, M., & Simola, J. (2004). Suppression of interference and artifacts by
the Signal Space Separation Method. Brain Topogr, 16(4), 269-275.
Taulu, S., & Kajola, M. (2005). Presentation of electromagnetic multichannel data:
The signal space separation method. J Appl Phys, 97(12), 124905–124910.
Taulu, S., & Simola, J. (2006). Spatiotemporal signal space separation method for
rejecting nearby interference in MEG measurements. Phys Med Biol, 51(7), 1759–
1768.
Toga, A.W., Thompson, P. M., Mori, S., Amunts, K., & Zilles, K. (2006). Towards
multimodal atlases of the human brain. Nat Rev Neurosci 7(12), 952–966.
Uusitalo, M., & Ilmoniemi, R. (1997). Signal-space projection method for separating
MEG or EEG into components. Med Biol Eng Comput 35(2), 135–140.
Van Essen, D. C. (2005). A Population-Average, Landmark- and Surface-based (PALS)
atlas of human cerebral cortex. Neuroimage, 28(3), 635–662.
Vigário, R., Särelä, J., Jousmäki V., Hämäläinen, M., & Oja, E. (2000). Independent
component approach to the analysis of EEG and MEG recordings. IEEE Trans
Biomed Eng 47(5), 589–593.
Volegov, P., Matlachov, A., Mosher, J., Espy, M. A., & Kraus, R. H. (2004). Noisefree
magnetoencephalography recordings of brain function. Phys Med Biol, 49(10),
2117–2128.
Vrba, J., & Robinson, S. E. (2001). Signal processing in magnetoencephalography.
Methods, 25(2), 249–271.
Zimmerman, J.E., Thiene, P., & Harding, J. T. (1970). Design and Operation of Stable
rf-Biased Superconducting Point-Contact Quantum Devices and a Note on the
Properties of Perfectly Clean Metal Contacts. J Appl Phys, 41(4), 1572–1580.
3
Measurements
Lauri Parkkonen and Riitta Salmelin
Introduction
Running successful MEG measurements requires not only understanding the
principles of MEG but also mastering various practical points which impact
the quality of the acquired data. Investing time and effort in the MEG recordings
most often pays off; nothing facilitates data analysis more than a well-planned
experiment and good-quality data.
This chapter addresses the practical issues encountered when conducting
MEG measurements.
65
66 MEG: An Introduction to Methods
Quality Assurance
The MEG environment and the system may change over time. New interfer-
ence sources (magnetic or radio frequency), flux traps in the MEG sensors,
incorrect or suboptimal settings of the MEG system or stimulators etc. may
all reduce the amount of useful information in the MEG recordings. Equipment
may also malfunction. In the worst case, such degradations could hamper the
detection and localization of neural sources.
Quality assurance measures are thus recommended. A typical procedure
comprises (i) noise measurements without a subject — “empty room mea-
surements” — to spot changes in the magnetic environment or MEG sensors,
(ii) verification of the localization accuracy with the help of a phantom head
that includes current sources, and (iii) checking the proper operation of the
stimulus delivery system. Quality assurance checks can be done on a daily
or weekly basis depending on the usage of the system and the risk factors
involved.
Monitoring the data quality should eventually save time by allowing the
experimenter to concentrate on the particulars of the experiment instead of
the system and setup. Sometimes it may be difficult to tell whether data are
compromised because of interference from the subject or from the environ-
ment or the system itself; having a recent “fingerprint” of the signals without
the contribution from a subject helps to resolve these cases. Similarly, an
apparently incorrect source localization cannot necessarily be traced to any
particular device or procedural step if regular checks are not performed.
In addition to periodic quality assurance measurements, it is good practice,
prior to each MEG experiment and before bringing in the subject, to reserve
ample time to check that all the hardware and software components are fully
functional. A test run of the MEG data acquisition also provides an opportu-
nity to record “empty room” data (2–3 min) for reference. If multiple groups
share the MEG system, it is particularly important to check that the stimula-
tion devices and the possible software scripts, as well as response buttons or
microphones, if needed, are connected as required by the experiment and
function as expected.
Careful pre-tests improve the data quality as the subject’s time and
attention can be focussed on the experiment instead of debugging the mea-
surement setup.
Subject Preparation
Before starting the actual MEG recording and even prior to guiding the
subject into the magnetically shielded room, several preparatory steps are
necessary. A typical procedure is outlined below.
1. Consent. The subject receives a clear explanation of the experimental
procedures and reads and signs the consent form.
Measurements 67
7. Coils. Head position indicator coils are attached on the scalp. Using three
coils is the absolute minimum, but for improved accuracy and desirable
redundancy at least four coils should be employed, if that option is supported
by the MEG system. With four or more coils the system is able to cope with
situations where one coil detaches or moves prior to the MEG measurement.
The locations should be chosen such that the coils can be firmly attached to
the scalp, not to the hair, and that they are all covered by the MEG sensor
array once the subject is seated in the system. At the same time, they should
be as far apart as possible to ensure most stable and accurate co-registration.
When using four coils, two of them are typically placed behind the earlobes as
high up as possible, and the other two wide apart on the forehead, again as
high up as possible so that they can be properly covered by the MEG sensor
array.
There are various ways to attach the coils: they can be embedded in the
EEG cap, attached to the scalp with tape, or even glued to the skin with
collodium. Note that the coils are electrically isolated, and they should not be
in an electrical contact with the scalp.
8. Digitization. The locations of the indicator coils must be known with
respect to the anatomy for co-registration, e.g., with MRI, and the locations
of the scalp EEG electrodes for EEG source modelling. This information is
obtained by a digitizer device that records the coordinates of a stylus in the
3D space. The digitization procedure begins by identifying the anatomical
landmarks that span the head coordinate system (see Chapter 2). After that,
the locations of the head position indicator coils and EEG electrodes can be
digitized, and both expressed in the head coordinate frame. If no scalp EEG
electrodes are digitized, the head shape should be digitized to allow for a more
accurate and verifiable co-registration with anatomical MRIs. Some tens of
scalp points taken along contours from the tip of the nose to the back of the
head, and from one ear to the other, already help in obtaining a better match
with the MRIs. Some systems support a continuous digitization mode which
allows quickly collecting thousands of points on the head surface.
These steps may take just a few minutes, or even up to an hour if a
high-density EEG cap is applied.
Measurement
After the preparation steps described above, the subject can be guided in to
the shielded room. The actual measurement is typically preceded by the
following steps.
1. Seating. Head position indicator coils and EEG electrode wires and cap cables
should be connected to the MEG/EEG system. Stimulator devices (e.g., auditory
or somatosensory), if used, should be connected to the subject and tested.
Measurements 69
The subject should also try out the response buttons and microphones, if they
are needed in the experiment.
The setup should be made as comfortable as possible for the subject; if
the position is not comfortable, being immobile even for a short time easily
becomes painful, which distracts the subject. Tense neck muscles add wide-
band noise to the MEG, and thus deteriorate the quality of the recording.
Unlike in fMRI, in MEG subjects may see the movement of, e.g., the hand
when using response keys. Looking at such a movement evokes visual
responses that do not vanish in the averaging, as they are synchronized to the
trials. Such additional responses may unnecessarily complicate the data anal-
ysis. Similarly, response buttons emitting an audible click may give rise to
time-locked auditory responses, and movements that are mechanically trans-
mitted to other limbs may lead to additional somatosensory responses. Thus,
care should be taken to prevent or mask unwanted sensory input.
When recording children, or certain kinds of patients, it may be advisable
to have an assistant in the shielded room with the subject. The above demag-
netization guidelines apply also to the assistant. During data collection, the
assistant should remain immobile and as far from the MEG sensor helmet
as possible.
2. Instructions. The subject is reminded of the task and given specific instruc-
tions. In most MEG experiments, the subject should also be asked not to
move during the recording, to avoid eye movements, and — depending on
the experiment — to try to blink only during certain periods of the stimulus
sequence. In addition, the subject should be told how and when to communi-
cate with the experimenter.
3. Checking for artifacts. After the door of the shielded room has been closed
and the data acquisition system started, the MEG traces should be examined
visually for any artifacts. Movement-related and biological artifacts are the
most common; refer to the following section to identify artifacts. Requesting
the subject to take a couple of deep breaths helps in verifying that there is no
magnetic material on the subject’s person; see the next section.
4. Checking the EEG signals. Scalp EEG signals should be inspected visually.
Excessive noise or line frequency interference are likely signatures of bad elec-
trode contacts. With high-density electrode caps, it is often unavoidable that
some channels lose proper contact to the skin; aiming to have every single
EEG cap electrode working may not be the most productive approach. On the
other hand, the reference and isoground electrodes must be fully functional,
as losing them spoils the whole EEG recording.
The operation of the EOG channels can be verified by asking the subject
to blink a few times, move the gaze to left and right, followed by up and down
movements, while the experimenter is watching the EOG traces: blinks and
vertical movements should evoke clear signals, in excess of 200 microvolts, in
the vertical EOG channel whereas horizontal movement is mainly visible in
70 MEG: An Introduction to Methods
the horizontal EOG channel. A single diagonal EOG channel picks up all these
movements but with a lower signal-to-noise ratio than separate horizontal
and vertical channels.
Contracting the muscle monitored by EMG should give rise to bursting
high-frequency activity. In studies of cortico-muscular coherence where the sub-
ject has to maintain a steady contraction, it is a good idea to give feedback of the
proper level of contraction (clear bursting on the EMG channel) at this stage.
5. Head position measurement. The position of the subject’s head with
respect to the MEG sensor is determined by briefly energizing the head posi-
tion indicator coils. Depending on the MEG system, the coils are activated
either sequentially or simultaneously at distinct frequencies. The signals emit-
ted by these coils are captured by the MEG sensors. Based on these signals, the
coil locations are estimated in the MEG device coordinate system (see above).
With the help of the information from the digitization, the transformation
between the head and MEG device coordinate systems is calculated.
Just prior to the head position measurement, the subject should be asked
to take a comfortable position and remain as immobile as possible until the
end of the measurement block. If continuous head position tracking is
enabled, small movements are acceptable, however, the subject should still
avoid large head movements, as the associated motor activity and sensory
input may have an effect on the data.
When measuring a subject for the first time in MEG, it may be advisable
to perform a short test run prior to the actual measurement and re-adjust the
position when the subject is more relaxed.
6. Data collection. After a successful head position measurement, MEG data
collection can be started. The experimenter should monitor the raw MEG
data throughout the experiment; simple visual inspection allows judging
whether the noise level is acceptable. All deviations should be detected as early
as possible in order to avoid recording useless data.
In addition, on-line averaging should be employed and the accumulating
average monitored. Even when the data are to be re-averaged off-line, the
online average reliably shows the response amplitude. This direct measure of
signal-to-noise ratio allows the researcher to decide when to stop data collec-
tion, instead of always acquiring a fixed number of trials. It also helps in
detecting unwanted trial-locked responses due to, e.g., movement.
7. Breaks between blocks. Subjects can typically concentrate on a task no
more than 10 to 15 minutes continuously. Therefore, longer experiments are
often split in multiple blocks (see Chapter 4). Between the blocks, the subject
can rest and relax, blink and move the eyes freely. Head movement between
the blocks may also be allowed, if the blocks are analyzed separately or if the
researcher is prepared to apply specific post-processing and analysis methods
to the data. In any case, it is advisable to measure the head position at the
beginning of each block.
Measurements 71
Artifacts
MEG signals can, unfortunately, be hampered by unwanted signals from several
sources. Contaminating interference can be a result of strong ambient magnetic
fields or by sources — biological or artifactual — in the subject. The ambient
sources include interference from power lines and electromotors (at 50 or 60 Hz
and harmonics) and from large moving magnetic objects such as cars, elevators,
and even hospital beds. The signals from moving objects are characterized by
temporal scales similar to the movement; the frequency content is predominantly
below 1 Hz. Reduction of this type of interference is discussed in Chapter 2.
The experimenter is more often confronted by interference from the
subject rather than from the environment. Common biological sources dis-
turbing MEG include the cardiac muscle, skeletal muscles, and the eyeballs.
Each of these sources has a distinct temporal waveform; see Figure 3–1.
Cardiac artifact reflects either the pulsation or the magnetic counterpart of
the QRS complex of the electrocardiogram (Jousmäki and Hari, 1996). It is
characterized by relatively brief pulses occuring at the heart rate, and it appears
more prominently on magnetometers than gradiometers. Other muscles,
when they contract, emit continuous or bursting high-frequency noise, which
is naturally stronger the closer the muscle is to the sensor array. Both head
and eye movements are associated with shifts of the steady (DC) magnetic
field level seen by the MEG sensors; eye movements and blinks are manifested
as sub-second deflections on the frontal channels, whereas large head move-
ments evoke such changes on most channels.
Artifacts may also arise from nonbiological sources in or on the subject.
Dental work, fillings and braces, often give rise to large-amplitude magnetic
signals that amplify when the subject is moving the jaw. These signals may be
so strong even after demagnetization that the subject has to be excluded from
research studies. However, cleaning the data with post-processing methods
72 MEG: An Introduction to Methods
cardiac artefact
Saccades
Blinks
muscle activity
0 1 2 3 4 5 6
Time (s)
first hold the breath, and then to take a couple of deep breaths while the
experimenter is monitoring the MEG signals: if the slow variation first disap-
pears and then increases, the source is likely related to the small movements
associated with breathing. To trace the source further, the subject can be
asked to move one limb at a time. Most often the magnetic material is in the
clothing.
When the underlying neural sources are strong and exhibit simple field
patterns, so that they are easy to model, co-registration procedures are respon-
sible for most of the localization errors. When the experiments are performed
carefully, such sources can repeatedly be localized to a few millimeters.
Differing sensor-level responses from two recordings of the same subject
may be simply due to different head positions. This problem can be rectified
either by preprocessing the data with movement compensation methods
(Uutela et al., 2001; Taulu et al., 2005) to align the signals to a reference head
position, or by performing the comparison in the source space instead of
sensor space.
The overall replicability of MEG responses is good. Figure 3–2 shows the
results of the same auditory experiment performed multiple times on a single
subject. The locations of the N100m sources, reflecting a response from the
primary auditory cortex, are within a few millimeters of each other, and the
source waveforms differ only by amplitude, partly attributable to variations
in the stimulus intensity.
References
Jasper, H. H. (1958). The ten-twenty electrode system of the International Federation
of Electroencephalography and Clinical Neurophysiology. Electroencephalogr Clin
Neurophysiol, 10, 371–375.
Jousmäki, V., & Hari, R. (1996). Cardiac artifacts in magnetoencephalogram. J Clin
Neurophysiol 13(2), 172–176.
Niedermeyer, E., & Lopes da Silva, F. (2004). Electroencephalography: Basic Principles,
Clinical Applications, and Related Fields(5th ed.). Philadelphia: Lippincott Williams
& Wilkins.
Oostenveld, R., & Praamstra, P. (2001). The five percent electrode system for high-
resolution EEG and ERP measurements. Clin Neurophysiol, 112(4), 713–719.
Taulu, S., & Kajola, M. (2005). Presentation of electromagnetic multichannel data:
The signal space separation method. J Appl Phys 97(12), 124905–124910.
Uutela, K., Taulu, S., & Hämäläinen, M. (2001). Detecting and correcting for head
movements in neuromagnetic measurements. Neuroimage, 14(6), 1424–1431.
4
Experimental Design
Riitta Salmelin and Lauri Parkkonen
• Experimental setups that work well for fMRI are often suboptimal for
MEG, and vice versa
• Designs used in behavioral studies may serve as good starting points
for MEG experiments
• Interstimulus interval, stimulus duration, and number of trials per
experimental condition, all influence the neural response
Introduction
In order to design efficient functional neuroimaging experiments, one needs
to consider (i) the dynamics of the measured neural variable, and (ii) the rate
at which this variable can be sampled. The bulk of the MEG (and EEG) signal
reflects synchronous post-synaptic current flow in a large number of neu-
rons. These are very fast processes, in the millisecond range, and the signal
can also be sampled at a very high rate, up to several kHz, i.e., in submillisec-
ond range. Accordingly, the response can be tracked with good temporal
accuracy in each trial (Figure 4–1). In contrast, fMRI signal reflects changes in
the oxygen consumption, which lags the neural activation by 5–10 s and var-
ies slowly, over a period of several seconds. Whole-head fMRI images with a
reasonable spatial resolution can usually be collected every 2 seconds, at best,
75
Figure 4–1. Comparison of fMRI BOLD and MEG/EEG signals; schematic responses to stimuli of 200 ms, 2 s and 15 s in duration.
A prominent BOLD effect is obtained only with stimulation persisting for several seconds, whereas MEG/EEG evoked responses are pri-
marily elicited by stimulus onsets and offsets, with possibly a relatively weak sustained response. Cortical rhythms seen by MEG/EEG
(mainly alpha around 10 Hz and mu with 10- and 20-Hz components) may undergo suppression and rebound modulation, whose
time course is more comparable to that of BOLD. Due to the sluggishness of the BOLD signal, it is sufficient to sample it 0.5–2 times
per second; here, the dots denote sampling at a rate of 1/second (repetition time TR = 1 s). The MEG/EEG onset responses typically last
for less than a second and change orders of magnitude more rapidly than BOLD, thus necessitating sampling rates above 300 Hz.
Experimental Design 77
with the sampling rate limited by the image acquisition process. Because of
that, single trials can typically not be sampled with high enough temporal
resolution. The shape of the response can be estimated by combining multiple
single trials which were sampled at different times with respect to stimulus or
task onset (event-related design with jittered stimulus timing). The com-
monly applied analysis to a block design experiment effectively sums up the
fMRI signal over 20-30-s periods when the subject continuously receives sim-
ilar stimuli, or performs the same task—thus, the result carries no temporal
information on the responses.
Thus, a major difference between neurophysiological and hemodynamic
neuroimaging methods is that the signatures of neural activity recorded
in MEG (and EEG) are fast and they are oversampled in time whereas those
detected with fMRI and PET are slow and they are undersampled (see
Figure 4–1). This difference affects experimental design to a large extent and,
in particular, often renders good fMRI designs suboptimal for MEG and vice
versa. In fact, the designs used in behavioral studies, where one collects man-
ual or oral reaction times, tend to be better suited as starting points for MEG/
EEG experiments than those typically used in fMRI/PET studies.
Owing to its combined temporal and spatial sensitivity, MEG can identify
different processing stages as they unfold. Because of that, negligence in stim-
ulus preparation tends to immediately manifest in the neural signals. For
example, differences in the fade-in envelope of sound stimuli may cause much
larger effects in the early auditory responses than behaviorally highly relevant
differences in the stimulus content, e.g., speech vs. nonspeech sounds.
Therefore, the basic physical properties of sensory input need to be controlled
carefully (either matched or completely randomized) in order to extract
meaningful information about different levels of processing. For this same
reason, MEG studies do not readily accommodate an approach frequently
applied in fMRI/PET studies in which one assumes that, e.g., when the same
type of manual response has been given in two tasks, those tasks can be directly
contrasted and the movement effects so removed. MEG data may well reveal
that particularly the timing of the neural processes leading to the manual
response, and the sensorimotor activation itself, are influenced by the experi-
mental condition. This great advantage (or curse, depending on the situation)
of the MEG method has to be kept in mind when designing and piloting new
experiments.
Evoked Responses
Most MEG studies, so far, have focused on evoked responses, i.e., neural
activation that occurs at the same time, phase-locked, with respect to stimu-
lus or task onset (or offset) from trial to trial. Evoked responses are typically
detected within about 1 s from the stimulus presentation or execution of the
task. Evoked single-trial responses may be detectable in favorable conditions;
however, most often some tens or hundreds of trials are collected to yield an
average evoked response with a good signal-to-noise ratio. The earliest salient
responses, i.e., those of shortest latency, are typically transient (short-lasting)
and tightly locked to the stimulus, and thus yield sharp responses even when
averaged across multiple trials. The longer-latency responses tend to progres-
sively increase in duration and are likely to exhibit more jitter with respect to
the stimulus timing; in the average, they appear as sustained responses with
slow fade-in and fade-out phases. Figure 4–1a illustrates these dynamics in a
schematic way.
Evoked responses can be considered to reflect changes in the sensory
input: a long-lasting stimulus gives rise typically only to a transient MEG/
EEG evoked response at the onset and offset of the stimulus, whereas the
fMRI BOLD response may persist throughout the entire duration of the stim-
ulus, although delayed overall due to the sluggishness of the BOLD signal.
Moderately long stimuli may generate sustained components in the evoked
response.
limit the full length of the experiment. In the visual domain, the risk for
saccades (and problematic artifacts) is significantly reduced by presenting the
stimulus for less than 150–200 ms. If one needs to use stimuli or tasks of very
long duration, on the order of seconds, it may be worthwhile to consider
whether, instead of phase-locked evoked responses, a more suitable approach
might be spectrotemporal analysis on averaged event-related data, or on
blocks of non-averaged data, and design the experiment accordingly.
The interstimulus interval (ISI) should be long enough to allow the neural
responses to return to the base level (‘rest’) for at least 200 ms in the case of
evoked responses, or for 500–1000 ms in the case of event-related modulation
of rhythmic activity. Thus, when the focus is solely on evoked responses,
stimulus onset asynchrony (SOA) falls typically within 1–3 s, depending on
the amount of sustained activity the stimulus or task elicits, and on the speed
at which the subject can process the stimuli or perform the task. SOAs up to
5–10 s may be required for comprehensive tracking of event-related modula-
tion of cortical rhythms. The choice between fixed or variable ISI/SOA is at
the discretion of the experimenter. MEG analysis does not set any require-
ments in this regard, owing to the high time resolution. In MEG (or EEG)
experiments the effective ISI/SOA within each stimulus or task category often
varies considerably because the order of the trials belonging to the different
categories is usually randomized. Randomization is recommended, as the
vigilance of the subject thus varies, on average, similarly for all experimental
conditions and renders them more comparable. Nevertheless, should it be
required by the neuroscience question, a blocked design is obviously equally
feasible from the MEG point of view. Even in that case, randomized mini-
blocks may often be a better option than one extended block of each stimulus/
task type. Blocked design is obviously the choice when using continuous tasks
or stimuli, and the analysis is performed on non-averaged data. In that case,
it may be useful to introduce short rest periods also within blocks (and not
just longer periods between blocks) to serve as a baseline condition.
As an example, let us assume that we are interested in how the left and
right auditory cortex respond to 1–kHz tones, 50 ms in duration, presented to
the left or right ear or to both ears simultaneously. Since the stimuli are short
tone pips, cortical activation is concentrated to the first 100–200 ms. Thus,
the SOA could, in principle, be as short as 600 ms. However, the auditory
cortex responds more strongly with increasing SOA. A reasonable setup might
be as follows: 3 conditions (randomized) x 100 stimuli x 2–s SOA = 600 s = 10
min of effective recording time. This is a very short experiment, and one could
easily double the number of stimuli, and thus test for reproducibility of the
responses. Alternatively, one could enhance the neural responses by choosing
a longer SOA, or the SOAs could be randomized, e.g., between 2 and 6 sec-
onds to reduce expectation. However, one also needs to keep in mind that this
is an extremely boring experiment for the subject and, therefore, best results
are probably obtained by keeping the experiment as short as possible.
Experimental Design 81
fingers of the same hand, to a somewhat lesser degree. In some cases, verbal
responses may work better, as the readiness field tends to be markedly weaker
than for hand movements, and the activation is bilateral and quite similar for
stereotyped responses such as “yes/no.”
The excellent time resolution of MEG facilitates a fairly free choice of
parameters. What we have presented here are guidelines, rather than rules,
for experimental design. They can be adjusted at will, as long as the experi-
menter is aware of the possible effects for subject performance, neural
responses, and data analysis.
5
The Dowser in the Fields: Searching for
MEG Sources
Sylvain Baillet
Introduction
First: A Metaphoric Detour
Leaving the dowser behind right from the beginning of this chapter, there are
plenty of other ecological situations where finding sources of some observa-
tions is not obvious to begin with. Imagine that you go bow-fishing in clear
waters for the first time ever. This is a well-known situation in many books of
83
84 MEG: An Introduction to Methods
lessons about things: “How come the bird will catch the fish before I can?”.
The answer stays almost immutably the same: “That’s because the bird knows
about light refraction at the surface of water”. As unsatisfactory as this answer
might sound, we have to admit at least that the bird (and humans) can learn
from experience that the fish, being underwater, cannot be caught in the same
way as the mouse, running on the ground. This countryside metaphor does
not take us too far from our problem here. Indeed, the observations the bird
makes are series of distorted images from its prey. The fish here is the source
of the observations, which the bird implicitly combines with his model of the
influence of the source medium on the observation (light refraction), and a
collection of a priori information on the expected fish behavior in stressful
circumstances.
Maybe the bird catches the fish much faster than we would ever do, but
he surely does not know he is solving forward and inverse problems with
every dinner.
This book chapter has no intent to cover an extensive technical review of all
methodological tricks relevant to the MEG source-estimation problem.
86 MEG: An Introduction to Methods
An inverse problem is something we are all facing in our personal and some-
times professional lives. The bow-fishing metaphor was an attempt to dem-
onstrate this statement. The concept of inverse problem was formalized by
physicists in experimental science, where a model is confronted to some
observations. Models derive from theories and are supposed to let us make
predictions on natural phenomena and, more generally, on the outputs of a
system (e.g., the Earth’s climate, the stock exchange, or the brain). Hence, a
system is a very general concept used in many different situations – from biol-
ogy to economics. The main motivation here is to formalize how inputs I are
transferred to observable outputs O via some transformation T() – which
might include optional feedback loops and additional nuisances – so that:
O = T (I ). (5–1)
O = T (I , q ). (5–2)
where q represents the set of originally implicit parameters in T() (i.e. q = {nw}
in the context of light refraction). Predicting observations from a theoretical
model with a given set of parameters is called solving the forward modeling
problem.
The reciprocal situation, where observations are used to estimate the val-
ues of some parameters of the system, corresponds to the inverse modeling
problem.
Figure 5–2. Modeling illustrated: unknown brain activity – (a) top view –
generates variations of magnetic fields, and electric potentials, at the
surface of the scalp. This is illustrated by time series representing mea-
surements at each sensor lead – (a) from bottom to top. Modeling of
the sources and of the physics of MEG and EEG is illustrated in (b). As
naively represented at the top of (b) forward modeling consists of a sim-
plification of the complex geometry and electromagnetic properties of
head tissues. Source models are presented with colored arrows. Their free
parameters – e.g. location, orientation and amplitude – are adjusted dur-
ing the inverse modeling procedure to optimize some quantitative index.
This is illustrated here in (c) where the residuals – i.e., the absolute differ-
ence between the original data and the measures predicted by a source
model – are minimized.
sphere, with its center at some given coordinates. The way the inverse modeling
phase will be approached has many branches, as we shall discuss later.
conclude that while two solutions to the forward problem might exactly fit
the observations by adjusting the free parameters in the model (here by
changing values in I, e.g., the size and the location of the fish), this would lead
to multiple – and sometimes radically different—solutions to the inverse
problem (a strong bias in fish localization from the naive bow-fisher).
Therefore – and this is a general principle – whereas the forward problem has
a unique solution in classical physics (as dictated by the causality principle),
the inverse problem might accept multiple solutions, which are models that
equivalently predict the observations.
In MEG (and EEG), the situation is critical, as it has been demonstrated
theoretically by von Helmoltz back in the 19th century that the general inverse
problem that consists in finding the sources of electromagnetic fields outside
a volume conductor has an infinite number of solutions. This issue is not
specific to MEG: geophysicists are also confronted to non-uniqueness of
inverse models in trying to determine the distribution of mass inside a planet
by measuring the gravity field in the space outside the globe. Therefore, theo-
retically, an infinite number of source models would equivalently fit MEG
and EEG observations, which reduces their predictive power on the system’s
behavior to null. Fortunately, this question has been addressed with the math-
ematics of ill-posedness and inverse modeling, which formalize the necessity
of bringing additional contextual information to complement a basic theo-
retical model, as we shall discuss in the next section. This brings us to the
second point, where one has to bear in mind that the inverse problem is a
modeling problem. If we follow the Popperian paradigm that states that a
model is valid as long as it is not falsified by failure in predicting new experi-
mental evidence (Popper, 1959), both the bird and naive solutions to the fish
problem are valid until the fish needs to be caught, which consists in itself of
a new experiment that will eventually invalidate the naive model. Transposed
to the MEG world, the fish becomes, e.g., an epileptogenic locus supposedly
identified using MEG source modeling and potentially subjected to surgical
resection, with terrible consequences if the wrong model was not falsified
beforehand.
As we shall now see, these considerations have both philosophical and
technical impacts on approaching the general theory and the practice of
inverse problems (Tarantola, 2004). An important approach to address this
caveat consists in obtaining a measure of the uncertainty on the values of the
parameters sought by the inverse model. Situations like the aforementioned,
where a large set of values for some of the parameters produce models that
equivalently account for the observations, may raise a red flag to both ques-
tion the quality of the experimental data and, most importantly, falsify the
theoretical model.
Therefore, because uncertainty and modeling come in a pair, solutions to
the inverse problem should necessarily be complemented or even directly
built from probabilistic appraisal or statistical inference. This point will be
reviewed below.
90 MEG: An Introduction to Methods
We refer to the localization approach when the global source model consid-
ered in the MEG inverse problem states that the observations are produced by
the activity of brain areas, whose locations can be estimated from the data. In
this paradigm, each source in the global model accounts for the activity of a
brain region which is explicitly separated in space from other active regions in
the model.
Imaging approaches have been developed more recently, and were
inspired by the plethoric research in image restoration and reconstruction in
other domains (early digital imaging, geophysics, and other biomedical imag-
ing techniques). The corresponding global source model does not attempt to
estimate location parameters from observations, but rather aims at recover-
ing the distribution of all mass neural currents, at a scale compatible with the
electrophysiological basis of MEG signals (Chapter 1). The general outcome
of this approach is truly a stack of images – hence imaging – in the sense that
the estimated parameters are restricted to the intensities of distributed ele-
mentary models of mass neural activity that spatially sample the brain, just
like the pixels of an image sample a region of 2D space. Contrarily to the
localization model, there is no sense of source separation in the imaging
92 MEG: An Introduction to Methods
Figure 5–3. Inverse modeling: the localization (a) vs. imaging (b) approaches.
Source modeling through localization consists in decomposing the MEG
generators in a handful of elementary source contributions; the simplest
source model in this situation being the equivalent current dipole (ECD).
This is illustrated here from experimental data testing the somatotopic
organization of primary cortical representations of hand fingers (Meu-
nier et al., 2001). The parameters of the single ECD have been adjusted
on the [20, 40] ms time window following stimulus delivery. The ECD was
found to localize along the contralateral central sulcus, as revealed from
the 3D rendering obtained after the source location has been registered
to the individual anatomy. In the imaging approach, the source model
is spatially-distributed using a large number of ECD’s. Here, a surface
model of MEG generators was constrained to the individual brain surface
extracted from T1-weighted MR images. Elemental source amplitudes are
interpolated onto the cortex, which yields an image-like distribution of
the amplitude and spatial extension of cortical currents.
approach per se. Explicit identification of activity issued from distinct brain
regions necessitates complementary analysis beyond inverse modeling.
We will now discuss these approaches in the following subsections with a
focus on concepts rather than on techniques, which details can be found in
the cited references.
O = T (I , q )+ e . (5–3)
I
( 2
)
Iˆ = arg min O − T (I , q ) = arg min e LS
I
( 2
). (5–4)
Hence, tuning the model parameters so that they perfectly fit the data would
also result in explaining the remaining nuisance components; a general issue
known as overfitting the observations. The LS approach defines a criterion to
select the unknown parameters so that the resulting model is able to predict
the signal part in the observations, without the nuisance contribution. This
criterion assumes that ε is IID under normal distributions across sensors.
This condition is violated for many types of perturbations such as physiologi-
cal artifacts (e.g., cardiac and/or ocular), which produce highly correlated
field patterns across a subset of head surface sensors (see Figure 5–5). Beyond
artifact rejection and/or reduction techniques, careful preprocessing of the
data may therefore also include a so-called pre-whitening procedure, which
insures that the remaining perturbations are indeed IID (Kay, 1993).
From a statistical point of view, a LS fit of model parameters to data
attempts to reduce the sample variance of the residuals – that is, the part in
the observations left unexplained by the model – to some expected value; ide-
ally the one of e. In practice, this value may be estimated from the sample
statistics of the MEG signals preceding the stimulation, or ideally from an
acquisition run with the MEG recording only the environmental noise and/or
with the subject resting quietly under the helmet (Huizenga et al., 2002; Jun
et al., 2006).
O = T (q )I . (5–5)
q is the set of orientation and location parameters and I is the set of dipole
amplitudes. (5–5) translates that the physics of MEG predicts that a current
96 MEG: An Introduction to Methods
Iˆ = T (q ) O.
−1
(5–6)
If observations are from real – i.e., noisy – MEG data, then we have
O = T(θ)I + ε, which we replace in (5–6) to finally obtain:
Iˆ = T (q ) (T (q )I + e ),
−1
(5–7)
Iˆ = I + T (q ) e .
−1
0.8
20
0.6
40
0.4
MEG sensors
60 0.2
0
80
−0.2
100 −0.4
120 −0.6
−0.8
140
−1
20 40 60 80 100 120 140
MEG sensors
Figure 5–5. Matrix array of correlation coefficients across 147 MEG sensors
during a 900 ms baseline recording of ongoing brain activity. Red and
blue colors represent strong patterns of spatial correlation between MEG
measurements, indicating – as expected – that supposed brain noise is not
spatially independent.
The Dowser in the Fields: Searching for MEG Sources 97
overfitting the data by producing a model Î in (5–7) that would also fully
account for noise in the data (i.e. eLS = 0 in (5–4)), which is of course not
desirable in practice.
Therefore, trying to fit as many parameters as there are unknowns does
not lift the ill-posed nature or the modeling problem in real, noisy conditions.
Here, although a solution to the inverse problem exists and is unique, it is
highly dependent on the noise components in the data and ends up violating
the third Hadamard condition to well-posedness (i.e., continuous depen-
dency). Now let us suppose for the sake of further demonstration that the
data is idealistically clean from any disturbance. Remember, we obtained
the canonical inverse in (5–6) by estimating the amplitude parameters I after
the other source parameters (orientations and locations) q have been fixed.
Now, if all 300 source parameters {I, q} are left unconstrained, (5–5) is still
linear in I but not in q, as imposed by the forward model of the electromag-
netics of MEG. Here, there are as many unknowns as there are instantaneous
data, but the fact that some of them have a nonlinear dependency just makes
them more difficult to estimate than in the situation where all unknowns
linearly depend on the observations as in (5–6).
We need to solve the full LS optimization problem:
Energy
Parameter
A general rule of thumb when the data is noisy and the optimization principle
is ruled by nonlinear dependency is to keep the complexity of the estimation
as low as possible. Taming the complexity starts with bringing the number of
unknowns to be estimated under the number of observations, thereby mak-
ing (5–5) become overdetermined. Overdeterminacy should not be consid-
ered as big an issue in experimental sciences as underdeterminacy. Indeed,
and very pragmatically speaking, the additional information brought by sup-
plementary sensors just brings more information to the estimation problem,
and might even give room to channel selection in case of noise or artifact
contamination of a subset of sensors.
Early dipole fitters – as we have seen above – have naturally realized that
fitting a single dipole to early-latency somatosensory data would be about the
The Dowser in the Fields: Searching for MEG Sources 99
right way to start with computational source fitting (Okada et al., 1984). The
somatotopy of primary somesthesic brain regions using single dipole models
has been, and still is, yielding a flourishing literature, especially in clinical
investigations (e.g., Meunier et al., 2003). This single equivalent dipole model
(ECD) for supposedly isolated activations during the early stage of brain
information processing was also conjectured and evaluated in auditory
(Zimmerman et al., 1981) and visual (Lehmann et al., 1982) primary responses.
Very early on, though, it appeared that most later components of evoked
fields would necessitate more parameters to both 1) bring the LS-error εLS
down to a level compatible with SNR, and 2) yield a source model with rea-
sonable stability across successive observations on a time window compatible
with the waveforms measured at the sensor level.
Scherg and von Cramon (1985) conceptualized the spatiotemporal (ST)
dipole model, which also requires to solve a LS-optimization problem, but on
a set of successive time samples. The ST dipole model was typically developed
to localize the sources of scalp waveforms that were assumed to be generated
by multiple and overlapping brain activations. The model therefore includes
the a priori that a source in the inverse model is expected to be activated for a
certain duration – with amplitude modulations – while staying at the same
location with the same orientation. The ST model is typical of the introduc-
tion of further information in the tricky estimation problem confronting us.
Alternatively, the orientation constraint may be relaxed by considering
that tiny displacements of the brain activity can be efficiently modeled by a
rotating dipole source at some fixed location. This approach was extensively
adopted in studies of the tonotopic organization of the auditory cortex
(Zimmerman et al., 1981). In practice, a rotating dipole can be efficiently
modeled by a triplet of orthogonal dipoles, which form a basis for a dipole
with any orientation at the same location. Hence, the orientation parameters
in q can be reduced to the amplitude parameters of the dipoles in the equiva-
lent triplet, i.e., parameters with linear dependency on the data.
When multiple sources are expected from the experiment, multiple-
dipole models need to be fitted to the observations. As we have seen previously,
fitting the entire set of parameters with too many dipoles is detrimental to the
numerical stability and significance of the inverse model. Hence, the number
of dipoles to be adjusted must also be estimated from the observations, and
thereby constitutes some hidden parameter in the forward model.
Adding the number of sources to the LS error functional leads to intrac-
table optimization. Other criteria based on information theory may be applied
to automatically estimate the number of sources in the data, but have been of
limited success in MEG because the statistics of the data do not match the
methodological assumptions for which they were developed initially (Waldorp
et al., 2005). Signal classification and spatial filtering techniques are efficient
solutions to this problem as we shall see in the next section.
Let us conclude by stating that the localization approach, consisting in
adjusting the parameters of a limited number of equivalent dipole sources,
100 MEG: An Introduction to Methods
1988) are by far the most popular, as they offer a reasonable tradeoff between
the attenuating performances outside the pass-band and the degrees of free-
dom available from observations to build the filter coefficients.
Though beamformers are convenient in translating the source localiza-
tion problem to a signal detection issue in a search space, they suffer from
drawbacks that are important to bear in mind.
(1) They are built from a model of the covariance statistics of the data, which
may be estimated from the data through sample statistics. However, even
though a greater number of sensors is beneficial to the synoptic measure-
ment of the brain magnetic fields, the more channels, the more data sam-
ples are necessary for robust – and numerically stable – estimation of the
covariance statistics. This is the reason MEG beamformers have been
evaluated on sweeps of ongoing, unaveraged data (see Chapter 9) in
experimental conditions where behavioral stationarity was a means to
ensure some stationarity in the data as well. In the context of evoked,
time-locked activity, some recent developments suggest to consider sam-
ples across single trials to build the statistics (Cheyne et al., 2006).
(2) They are quite sensitive to errors in the head model. The filter outputs are
usually normalized by local noise contributions evaluated from some
baseline time window. However, SNR is not homogeneous everywhere in
the source space, which results in sidelobe leakages from interfering
sources nearby, which impedes the filter selectivity and, therefore, the
specificity of source detection (Wax & Anu, 1996).
(3) They are fooled by simultaneous highly correlated activations that are
interpreted by the beamformers as emerging from a single source, hence
identified with an uncontrolled location bias.
Signal processors had long identified these issues, and consequently
developed an alternative point of view on the data as being signal or noise, as
an alternative technique to beamformers. Multiple signal classification
(MUSIC) algorithm (Schmidt, 1986) starts by considering that the signal and
noise components within observations are uncorrelated. Strong results in sig-
nal subspace theory show that these components live in separate subspaces
which can be identified using, e.g., principal component analysis (PCA) of the
data time series (Golub, 1996).
Mosher et al. (1999) gives an extensive review of signal classification
approaches to MEG and EEG source localization. Their practice in various
experimental conditions remains limited by their sensitivity in the definition
of the signal vs. noise subspaces, which rules completely the subsequent clas-
sification and, therefore, the performances of source identification. Reasons
for this come from the quite limited instrumental noise in EEG and MEG
compared with the massive background brain activity which is very perturb-
ing, because it is structured in a way very comparable with the signal of
interest. An interesting application of MUSIC-like powerful discrimination
102 MEG: An Introduction to Methods
The imaging approach estimates the source amplitudes from the observations
while constraining the locations and orientations to the brain volume or
surface. In the volumic case, the brain is gridded using a 3D lattice of
voxels, which might be either generic – e.g., inferred from an MRI template –
or obtained directly from the subject’s individual MRI and confined to a
mask of the grey matter by using the appropriate software solution, e.g.,
SPM(http://www.fil.ion.ucl.ac.uk/spm/), brainVISA(http://brainvisa.info) or
BrainSuite(http://neuroimage.usc.edu/brainsuite/).
The cortically-constrained image model derives from the assumption
that MEG data originates essentially from large cortical assemblies of pyrami-
dal cells, with currents from post-synaptic potentials (PSP) flowing orthogo-
nally to the local cortical surface. The orientation constraint can either be
strict (Dale and Sereno, 1993) or relaxed by authorizing some controlled
deviation from the surface normal (Lin et al., 2006a).
The Dowser in the Fields: Searching for MEG Sources 103
(a) (b)
(c) (d)
source model – and the a priori probability of a given source model p(I)
following the Bayes rule that states.
p (I | O )∝ p (O | I ) p (I ). (5–9)
I
( 2
)
IˆRLS = arg min O − TI + lf (I ) = arg min e RLS .
I
(5–10)
Note that we have removed the nonlinear parameters q (e.g., the elemen-
tary source locations) from the optimization, as they are now all considered
as fixed and predetermined. T is now entirely defined as the solution to the
MEG forward problem for all elementary sources in the distributed model
with arbitrary unit current amplitudes, and is generally referred to as the gain
matrix; f(I) is typically a positive monotonic function of source amplitudes,
which takes large values when I deviates from the expected type of source
distribution a priori; λ is a positive scalar that helps balance the parameter
optimization between unregularized OLS prediction error e LS (l 1) and
The Dowser in the Fields: Searching for MEG Sources 105
Likeihood
Prior
Posterior
increasing λ
A Collection of Priors
A widely used prior in the field of image reconstruction considers that the
expected source amplitudes I be as small as possible on average. This is the
well-described minimum-norm (MN) model where:
f (I ) = I .
2
(5–11)
where T′ stands for the matrix transpose of T, and Id the identity matrix. The
computational simplicity of the MN estimate of distributed current ampli-
tudes has been very attractive in MEG (Wang et al., 1992), because it demon-
strated that a unique solution to the underdetermined MEG inverse problem
could be elaborated.
The basic MN estimate has been demonstrated to be problematic, though,
as it tends to favor the most superficial brain regions (e.g., gyral crowns) and
underestimate the contribution of deeper source areas (such as sulcal fundi)
(Fuchs et al., 1999) (Figure 5–9).
As a remedy, a slight alteration of the basic MN estimator consists in
weighting each elementary source amplitude by the inverse of the norm of its
contribution to sensors (i.e. of the corresponding column of the gain matrix
(T)). This depth weighting yields a weighted MN (WMN) estimate which still
benefits from unicity and linearity in the observations as the basic MN in
(5–12) (Lin et al., 2006b).
Despite their robustness to noise and simple computation, it is relevant to
question the neurophysiological validity of MN priors – though it would have
been more rigorous to do so beforehand. Indeed – though reasonably intuitive -
there is no evidence that neural currents would systematically match the
principle of minimal energy. Some authors have speculated that a more physi-
ologically relevant prior would be that the norm of spatial derivatives (e.g.,
gradient or Laplacian) of the current map be minimized (as in LORETA;
Pascual-Marqui et al., 1994). As a general rule of thumb, however, all MN-based
source imaging approaches greatly overestimate the smoothness of the spatial
distribution of neural currents, while quantitative and qualitative empirical
evidence demonstrate spatial discrimination of reasonable range at the sub-
lobar brain scale (Darvas et al., 2004; Sergent et al., 2005). Refer to Chapters 7
and 8 for specific discussions on distributed-source modeling.
In previous sections, we have taken for granted that the forward model of
MEG was solved prior to running into the questions of inverse modeling. We
shall speak briefly of state-of-the-art approaches to modeling the neural cur-
rents and the magnetostatics first, always from a pragmatic perspective, and
cite useful references for complete reading.
The solution to the forward problem in MEG (and, again, EEG) concerns the
choice of two models that are bound to work together very complementarily:
the source model and the prediction (modeling) of MEG/EEG surface
measurements produced by such a model.
The canonical source model we have been using the all way through so
far, is the electric current dipole. This model was initially motivated by the
dipolar pattern of magnetic fields observed outside the scalp. Further, the
depolarization process in neural cell assemblies is quite intuitively depicted as
a current flow between a source and a sink over a limited distance, which is
also well-accounted by an equivalent current dipole.
Indeed, most topographical patterns from evoked activity at the sensor
level are essentially made of dipolar shapes of field distributions, thereby rein-
forcing the fitness of such a simple model for current density distribution in
cell assemblies.
Some authors however, have questioned the predictive value of the dipo-
lar model – especially in the context of ECD fits for source localization – when
brain activity is susceptible to extending up to about 10cm2 to be detectable at
the scalp surface, as evidenced by electrocorticographic (ECoG) recordings
(Tao et al., 2005).
To understand this, we need to refer to the physics of magnetostatics,
which state that the magnetic fields produced by any arbitrary current distri-
bution may be decomposed as the sum of more elementary terms; the first
108 MEG: An Introduction to Methods
being the dipolar contribution of the source, and the rest being built from the
contributions of elementary multipolar current sources: this is called the
‘multipolar expansion’ of the field (Karp et al., 1980). When the source exten-
sion is small compared to its distance to sensors – e.g. from small (~< 1cm2)
shallow brain activations to deeper large source (~> 10cm2), the multipolar
contributions tend to be negligible compared with dipole’s. However, when
the extension of the local brain activity tends to increase, and/or when the
distribution of neural currents within an active areas departs from uniformity –
e.g., when local synchronization rate amongst cell assemblies is rather poor –
multipolar components tend to become prominent in the measured field
(Jerbi et al., 2002).
As a matter of fact, measures from simultaneous ECoG and EEG in epi-
lepsy have shown that on the order of 10cm2 active brain surface is sometimes
necessary, to be detected at the scalp level. Further, studies have shown that
the ECD model in localization approaches for such large surface of brain
activity, localize the source of the signals with a distance bias of few centime-
ters away from the patch centroid (Jerbi et al., 2004). Recently, federative
efforts for compact multipolar modeling of large brain areas have been
achieved (Jerbi et al., 2002) and triggered research on efficient localization
using source models from multipolar expansions up to order 4 (quadrupole
source model), which has shown very encouraging gain on localization bias
(Jerbi et al., 2004). In the near future, this may be applied for further compact
modeling in distributed source models and yield better-posed inverse model-
ing problems with more efficient numerical resolution.
Once a generic model of neural mass activity has been selected, the forward
problem is further resolved by describing the source environment – i.e., the
geometrical and the electromagnetic properties of head tissues – and the sensor
array, both having influence on the fields produced by the source model.
In EEG for instance, it is quite intuitive that the skull would form a
barrier of lower conductance that strongly distorts and attenuates the electric
potentials at the scalp vs. at the cortical surface. This is also the case for MEG
in an arbitrary head geometry, as the electric potentials and magnetic induc-
tion are coupled by the Maxwell’s equations (Hämäläinen et al., 1993).
Simple geometrical models of the head have been extensively investi-
gated; the most popular being concentric layers arranged in spheres, one
sphere per major category of tissue (scalp, skull and brain). The reasons for
this are both historical and pragmatic. From the historical point of view, the
sphere was – again – initially investigated in MCG, and demonstrated an
interesting trade-off between oversimplified models of body sources floating
in an homogeneous medium, and the computationally-demanding numeri-
cal solutions required by realistic geometry, which basic computers could not
handle routinely at the time.
From the pragmatic point of view, one can notice that most heads fit
reasonably well inside a sphere centered about 5 cm above the plane defined
by the usual anatomical fiducials used in MEG: the nasion, and both bilateral
pre-auricular points.
The spherical geometry has demonstrated very attractive properties in
MEG (Sarvas, 1987). Indeed, remarkably, MEG spherical solutions are abso-
lutely insensitive to the number of shells, such that a single homogeneous
sphere generates the same MEG fields as a set of concentric spheres of different
conductivities. Therefore, for MEG purposes, only the center of symmetry,
i.e., the common center of the concentric spheres, is important; the conduc-
tivity profile plays no role in the solution, nor do the radii of the spheres.
Implicitly, it is assumed that the radius of the outermost sphere is smaller
than the distance of any of the sensors from the center of symmetry. The
sphere can be fit to the entire head, or restricted to regions of interest, such as
parieto-occipital regions for visual studies.
Another remarkable consequence of the spherical symmetry is that the
radial component of the magnetic field is not affected by the volume currents,
and therefore only depends on the primary current source. However, this fact
is of minor practical importance, since all field components can be easily
computed from an analytical formula, which takes the contribution of the
volume currents correctly into account (Sarvas, 1987). More importantly,
however, radially oriented currents produce no magnetic field outside a
spherically symmetric conductor. Therefore, signals from currents at the
crests of the gyri and depth of the sulci are attenuated in the MEG data,
whereas the contribution of the former is very prominent in EEG (see
Figure 5.9, and Hillebrand & Barnes, 2002).
This has led researchers to investigate more realistic head geometries to
enhance the sensitivity of MEG toward pseudo-radial sources. Boundary
Element (BEM) and Finite Element (FEM) methods are generic numerical
approaches to the resolution of continuous equations over discrete space, and
have naturally been applied to MEG. Both approaches necessitate a geometric
The Dowser in the Fields: Searching for MEG Sources 111
(a) (b)
(c)
e.g., the skull – which is invisible on conventional T1-weighted MRI (but see
Dogdas et al., 2005) for an efficient estimation procedure).
The computational times also remain extremely long (several hours on a
conventional workstation) and are detrimental to source localization meth-
ods where the source location is optimized, which necessitates the update of
the corresponding forward model on the fly. Pragmatic solutions to this
problem have been proposed recently. These include the interpolative approx-
imation of the source lead-field from a dense precomputed volumic grid
(Ermer et al., 2001), or from an atlas-based approach, which deforms a pre-
computed FEM tessellation to match the individual scalp of the subject
(Darvas et al., 2006).
An alternative solution to these numerical approaches considers a set of
overlapping spheres to compute the individual lead-fields at each sensor in
the array (Huang et al., 1999) (Figure 5–11). In other words, a spherical head-
model is designed locally to compute the led-field at each individual sensor
location. If the tessellations of the three primary tissue layers are available, the
initial decision is which layer to use for the fitting; practical experience is that
the inner skull boundary is most useful for this fitting in the MEG case, since
the currents are most affected by the inner skull boundary, and because the
scalp layer in the inferior regions basically begins to flatten out as the scalp
joins the neck regions. Fitting local spheres to these regions tends to create
unrealistically large spheres, and fitting instead to the inner skull yields spheres
that are more naturally fit to the brain regions (See BrainStorm).
Finally, let us close this section by mentioning that any realistic head
model requires an estimation of the conductivity values of the tissues of the
encephalon. Though solutions for impedance tomography using MRI (Tuch
et al., 2001) and EEG (Goncalves et al., 2003) have been suggested, their
practical impact is yet to be matured before entering the daily practice of
(a) (b)
Figure 5–11. Side (a) and front (b) views of a selection locally-fitted spheres
to the individual anatomy of a subject. Each sphere is adjusted to a
given sensor location and the neighboring scalp surface to compute the
corresponding lead field.
The Dowser in the Fields: Searching for MEG Sources 113
MEG researchers. So far, conductivity values from ex-vivo studies are con-
ventionally integrated in all BEM and FEM models (Geddes & Baker, 1967).
Confidence Intervals
We have seen before that fitting dipoles to a time chunk of data may be quite
sensitive to their initial locations prior to the search. Similarly, imaging meth-
ods suggest that each brain location is active, potentially. It would be quite
relevant to understand the confidence the researcher may have in the ampli-
tudes suggested by the distributed source model.
In other words, we are now looking for error bars that would define a
confidence interval about the estimated values of a source model.
Signal processors have long developed a systematic approach to what
they have coined as ‘detection and estimation theories’ (Kay, 1993). The
general stake consists in understanding how certain one can be about the esti-
mated parameters of a model, given a model for the noise in the data. The
basic approach consists in considering the estimated parameters (e.g., source
location) as a random variable. The parametric approach to the estimation of
error bounds on the source parameters consists in estimating their bias and
variance.
Bias is an estimation of the distance between the true value and the expec-
tancy of the average of a parameter due to perturbations. The definition of
variance follows immediately. Cramer-Rao lower bounds (CRLB) on the
estimator’s variance can be explicitly computed using an analytical solution
to the forward model, and given a model for perturbations (e.g., with distri-
bution under a normal law). In a nutshell, the tighter the CRLB, the more
confident one can be about the estimated values. Mosher et al. (1993) have
investigated this approach using extensive Monte Carlo (MC) simulations,
which evidenced a resolution of a few millimeters in single dipole models.
These results were later confirmed by phantom studies (Baillet et al., 2001b;
Leahy et al., 1998).
114 MEG: An Introduction to Methods
Statistical Inference
Questions like: “How different is the dipole location between these two exper-
imental conditions?” and “Are source amplitudes larger in such condition
that in a control condition?” belong to statistical inference from experimental
data. The basic problem of interest here is hypothesis testing, which is sup-
posed to potentially invalidate a model under investigation. Here, the model
must be understood at a higher hierarchical level than when talking about,
e.g., a basic source model. It is supposed to address the neuroscience question
that has motivated data acquisition and some experimental design (Guilford
et al., 1978). Readers are also encouraged to refer to Chapter 10 in this book,
which extensively documents this question.
In the context of MEG, population samples that will support the infer-
ence are either trials or subjects. Testing can therefore be run at the individual
and group levels, respectively.
The Dowser in the Fields: Searching for MEG Sources 115
(a)
(b)
nonparametric statistical tests take the data as they are, and are robust to
departures from gaussianity.
In brief, hypothesis testing forms an assumption about the data that the
researcher is interested in questioning. This basic hypothesis is called the null
hypothesis, H0, and is traditionally formulated to translate no significant
finding in the data, e.g., ‘there are no differences in the source model in both
experimental conditions’. The statistical test will express the significance of
this hypothesis and evaluate the probability that the statistics in question
would be obtained just by chance. In other words, the data from both condi-
tions are interchangeable under the H0 hypothesis. This is literally what per-
mutation testing is doing. It computes the sample distribution of estimated
parameters under the null hypothesis, and verifies whether a statistics of the
original parameter estimates was likely to be generated by this law.
We shall now review quickly the principle of multiple hypothesis testing
from the same sample of measurements,which induces errors when multiple
parameters are being tested at once. This issue pertains to statistical inference
both at the individual and at the group levels. Samples will be therefore
formed of repetitions (trials) of the same experiment in the same subject, or
the results from the same experiment within a set of subjects, respectively.
This distinction is not crucial at this point. We shall, however, highlight
the issue of spatial normalization of the brain across a sample of subjects
either by applying normalization procedures (Ashburner & Friston, 1997) or
by the definition of a generic coordinate system onto the cortical surface
(Mangin et al., 2004; Fischl et al., 1999).
The outcome of a test will evaluate the probability p that the statistics
computed from the data samples will be issued from complete chance as
expressed by the null hypothesis. The researcher fixes a threshold on p, above
which H0 cannot be reasonably rejected, thereby corroborating H0. Tests are
designed to be computed once from the data sample so that the error – called
the type I error – consisting in accepting H0, while it is invalid, stays below the
predefined p-value.
If the same data sample is used several times for several tests, we multiply
the chances that we commit a type I error. This is particularly critical when
running tests in sensors or source amplitudes of an imaging model, as the
number of tests is on the order of 100 and even 10,000, respectively. In this
latter case, a 5% error over 10,000 tests is likely to generate 500 occurrences of
false positives by wrongly rejecting H0. This is obviously not desirable, and
this is the reason why this so-called family-wise error rate (FWER) should be
kept under control.
Parametric approaches to address this issue have been elaborated using
the theory of random fields, and have gained tremendous popularity through
the SPM software. These techniques have been extended to electromagnetic
The Dowser in the Fields: Searching for MEG Sources 117
source imaging, but are less robust to departure from normality than non-
parametric solutions. The FWER in nonparametric testing can be controlled
by using the statistics of the maximum over the entire source image, or topog-
raphy at the sensor level – possibly across time (Pantazis et al., 2005).
The emergence of statistical inference solutions adapted to MEG has
brought electromagnetic source localization and imaging to a considerable
degree of maturity, quite comparable to other neuroimaging techniques (see
Figure 5–13 for an example). Most software solutions now integrate sound
solutions to statistical inference for MEG and EEG data, and this is a field that
is still growing rapidly.
Figure 5–13. MEG statistical inference at the group level illustrated: Jerbi
et al. (2007) have revealed a cortical functional network involved in hand
movement coordination at low frequency (4Hz). The statistical group
inference first consisted on fitting, for each trial in the experiment, a
distributed source model constrained to the individual anatomy of each
of 14 subjects involved. The brain area with maximum coherent activa-
tion with instantaneous hand speed was identified. The traces at the top
illustrate excellent coherence in the [3,5]Hz range between these mea-
surements (hand speed in green and M1 motor activity in blue). Secondly,
the search for brain areas with activity in significant coherence with M1
revealed a larger distributed network of regions. All subjects were coreg-
istered to a brain surface template in Talairach normalized space, with the
corresponding activations interpolated onto the template surface. A non-
parametric t-test contrast was completed using permutations between
rest and task conditions (p < 0.01); adapted from (Jerbi et al., 2007).
118 MEG: An Introduction to Methods
Notes
1 To estimate retrospectively what was the distribution of temperature in a
medium for instance.
2 3 for location per se, 2 for orientation and 1 for amplitude.
References
Arthur, R. M., & Geselowitz, D. B. (1970). Effect of inhomogeneities on the apparent
location and magnitude of a cardiac current dipole source. IEEE Trans Biomed Eng,
17(2), 141–146.
Ashburner, J., & Friston, K. (1997). Multimodal image coregistration and partitio-
ning–a unified framework. Neuroimage, 6(3), 209–217.
Badia, A. E. (2004). Summary of some results on an EEG inverse problem. Neurol Clin
Neurophysiol, 2004, 102.
Baillet, S., Mosher, J.C., & Leahy, R.M. (2001a). Electromagnetic brain mapping. IEEE
Signal Processing Magazine, 18(6), 14–30.
Baillet, S., Riera, J. J., Marin, G., Mangin, J. F., Aubert, J., & Garnero, L. (2001b).
Evaluation of inverse methods and head models for EEG source localization using
a human skull phantom. Phys Med Biol, 46(1), 77–96.
Baryshnikov, B. V., Van Veen, B. D., & Wakai, R. T. (2004). Maximum likelihood
dipole fitting in spatially colored noise. Neurol Clin Neurophysiol, 2004, 53.
Biomag. (1987). The biomagnetic inverse problem. The biomagnetic inverse problem
conference proceedings. Milton Keynes, U.K., April 1986. In: Phys Med Biol, 32(1),
1–146.
Cheyne, D., Bakhtazad, L., & Gaetz, W. (2006). Spatiotemporal mapping of cortical
activity accompanying voluntary movements using an event-related beamforming
approach. Human Brain Mapping, 27(3), 213–229.
Cohen, D., & Hosaka, H. (1976). Part ii: magnetic field produced by a current dipole.
J Electrocardiol, 9(4), 409–417.
Cuffn, B. N., & Geselowitz, D. B. (1977). Computer model studies of the magneto-
cardiogram. Ann Biomed Eng, 5(2), 164–178.
Dale, A. M., & Sereno, M. I. (1993). Improved localization of cortical activity by com-
bining EEG and MEG with MRI cortical surface reconstruction: A linear approach.
Journal of Cognitive Neuroscience, 5, 162–176.
Darvas, F., Pantazis, D., Kucukaltun-Yildirim, E., & Leahy, R. M. (2004). Mapping
human brain function with MEG and EEG: methods and validation. Neuroimage,
23, S289–S299.
Darvas, F., Rautiainen, M., Pantazis, D., Baillet, S., Benali, H., Mosher, J. C., et al. (2005).
Investigations of dipole localization accuracy in MEG using the bootstrap. Neuro-
image, 25(2), 355–368.
120 MEG: An Introduction to Methods
Darvas, F., Ermer, J. J., Mosher, J. C., & Leahy, R. M. (2006). Generic head models for
atlas-based eeg source analysis. Hum Brain Mapp, 27(2), 129–143.
Daunizeau, J., Mattout, J., Clonda, D., Goulard, D., Benali, H., & Lina, J-M. (2006).
Bayesian spatio-temporal approach for eeg source reconstruction: conciliating ecd
and distributed models. IEEE Trans Biomed Eng, 53(3), 503–516.
David, O., Kiebel, S. J., Harrison, L. M., Mattout, J., Kilner, J. M., & Friston, K. J. (2006)
Dynamic causal modeling of evoked responses in EEG and MEG. Neuroimage, 30(4),
1255–1272.
Davison, A. C., & Hinkley, D. V. (1997). Bootstrap methods and their application. Cam-
bridge series on statistical and probabilistic mathematics. Cambridge: Cambridge
University Press.
Demoment, G. (1989). Image reconstruction and restoration: overview of common
estimation structures and problems. IEEE Transactions on Signal Processing, 37(12),
2024–2036.
Dogdas, V., Shattuck, D. W., & Leahy, R. M. (2005). Segmentation of skull and scalp in
3-d human MRI using mathematical morphology. Hum Brain Mapp, 26(4), 273–285.
Ermer, J. J., Mosher, J. C., Baillet, S., & Leahy, R. M. (2001). Rapidly recomputable
EEG forward models for realistic head shapes. Phys Med Biol, 46(4), 1265–1281.
Feynman, R. P. The Feynman Lectures on Physics. (Volume 2). Reading, Massachus-
etts: Addison-Wesley.
Fischl, B., Sereno, M. I., & Dale, A. M. (1999). Cortical surface-based analysis.
ii: Inflation, flattening, and a surface-based coordinate system. Neuroimage, 9(2),
195–207.
Fuchs, M., Wagner, M., Köhler, T., & Wischmann, H. A. (1999). Linear and nonlinear
current density reconstructions. J Clin Neurophysiol, 16(3), 267–295.
Geddes, L. A., & Baker, L. E. (1967). The specific resistance of biological material–a
compendium of data for the biomedical engineer and physiologist. Med Biol Eng,
5(3), 271–293.
Geselowitz, D. B. (1963). The concept of an equivalent cardiac generator. Biomed Sci
Instrum, 1, 325–330.
Geselowitz, D. B. (1964). Dipole theory in electrocardiography. Am J Cardiol, 14,
301–306.
Geselowitz, D. B. (1973). Electric and magnetic field of the heart. Annu Rev Biophys
Bioeng, 2, 37–64.
Golub, C. F., & van Loan, G. H. (1996). Matrix Computations. (3rd ed.). Baltimore:
Johns Hopkins University Press.
Goncalves, S. I., de Munck, J. C., Verbunt, J. P. A., Bijma, F., Heethaar, R. M., & Lopes
da Silva, F. (2003). In vivo measurement of the brain and skull resistivities using an
EIT-based method and realistic models for the head. IEEE Transactions On Biome-
dical Engineering, 50(6), 754–767.
Guilford, J. P., & Fruchter, B. (1978). Fundamental Statistics in Psychology and Educa-
tion. (6th ed.). New York: McGraw-Hill.
Hadamard, J. (1902). Sur les problemes aux derivees partielles et leur signification
physique. Princeton University Bulletin, pp. 49–52.
Hämäläinen, M., Hari, R., Ilmoniemi, R., Knuutila, J., and Lounasmaa, O. (1993).
Magnetoencephalography: Theory, instrumentation and applications to the non-
invasive study of human brain function. Rev Mod Phys, 65, 413–497.
Hillebrand, A., & Barnes, G. R. (2002). A quantitative assessment of the sensitivity of
whole-head MEG to activity in the adult human cortex. NeuroImage, 16, 638–50.
The Dowser in the Fields: Searching for MEG Sources 121
Hillebrand, A. Singh, K. D., Holliday, I. E., Furlong, P. L., & Barnes, G. R. (2005).
A new approach to neuroimaging with magnetoencephalography. Hum Brain
Mapp, 25(2), 199–211.
Hosaka, H., & Cohen, D. (1976). Part iv: visual determination of generators of the
magnetocardiogram. J Electrocardiol, 9(4), 426–432.
Huang, M. X., Mosher, J. C., & Leahy, R. M. (1999). A sensor-weighted overlapping-
sphere head model and exhaustive head model comparison for meg. Phys Med Biol,
44(2), 423–440.
Huizenga, H. H., de Munck, J. C., Waldorp, L. J., & Grasman, R. P. (2002). Spatiotem-
poral eeg/meg source analysis based on a parametric noise covariance model. IEEE
Trans Biomed Eng, 49(6), 533–539.
Jerbi, K., Mosher, J. C., Baillet, S., & Leahy, R. M. (2002). On MEG forward modelling
using multipolar expansions. Phys Med Biol, 47(4), 523–555.
Jerbi, K., Baillet, S., Mosher, J. C., Nolte, G., Garnero, L., & Leahy, R. M. (2004). Loca-
lization of realistic cortical activity in MEG using current multipoles. Neuroimage,
22(2), 779–793.
Jerbi, K., Lachaux, J.P., NDiaye, K., Pantazis, D., Leahy, R.M., Garnero, L., et al.
(2007). Coherent neural representation of hand speed in humans revealed by MEG
imaging. Proc Natl Acad Sci USA, 104(18), 7676–7681.
Jun, S. C., Plis, S. M., Ranken, D. M., & Schmidt, D. M. (2006). Spa tiotemporal noise
covariance estimation from limited empirical magnetoencephalographic data. Phys
Med Biol, 51(21), 5549–5564.
Karp, P. J., Katila, T. E., Saarinen, M., Siltanen, P., & Varpula, T. T. (1980). The nor-
mal human magnetocardiogram. ii. a multipole analysis. Circ Res, 47(7), 117–130.
Kay, S. M. (1993). Fundamentals of Statistical Signal Processing: Estimation Theory.
Englewood Cliffs, NJ: Prentice Hall.
Kiebel, S. J., Tallon-Baudry, C., & Friston, K. J. (2005). Parametric analysis of oscilla-
tory activity as measured with eeg/meg. Hum Brain Mapp, 26(3), 170–177.
Kooi, K. A., Holland-Moritz, E. K., & Marshall, R. E. (1969). Oscillatory motion of a
dipole as a model for the generation of cerebral rhythms having asymmetrical wave
forms including the “14 and 6/sec” pattern. Electroencephalogr Clin Neurophysiol,
26(1), 116.
RLeahy, M., Mosher, J. C., Spencer, M. E., Huang, M. X., & Lewine, J. D. (1998).
A study of dipole localization accuracy for MEG and EEG using a human skull
phantom. Electroencephalogr Clin Neurophysiol, 107(2), 159–173.
Lehmann, D., Darcey, T. M. & Skrandies, W. (1982). Intracerebral and scalp fields
evoked by hemiretinal checkerboard reversal, and modeling of their dipole genera-
tors. Adv Neurol, 32, 41–48.
Lin, F. H., Belliveau, J. W., Dale, A. M., & Hämäläinen, M. S. (2006a). Distributed
current estimates using cortical orientation constraints. Human brain mapping,
27(1), 1–13.
Lin, F. H., Witzel, T., Ahlfors, S. P., Stuffebeam, S. M., Belliveau, J. W., & Hämäläinen,
M. S. (2006b). Assessing and improving the spatial accuracy in meg source
localization by depth-weighted minimum-norm estimates. Neuroimage, 31(1),
160–171.
Mangin, J-F., Rivière, D., Coulon, O., Poupon, C., Cachia, A., Cointepas, Y., Poline,
J-B., et al. (2004). Coordinate-based versus structural approaches to brain image
analysis. Artif Intell Med, 30(2), 177–197.
122 MEG: An Introduction to Methods
Mattout, J., Phillips, C., Penny, W. D., Rugg, M. D., & Friston, K. J. (2006). Meg
source localization under multiple constraints: an extended bayesian framework.
Neuroimage, 30(3), 753–767.
McIntosh, A. R. & Lobaugh, N. J. (2004). Partial least squares analysis of neuroima-
ging data: applications and advances. Neuroimage, 23 Suppl 1, S250–S263.
Meunier, S., Garnero, L., Ducorps, A., Mazières, L., Lehéricy, S., du Montcel, S. T.,
et al. (2001). Human brain mapping in dystonia reveals both endophenotypic traits
and adaptive reorganization. Ann Neurol, 50(4), 521–527.
Meunier, S., Lehéricy, S., Garnero, L., & Vidailhet, M. (2003). Dystonia: lessons from
brain mapping. Neuroscientist, 9(1), 76–81.
Miller, W. T., & Geselowitz, D. B. (1974). Use of electric and magnetic data to obtain
a multiple dipole inverse cardiac generator: a spherical model study. Ann Biomed
Eng, 2(4), 343–360.
Mosher, J. C., Spencer, M. E., Leahy, R. M., & Lewis, P. S. (1993). Error bounds for
eeg and meg dipole source localization. Electroencephalogr Clin Neurophysiol, 86(5),
303–321.
Mosher, J. C., Baillet, S., & Leahy, R. M. (1999). EEG source localization and ima-
ging using multiple signal classification approaches. J Clin Neurophysiol, 16(3),
225–238.
Mosher, J.C., Baillet, S., Leahy, & R.M. (2003). Equivalence of linear approaches in
bioelectromagnetic inverse solutions. In: Proceedings of the 2003 IEEE Workshop on
Statistical Signal Processing, pp. 294–7.
Nichols, T. E., & Holmes, A. P. (2002). Nonparametric permutation tests for functio-
nal neuroimaging: a primer with examples. Hum Brain Mapp, 15(1), 1–25.
Nolte, G., & Curio, G. (1999). The effect of artifact rejection by signal-space pro-
jection on source localization accuracy in meg measurements. IEEE Trans Biomed
Eng, 46(4), 400–408.
Okada, Y. C., Tanenbaum, R., Williamson, S. J., & Kaufman, L. (1984). Somatoto-
pic organization of the human somatosensory cortex revealed by neuromagnetic
measurements. Exp Brain Res, 56(2), 197–205.
Ossadtchi, A., Baillet, S., Mosher, J. C., Thyerlei, D., Sutherling, W., & Leahy,
R. M. (2004). Automated interictal spike detection and source localization in
magnetoencephalography using independent components analysis and spatio-
temporal clustering. Clin Neurophysiol, 115(3), 508–522.
Pantazis, D., Nichols, T. E., Baillet, S., & Leahy, R. M. (2005). A comparison of ran-
dom field theory and permutation methods for the statistical analysis of MEG data.
Neuroimage, 25(2), 383–394.
Pascual-Marqui, R. D., Michel, C. M., & Lehmann, D. (1994). Low resolution electro-
magnetic tomography: a new method for localizing electrical activity in the brain.
Int J Psychophysiol, 18(1), 49–65.
Penfield, W. G., & Boldrey, E. (1937). Somatic motor and sensory representation in
the cerebral cortex of man as studied by electrical stimulation. Brain, 60, 389–443.
Perrin, F., Pernier, J., Bertrand, O., Giard, M. H., & Echallier, J. F. (1987). Mapping
of scalp potentials by surface spline interpolation. Electroencephalogr Clin Neuro-
physiol, 66(1), 75–81.
Petit-Dutaillis, D., Fischgold, H., & Rougerie, J. (1952). Progress made at the Neuro-
surgical Clinic of the Hôpital de la pitié from 1949 to 1952 in the localization and
diagnosis of cerebral tumors by electroencephalography and arteriography. Bull
Acad Natl Med, 136(24-26), 444–448.
Popper, K. R. (1959) The Logic of Scientific Discovery. New York: Basic Books.
The Dowser in the Fields: Searching for MEG Sources 123
Press, W. H., Flannery, B. P., Teukolsky, S. A., & Vetterling, W. T. (1986). Numerical
Recipes: The Art of Scientific Computing. Cambridge: Cambridge University Press.
Sarvas, J. (1987). Basic mathematical and electromagnetic concepts of the biomagne-
tic inverse problem. Phys Med Biol, 32(1), 11–22.
Scherg, M. & von Cramon, D. (1985). Two bilateral sources of the late aep as iden-
tified by a spatio-temporal dipole model. Electroencephalogr Clin Neurophysiol,
62(1), 32–44.
Schmidt, R. O. (1986). Multiple emitter location and signal parameter estimation.
IEEE Transactions on Antennas and Propagation, 34, 276–280.
Schwartz, D., Poiseau, E., Lemoine, D., & Barillot, C. (1996). Registration of MEG/EEG
data with MRI: Methodology and precision issues. Brain Topography, 9, 101–116.
Sergent, C., Baillet, S., & Dehaene, S. (2005). Timing of the brain events underlying
access to consciousness during the attentional blink. Nature Neuroscience, 8(10),
1391–1400.
Spencer, M.E., Leahy, R.M., Mosher, J.C., & Lewis, P.S. (1992). Adaptive filters for
monitoring localized brain activity from surface potential time series. In: IEEE, ed.,
Conference Record of The Twenty-Sixth Asilomar Conference on Signals, Systems and
Computers, Vol. 1, pp. 156–161.
Tao, J. X., Ray, A., Hawes-Ebersole, S., & Ebersole, J. S. (2005). In tracranial eeg
substrates of scalp eeg interictal spikes. Epilepsia, 46(5), 669–676.
Tarantola, A. (2004). Inverse Problem Theory and Methods for Model Parameter
Estimation. Philadelphia: SIAM Books.
Tarantola, A. (2006). Popper, Bayes and the inverse problem. Nat Phys, 2(8), 92–494.
Tikhonov, A., & Arsenin, V. (1977). Solutions of Ill-Posed Problems. Washington, D.C.:
Winston & Sons.
Tuch, D. S., Wedeen, V. J., Dale, A. M., George, J. S., & Belliveau, J. W. (2001).
Conductivity tensor mapping of the human brain using diffusion tensor MRI.
Proc Natl Acad Sci USA, 98(20), 11697–11701.
Turner, R., Howseman, A., Rees, G. E., Josephs, O., & Friston, K. (1998). Functional
magnetic resonance imaging of the human brain: Data acquisition and analysis.
Exp Brain Res, 1-2(123), 5–12.
U.S. National Library of Medicine. The pubmed online database. Retrieved from:
http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=PubMed.
Uutela, K. Taulu, S., & Hämäläinen, M. (2001). Detecting and correcting for head
movements in neuromagnetic measurements. Neuroimage, 14(6), 1424–1431.
van Veen, D., & Buckley, K. M. (1988). Beamforming: a versatile approach to spatial
filtering. ASSP Magazine, IEEE Signal Processing Magazine, 5, 4–24.
Waldorp, L. J., Huizenga, H. M., Nehorai, A., Grasman, R. P., & Molenaar, P. C. M.
(2005). Model selection in spatio-temporal electromagnetic source analysis. IEEE
Trans Biomed Eng, 52(3), 414–420.
Wang, J. Z., Williamson, S. J., & Kaufman, L. (1992). Magnetic source images deter-
mined by a lead-field analysis: the unique minimum-norm least-squares estimation.
IEEE Trans Biomed Eng, 39(7), 665–675.
Wax, M., & Y. Anu, Y. (1996). Performance analysis of the minimum variance beam-
former. IEEE Transactions on Signal Processing, 44, 928–937.
Wood, C. C., Cohen, D., Cuffn, B. N., Yarita, M., & Allison, T. (1985). Electrical
sources in human somatosensory cortex: identification by combined magnetic and
potential recordings. Science, 227(4690), 1051–1053.
Zimmerman, J. T., Reite, M., & Zimmerman, J. E. (1981). Magnetic auditory evoked
fields: dipole orientation. Electroencephalogr Clin Neurophysiol, 52, 151–156.
6
Multi-Dipole Modeling in MEG
Riitta Salmelin
• MEG data are usually not ambiguous; it is mostly obvious where the
active areas are located
• Diligence in identifying clear dipolar field patterns yields well-behaved
models
• Parametric variation of stimuli is essential for functional localization
and helpful in source modeling
• The proposed solution should be obvious also in the original sensor
signals – always check
Introduction
The types of information one would typically like to extract from MEG data
are approximate location of active brain areas, time course of activation in
those areas, and effect of task or stimulus category on the neural timing and/
or strength of activation. The Equivalent Current Dipole (ECD) model is an
excellent tool in this endeavor. It is simple, well defined, and robust, as it
requires the minimum number of assumptions and, importantly, these
assumptions are entirely transparent to the user.
As an example of a neuroscience question with multiple experimental
conditions, let us consider silent reading of words and nonwords. Figure 6–1
displays MEG responses in one subject to short and long words, and nonwords.
124
Multi-Dipole Modeling in MEG 125
Figure 6–1. MEG signals to short and long words and nonwords, in one
subject. The measurement helmet is viewed from above, flattened onto
the plane, with the nose pointing upwards. In this Vectorview™ system,
sensors are arranged in 102 locations along the helmet. In each location,
there are two planar gradiometers that are most sensitive to orthogo-
nal orientations of current flow (see schematic heads in the upper right
corner) and one magnetometer (data not shown). Time runs along the
horizontal axis, from 200 ms before stimulus onset to 800 ms after it.
Variation of magnetic field is shown on the vertical axis.
The displayed time interval extends from 200 ms before stimulus presentation
to 800 ms after it. Each curve shows the signal recorded by one MEG sensor.
This particular MEG system (Vectorview™, Elekta-Neuromag Ltd.) uses two
orthogonally oriented planar gradiometers at each recording site, which
detect the maximum signal directly above an active cortical area. There are
deflections in the curves essentially everywhere over the helmet, reflecting
multiple active areas in the brain. In a number of sensors, the responses vary
by stimulus type.
By using ECDs one can identify the multiple areas that generate these
signals and their time courses of activation. In this experiment, there were
four experimental conditions and eight subjects. The analysis proceeds
through a number of steps, and we will start with single-subject analysis. First,
we focus on each experimental condition separately, and construct a source
model. Then, in order to compare the stimuli, we form a single, combined
source model that works for all conditions in this subject. Based on this com-
mon model, we can identify cortical areas and time windows that show sig-
nificant differences between conditions. Then we move onto group analysis.
We first group the individual source areas according to function, location,
direction of current flow, and/or timing, and then test for significant effects at
the group level.
126 MEG: An Introduction to Methods
This is a straightforward but not trivial task. In this chapter we will begin
with simpler paradigms in order to illustrate the basics of ECD source model-
ing, then work our way to increasingly complex data sets, and eventually
return to the question of reading words and nonwords. In addition, we will
touch upon the issue of mouth movements in an MEG experiment. Finally,
we will briefly discuss the use of ECD modeling in the analysis of cortical
rhythms and their task-related modulation.
Figure 6–3. Distribution of magnetic field over the course of the auditory
response. The red area indicates magnetic field emerging from the brain,
and the blue area the reentering field. Electric current flows in the mid-
dle, along the zero line (black curve).
As the fields are clearly dipolar, it seems reasonable to model the underly-
ing sources with Equivalent Current Dipoles that represent the center of an
active cortical patch, and the direction and magnitude of electric current
therein (Hämäläinen et al., 1993). By scanning through the N100m response,
we find a time point at which the field is as closely dipolar as possible
(Figure 6–4a). The density curves should then be fairly symmetrical, and the
line connecting the maxima should be perpendicular to the zero field line,
indicated by the black curve. The center of the dipole should fall on the cross-
ing of these two lines. In this way, one should already have a feeling of what to
expect before computing the ECD solution.
To find the source of the left N100m response, we select a subset of sensors,
covering the local field maxima (Figure 6–4b, solid line). The selection is also
depicted on the helmet as light gray squares. The source is located in the lower
lip of the Sylvian fissure, with the current flowing downwards, away from the
cortical surface (Figure 6–4c). As there is only one clear field pattern in the left
hemisphere, the exact selection of MEG sensors is not critical.
A model using this one source, plotted in purple in Figure 6–4b, allows
comparison with the original signals, which are plotted in black. This source
accounts adequately for the N100m in the left hemisphere but less so for the
later component, which had a clearly different field pattern (cf. Figure 6–3).
The right-hemisphere N100m remains unexplained, as it should. One can
determine the right-hemisphere N100m source in the same way, by finding a
clear dipolar field pattern and using another subset of sensors, as shown by
the dashed line in Figure 6–4b.
If we only use the left-hemisphere N100m source to account for the data
recorded by all MEG sensors, we will find that the waveform depicted in
Figure 6–5a describes the time course of activation in the left auditory cortex.
The overall level of explanation for the whole helmet, the goodness-of-fit
value, remains quite low, as one would expect. When we only include the
right-hemisphere N100m source (Figure 6–5b), the result is very similar, with
a clear time course of activation locally but poor explanation overall. When
we include both left and right N100m sources (Figure 6–5c), the explanation
rises dramatically and exceeds 80% around the peak of activation. Importantly,
Figure 6–5. Source analysis of the N100m response in one subject. (a) Time
course of activation in the left auditory cortex (source waveform) and
goodness-of-fit over the whole helmet when only one ECD in the left
auditory cortex was included. (b) Time course of activation in the right
auditory cortex when only one ECD in the right auditory cortex was
included. (c) Time courses of activation in the bilateral auditory cortex
when two ECDs were included, one in each hemisphere.
Multi-Dipole Modeling in MEG 129
Somatosensory Responses
Let us now move on to a slightly more complex data set. Figure 6–7 displays
responses to right median nerve stimulation (e.g., Forss et al., 1994; Schnitzler
et al., 1995). Again, the displayed time interval is 300 ms. Obviously, there is
plenty of activity, with different time behaviors: early components are
observed medially over the left central sulcus, and later components medially
in the left hemisphere and more laterally in both hemispheres.
Also in this case, by scanning through the field patterns one can recog-
nize the different source areas at distinct time points (Figure 6–8). At about
20 ms after stimulus onset, there is a dipolar field pattern forming over the left
hand area in the central sulcus. The picture becomes even clearer at about
30 ms, but the current now flows in the opposite direction. Until about 50 ms,
the field pattern remains fairly unchanged. At about 80 ms, the left parietal
cortex produces a pronounced dipolar field pattern. At around 100 ms, it is
accompanied by activations laterally in the left and right hemisphere. The
location is very similar to that of the auditory responses, but the opposite
direction of current flow now indicates activation in the upper lip of the
Sylvian fissure.
When several source areas are active rather simultaneously, one has to
put some effort into selecting the subsets of sensors used for source localiza-
tion. For the early sources in the upper row of Figure 6–8, almost any selec-
tion will do, as there is no other simultaneous activation. For the parietal
source, one should exclude the left frontal sensors, which detect the lateral
activation. Conversely, for the left lateral source, the sensor selection should
avoid the parietal region. As in the auditory responses, the question here is
Figure 6–8. Magnetic field patterns at different times after right median
nerve stimulation. The arrows represent the ECDs that best account for
each field pattern.
not where the active areas are, but how to find such unequivocal field patterns
that it is possible to reliably localize the sources. It is important to find the
time point at which each field pattern is optimally dipolar. It is worth noting
that this optimal time point may, or may not, coincide with a peak in the
signal.
In Figure 6–9, the sources are shown on the subject’s MR images. The
20-ms and 30-ms responses are generated by slightly different sources in the
hand area. However, they are so similar in location and orientation that both
cannot be included in the multidipole model because they interact too much.
Therefore, the strong 30-ms source in the primary somatosensory cortex, SI,
Figure 6–9. Final source model for right median nerve stimulation in the
same subject for whom original MEG data was displayed in Figures 6–7
and 6–8. Left: Locations and directions of the ECDs displayed on the sub-
ject’s MRI. Right: Time courses of activation in the brain. SI = primary
somatosensory cortex, PPC = posterior parietal cortex, SII = second soma-
tosensory cortex contralateral (c) and ipsilateral (i) to stimulation. Because
of the close similarity of the ECDs determined at 20 ms and 30 ms, they
are both represented in the SI waveform (negative deflection at 20 ms,
positive deflection at 30 ms).
132 MEG: An Introduction to Methods
represents the hand area activation in the model. In the source waveform, the
peak at 30 ms is preceded by a small negative deflection at 20 ms. Note that
this is not deactivation but simply reflects the opposite direction of current
flow in almost the same cortical location. Usually, one would be concerned
about a waveform that shows both positive and negative values, but here we
know that this is a real effect.
The other sources are located in the posterior parietal cortex (PPC) and
in the left and right second somatosensory cortices in the upper lip of the
Sylvian fissure (SII). We now have a sequence from SI to posterior parietal,
and further to the ipsi- and contralateral SII cortices, with all activations
partly overlapping in time. Again, one may check for possible interactions
between sources by leaving out one source at a time and observing whether it
affects the other waveforms.
For this high-quality data set, other analysis approaches will provide
essentially the same sequence of activation. The active areas can also be visu-
alized using so-called distributed models, which produce probability maps of
current distribution (e.g., Dale et al., 2000; Lin et al., 2004; Uutela et al., 1999).
Figure 6–10 displays Minimum Current Estimate (MCE; Uutela et al., 1999)
of this data set. From 20–50 ms, activity is concentrated to the SI cortex.
At about 80 ms, both the posterior parietal cortex and the contralateral
SII show activation. At about 100 ms, the SII activations have reached their
maximum in both hemispheres.
It is important to realize that the focal ECDs and the distributed proba-
bility maps produce exactly the same electromagnetic field outside of the
head, so both are equally correct. The appearance of the result is determined
by the choice of analysis method (model), not by the structure of active areas
in the brain. The bottom line is that MEG (or EEG) gives an estimate of
the center of an active area but—at least in typical experimental setups and
signal-to-noise ratios—no direct information about its spatial extent.
‘soijinto’). For real words, the lexical route should dominate, and length
should have little effect. Processing of nonwords, on the other hand, would
rely on the letter-level grapheme-to-phoneme conversion, and letter-string
length should have a strong effect. There were 100 stimuli of each type, shown
in a randomized sequence. The letter-strings were presented for 400 ms, and
each stimulus was followed by a blank interval of 2.6 s. The subjects were
reading the letter-strings silently, but, occasionally, a question mark prompted
them to read aloud the preceding word (4% of trials). The task coerced the
subjects to process the words until pronunciation. Movement artifacts were
avoided by performing the analysis on the silent trials (96%).
Figure 6–12 displays enlarged views of MEG signals recorded by three
sensors over the left temporal, parietal, and occipital areas. There are some
interesting effects. Over the left temporal area, the response is markedly
different to long nonwords than to the other word types. In the posterior
parietal cortex, on the other hand, there is a particularly strong response to
long real words. The early posterior visual response is similar for all stimuli.
Figure 6–13 illustrates the field patterns that produce the occipital and
left temporal signals in the different experimental conditions: short words,
long words, short nonwords, and long nonwords. The occipital field pattern
at about 200 ms after stimulus onset is very similar for all stimulus types.
Therefore, it is easy to represent these field patterns by a single ECD that
works well for all experimental conditions. It is worth noting that there are
small differences in location and orientation, which may result from noise in
the data or reflect real variability between the conditions. For example, there
may be a slight inferior–superior shift in location for long vs. short letter-
strings. However, the difference is so small that the sources would not appear
as separate ECDs in multidipole modeling (cf. discussion of auditory and
somatosensory data above). The same is true for the field pattern over the left
temporal cortex at 400–500 ms. The pattern is essentially the same for all
conditions, except that it is particularly strong for the long nonwords, and
most difficult to identify for long real words when, at the same time, there is
Figure 6–12. Examples of MEG signals when reading short and long words,
and nonwords. Enlarged views of three sensors. The complete data set of
this subject is displayed in Figure 6–1.
Figure 6–13. Field patterns and the best-fitting ECDs (arrows) correspond-
ing to the MEG signals recorded over the occipital and left temporal areas
(cf. Figure 6–12) in the different experimental conditions.
135
136 MEG: An Introduction to Methods
another field pattern nearby in the parietal cortex. How should one proceed
with the analysis?
Just as for the auditory and somatosensory data above, source areas generating
these signals are determined one by one. The localization is done by scanning
through the response to identify time points at which each distinct field pat-
tern is as closely dipolar as possible, selecting a subset of sensors that covers
the local field maximum, and computing the ECD that best accounts for the
signal measured by these sensors. The resulting ECDs are then brought into a
multidipole model where the locations and orientations of the ECDs are kept
fixed, while their amplitudes are allowed to vary in order to best account for
the data recorded by all MEG sensors over the time interval of interest. The
resulting source waveforms are then used for estimating cortical dynamics
within and across experimental conditions.
There are at least two possible approaches:
(i) Analyze each experimental condition independently, and then combine
the sources into a single multidipole model that works for all conditions.
• Start, e.g., from the most prominent response pattern or from the
earliest systematic response and then work through the data to find
the weaker/later sources.
• For the combined model, choose the sources (or sets of nearby
sources) with clearest field patterns and/or best confidence values.
Check that the combined model accounts for the data in each experi-
mental condition as adequately, and in the same way, as the models
constructed specifically for those conditions.
(ii) Equally well, one can start from one experimental condition, use those
sources as a starting point when one sets off to analyze the next condition,
and add and modify the sources while one works through the data sets.
In fact, it would be best to analyze the data (at least) two or three times,
using different approaches. It is important that one learns to know the data.
Separate sets of sources for each condition are useful if one expects to find
small but systematic differences in location or orientation of current flow—
say, for short vs. long words (cf. Figure 6–13). However, for controlled com-
parison of activation strengths and timing between conditions, one will want
to compose a single set of ECDs (if possible). The minimum number of
sources should be used that can explain the data. It may take some effort to
find the time points and experimental conditions in which a consistently
observed field pattern is most dipolar, i.e., where the signal is strong and
interference from other source areas is minimal. However, the endeavor is
worth the effort, as clean dipolar field patterns facilitate reliable identification
of the underlying sources that normally account both adequately and robustly
for the entire data set.
Multi-Dipole Modeling in MEG 137
In the reading experiment, we may start to look for sources in the long
nonword condition where there was an exceptionally strong left temporal
activation (Figure 6–13). By selecting the frontotemporal sensors painted in
light gray (Figure 6–14a) we can avoid the unwanted effect of other active
areas. The source is located in the superior temporal cortex (Figure 6–14b).
A model that only includes this one source accounts well for the sustained
left temporal signals (Figure 6–14c,d). Characterization of one clear dipolar
field pattern also helps to further recognize systematic unexplained signals in
other areas. The word ‘systematic’ here refers to appearance of the same
waveform on several adjacent sensors. For example, the same unexplained
waveform is seen on three neighboring sensors over the right occipitotemporal
cortex (Figure 6–14c,e).
In cognitive data sets, signals are often weaker overall than for basic
sensory/motor responses, and the large number of active brain areas that are
spatially and temporally quite close to each other appear as additional ‘noise’
when localizing the sources. Accordingly, one must often tolerate focal field
patterns that are not perfectly dipolar. The emerging and re-entering field
patterns may not be fully symmetrical and the resulting ECD will, therefore,
not fall exactly on the zero field line. Nevertheless, when one makes the effort
to find time points at which the different field patterns are as dipolar as possible,
ECD modeling is entirely feasible for cognitive data sets as well.
An effective decrease in signal-to-noise ratio will increase the potential
risk of mislocalizing the source unphysiologically deep. Normally, the local-
ization can be improved by testing other time points and sensor selections, in
order to minimize the effect of other nearby sources. Occasionally, a spatially
extended field pattern may seem to be generated by an unphysiologically deep
ECD when, in reality, it is generated by two source areas that are active at the
same time. Such sources could be located quite close to each other, with fairly
similar directions of current flow, yet separable in space (e.g., distance within
2–3 cm, difference in direction less than 20–30 deg; see Helenius et al., 1999,
for an example of superior temporal activation). This type of situation can
usually be solved by scanning through the different experimental conditions
and identifying the sources at time points where one or the other field pattern
is emphasized.
After the 2–3 analysis rounds, with different approaches, we have a model
composed of 9 ECDs in this subject, with sources in the occipital lobe, left
temporal and parietal, and right occipitotemporal cortex (Figure 6–15a). One
should always check that each individual source makes sense, by comparing
the source waveforms with the original signals. As an example, we could focus
on the left temporal area (Figure 6–15b). The signals in the upper and lower
row of the sensor pairs display quite different stimulus dependence. This
device has two orthogonal sensors in each measurement location. Therefore,
we would expect to see these differential effects on the sensors map onto two
different source areas in the brain that are fairly close to each other but have
almost orthogonal directions of current flow. Indeed, our solution suggests
that sources 5 and 9 produce these signals. The long nonwords differ from the
other stimuli in source 9 in the superior temporal cortex, while in the more
inferior, rather horizontally oriented source 5, the response is strongest for
the short nonwords, in agreement with the original measured signals.
In order to check that the model is really meaningful, it is very useful to
compare two stimulus conditions at a time. Figure 6–16 depicts responses to
long words and long nonwords. There are at least two very obvious differ-
ences in the original waveforms—one over the left temporal and the other
over the posterior parietal cortex (Figure 6–16b). These areas apparently cor-
respond to sources 9 and 8, respectively (Figure 6–16a). Indeed, source 9
shows the strong response to nonwords, whereas source 8 displays the
stronger activation to real words, with the time behavior matching that of the
original sensor waveforms. Accordingly, we can be reasonably satisfied with
this model.
Multi-Dipole Modeling in MEG 139
Figure 6–15. Source analysis in the reading task in one subject. (a) ECDs
shown on the surface rendering of the subject’s MRI and time courses of
activation in those areas for the different stimulus types. The sources are
ordered according to latency of activation. Goodness-of-fit value (g) is
plotted at the bottom. (b) Selection of MEG sensors over the left temporo-
parietal cortex. Enlarged view of two orthogonally oriented sensors is
plotted below.
After an acceptable source model has been compiled, one may collect any
descriptive values from the waveforms (e.g., onset/offset/peak timing, peak
amplitude, mean amplitude, or integral over specific time intervals) and test
them for significant stimulus/task effects.
Here, as an example, we will describe one possible approach for the data
set on word/nonword reading. Behaviorally, it has been reported that naming
latencies are shorter for short than long words, and that this length effect is
markedly stronger for naming nonwords than real words (Weekes, 1997).
A comparison between short words and long nonwords should encompass all
these effects: length effect, lexicality effect, and their interaction.
First, we search for significant differences in individual subjects. Figure 6–17
displays the time course of activation in the left temporal cortex for those two
conditions, in one subject. The signal variation in the prestimulus baseline
interval carries information about the noise level in this area, in this specific
source model and in these experimental conditions. A rather conservative
approach is to estimate, for example, the level of 2.58 times standard devia-
tion (corresponding to p<0.01), represented by the gray box, and only accept
as significant those differences between the two waveforms that exceed this
140 MEG: An Introduction to Methods
level (cf. Tarkiainen et al., 1999). In this case, we would identify a significantly
stronger activation to long nonwords than short words in the left superior
temporal cortex at about 400 to 700 ms after stimulus onset.
Cortical areas with stimulus effects were identified in the same way in all
subjects. Figure 6–18 collects the subsets of source areas that showed signifi-
cant differences between the extreme conditions of long nonwords and short
real words. The source clusters are presented separately for differences
detected within the first 200 ms after stimulus onset (Figure 6–18a) and after
200 ms (Figure 6–18b). Within 200 ms after stimulus, differences in the peak
amplitude were found in 6 of 8 subjects, with the sources clustered around the
occipital midline. After 200 ms, the most salient cluster of such sources (6 of
8 subjects) was found in the left superior temporal cortex.
Here, we used functional criteria combined with location information to
identify sources that were observed consistently across subjects. Within these
occipital and left temporal clusters, we now test for group-level statistical
effects between all four stimulus categories. The strength of the early occipital
activation (Figure 6–19a) showed a pure length effect, and most probably
reflects basic visual feature analysis within the first 200 ms (cf. Tarkiainen
et al., 1999). In the left temporal cortex, the duration of activation showed
interesting behavior—namely, a weak effect of length for real words but a
marked effect for nonwords; the duration for the 8-letter nonwords was twice
that for the 4-letter nonwords (Figure 6–19b). The peak amplitude and mean
signal strength also showed the same interaction between length and lexical-
ity. This pattern parallels that reported for reaction times in word and non-
word reading (Weekes, 1997). If we accept the dual-route model of reading,
this pattern should be interpreted as reflecting phonological analysis.
Considering that activation of the left superior temporal cortex in this very
142 MEG: An Introduction to Methods
Figure 6–19. Significant stimulus effects at the group level. (a) Maximum
amplitude (mean ± SEM) of the early occipital activation was significantly
stronger to long than short letter-strings, regardless of lexicality. (b) Dura-
tion of the left temporal activation was also influenced by letter-string
length but significantly more for nonwords than real words.
Modified from Wydell et al. (2003).
the experiment. The signals are aligned to the stimulus onset, which is shown
by the continuous vertical line. The mouth movement and speech onset vary
by about 100 to 200 ms from trial to trial, both with respect to the stimulus
onset and with respect to each other. This jitter makes it possible to dissociate
the artifact signal from the cortical activity of interest.
In Figure 6–22, the original MEG data were averaged with respect to
speech onset, recorded with a microphone. The speech artifact is usually quite
accurately time-locked to microphone onset and, therefore, one can get a very
clean artifact pattern with this averaging procedure. The signal concentrates
along the rim of the helmet. Task-related cortical activity has faded out
because of the jitter with respect to stimulus onset. By emphasizing the arti-
fact this way, one can then remove the disturbing field pattern from the
responses. This can be done with the help of, e.g., the Signal Space Projection
(SSP) method (Uusitalo and Ilmoniemi, 1997).
Figure 6–22. MEG data averaged with respect to speech onset, determined
from the microphone record. The helmet views illustrate the magnetic
field pattern at speech onset (t = 0 ms).
The same data can also be averaged with respect to the EMG signal, that
is, mouth movement onset (Figure 6–23). This procedure allows focus on the
motor control of speech production. The speech artifact field (Figure 6–22)
has been removed from the averaged signals. Clear dipolar field patterns
emerge over the left and right frontal lobes, approximately at the time when
the mouth movement starts. The sources of these field patterns are readily
localized to the bilateral face motor cortex.
Figure 6–23. MEG data averaged with respect to mouth movement onset,
determined from the EMG record. The helmet views illustrate the mag-
netic field pattern approximately at mouth movement onset (t = 20 ms)
and the arrows the best-fitting ECDs. The planar gradiometers that are
maximally sensitive to activation in these source areas are marked with
rectangles.
Multi-Dipole Modeling in MEG 145
Figure 6–24 displays the same data finally averaged in the typical fashion,
with respect to the stimulus onset. Again, the speech artifact field has been
removed. In this data set, bilateral frontal activations are observed best at
about 500 ms after word presentation. Over the left hemisphere the field pat-
tern is clearly dipolar, and the source in the face motor cortex can be deter-
mined reliably. However, in the right hemisphere, the field pattern points to
activation of the motor cortex but the underlying source area is practically
impossible to localize; the helmet view in Figure 6–24 represents the most
dipolar field pattern one can detect. Activation of the right motor cortex in
this subject thus seems to be more strongly time-locked to mouth movement
onset than to stimulus onset. A reasonable approach would be to determine
this source from the signals averaged with respect to mouth movement onset
(Figure 6–23) and then use that ECD to account for activity in the signals
averaged with respect to stimulus onset (Figure 6–24).
Figure 6–25 illustrates the cortical activation sequence at the group level.
In this experiment there was no systematic functional variation in the stimuli
or task. The main aim was to compare fluent speakers and stutterers in the
basic task of overt reading of single words. The source areas from individual
subjects were grouped together primarily by similarity in location. The curves
give the mean time course of activation in those areas, averaged across
fluently speaking subjects. The first vertical line indicates the word presenta-
tion and the second vertical line the appearance of the vocalization prompt
(question mark).
The occipital and the left and right inferior occipitotemporal cortices
were active within the first 200 ms. Next, there was activation in the left superior
Figure 6–24. MEG data averaged with respect to stimulus onset. The helmet
views illustrate the magnetic field pattern when the subject is waiting for
the vocalization prompt (t = 480 ms) and the arrow the best-fitting ECD
in the left hemisphere. The planar gradiometer that is maximally sensitive
to activation in this source area is marked with a rectangle.
146 MEG: An Introduction to Methods
temporal and inferior parietal cortices, starting about 200 ms after word onset
and reaching the maximum at about 400 ms. The left inferior frontal cortex,
approximately Broca’s area, showed activation in the same time window. The
sources assigned to the temporal cluster are spatially quite close to those
assigned to the inferior frontal cluster, on the one hand, and sources belong-
ing to the inferior parietal cluster, on the other hand. Here, the orientation of
current flow was used as an additional criterion (white bars plotted on the
three source clusters in Figure 6–25, middle column) which clearly distin-
guished between the adjacent clusters.
All the activations listed in the two left-most columns return to baseline
before the vocalization prompt. The signals depicted in the right-most column
begin at about 200 ms after word onset, and persist until actual vocalization
and even beyond it. This is quite reasonable, as those signals arise from the left
and right sensorimotor and premotor cortices and, apparently, from the
supplementary motor area.
Differences between the time courses of activation in fluently speaking
individuals and stutterers are collected in Figure 6–26. There were group dif-
ferences in three brain areas and time windows. Within the first 400 ms, while
preparing for vocalization the fluent speakers first activated Broca’s area, and
then the left motor/premotor cortex, which appears to be a natural order of
events. In the stutterers, however, the sequence was reversed. They showed
exceptionally early activation in the left sensorimotor/premotor area, already
within the first 200 ms, whereas activation of Broca’s area was delayed with
Figure 6–25. Group-level clusters of active cortical areas and their mean
time courses of activation across 10 fluently speaking subjects. In the mid-
dle column, the mean orientation of current flow in the source clusters is
shown as well (white bars).
Modified from Salmelin (2007).
Multi-Dipole Modeling in MEG 147
respect to that observed in fluent speakers. After the vocalization prompt, the
right motor cortex showed significantly stronger activation in fluent speakers
than in stutterers. The combination of MEG data and ECD analysis thus made
it possible to compare the groups at three different levels: the locations of the
active areas were fairly similar, but differences emerged in timing within the
first 400 ms and in activation strength during overt vocalization.
with test calculation of ECDs for a number of time points, provides an estimate
for setting acceptance criteria (e.g., goodness-of-fit value, confidence values
for localization) for the large set of ECDs obtained with automatic localiza-
tion over long stretches of data. Typically less than 5% of the original ECDs
represent reliable dipolar sources within the brain region covered by the selec-
tion of sensors (Salmelin & Hari, 1994a; Salmelin et al., 1995).
Figure 6–28 gives examples of source clusters along the central sulcus in
one subject, obtained from data filtered around 10 Hz and around 20 Hz. The
clusters may also be conveniently displayed as dipole density plots (e.g.,
Lehtelä et al., 1997; Liljeström et al., 2005; Vieth et al., 1996). In the resting
brain, the sources of mu rhythm are concentrated in and around the hand
representation area in the central sulcus. The 10-Hz component originates
largely in the somatosensory cortex—but also precentrally—whereas the
20-Hz component seems to be predominantly a motor cortical rhythm
(Salenius et al., 1997; Salmelin & Hari, 1994b; Salmelin et al., 1995).
In voluntary movements the 20-Hz activity shows somatotopic organization.
Figure 6–29 illustrates the results of a study in which subjects made self-paced
movements of the left and right toes, index fingers, and mouth. The colored
dots show foot, hand, and mouth representation areas along the central sul-
cus, determined with electric stimulation of the tibial and median nerve and
the lower lip, respectively. The colored blobs are sources of 20-Hz oscilla-
tions. For movement of left and right toes, the sources were concentrated
close to the foot area; for left and right index finger flexion, close to the con-
tralateral hand area; and for mouth movements, sources extended laterally to
the mouth area—in this subject, mainly in the right hemisphere. Note that the
sources again cluster frontally with respect to the central sulcus. This result
implies that one may use 20-Hz activity to evaluate functionality of different
parts of the motor cortex, not only the hand area.
Figure 6–30 presents the whole-head view of one subject when he was
reading words aloud. The curves display the mean amplitude of 20-Hz oscil-
lations from 1 second before word onset to 5 seconds after it (Salmelin et al.,
2000). The MEG signals were filtered to 16–24 Hz, their absolute value taken,
and the resulting signals averaged with respect to word onset (TSE, Temporal
Spectral Evolution; Salmelin & Hari, 1994b). This approach reveals modula-
tion of activity that occurs systematically, but not exactly at the same time,
from trial to trial (event-related vs. phase-locked activity). There is a clear
suppression of 20-Hz activity quite locally over the lateral areas, but also more
medially, with different time behaviors.
Figure 6–31a illustrates the localization of the 20-Hz oscillations—a gen-
eral pattern, here illustrated in one (typical) subject—to the hand and mouth
areas of the right and left hemispheres. With these candidate source areas one
may proceed, in principle, through the same steps as in the multidipole analy-
sis of evoked responses described above. One can estimate the time course of
activity in the proposed ECDs in the left and right hand and mouth areas
(Figure 6–31b). In order to evaluate, in individual subjects, how well the
sources account for the MEG signals measured over the whole head, we can
perform a forward calculation to map the estimated source activity onto the
MEG sensors (model signals). In this study, the interest was in localizing the
sources of the 20-Hz modulation. Accordingly, TSE curves of modeled MEG
signals were compared with TSE curves of the original measured MEG
signals. The four sources (Figure 6–31a) were found to account well for the
event-related modulation of 20-Hz activity (Figure 6–30).
Figure 6–31. Source analysis of 20-Hz activity. (a) ECDs in the left and right
hand and mouth areas along the central sulcus in one subject. (b) Time
course of 20-Hz activation in those areas. Vertical lines denote a new trial
(word presentation). (c) Grand average time course of 20-Hz modulation,
calculated across 10 subjects.
Modified from Salmelin et al. (2000).
From the continuous source activity (Figure 6–31b), one can compute
the mean level of activity (TSE) in those brain areas. Figure 6–31c displays
grand-averaged source-level TSE curves (10 subjects) in the right hand and
mouth, and left hand and mouth areas. The first vertical line indicates word
presentation, and the second line indicates the vocalization prompt. The gray
bar indicates actual speech production. Just as in the analysis of phase-locked
evoked responses, it is important to check that the ECDs included in the mul-
tidipole model do not interact—i.e., that they are not too close to each other,
and that the orientation of current flow in those areas is not too similar. In the
case of rhythmic activity, possible interaction is harder to evaluate from the
source waveforms than it is in the case of evoked responses. However, clearly
different time behaviors, like those between the hand and mouth areas within
each hemisphere in this data set, speak in favor of independence of the ECDs.
One may, for example, estimate the space angle between the magnetic field
distributions produced by the ECDs (should exceed 30–40 degrees), or simu-
late activity in those areas and evaluate how much of the signal generated by
one ECD is accounted for by the other ECD. In reality, the possible interac-
tions are not determined only by pairs of sources but by all sources included
in the model. Nevertheless, comparison of the most suspicious pairs of ECDs
provides a good approximation of the potential risks.
Equipped with reasonable source models, one may proceed to comparison
of different experimental conditions or subject groups, in the same way as for
phase-locked evoked responses. Figure 6–31c shows that in the mouth motor
cortex the 20-Hz rhythm was strongly suppressed well before vocalization.
152 MEG: An Introduction to Methods
Discussion
This chapter has, hopefully, emphasized to the reader that MEG data are usu-
ally not ambiguous. It is mostly quite obvious which areas are active. In that
sense, the infamous inverse problem is not really the problem in MEG data
analysis. It is more a question of identifying the clear field patterns, localizing
the sources that generate them, and constructing a clean, well-behaved model
with those sources.
Careful experimental design and high-quality data are the prerequisite
for successful data analysis. The experimenter should make sure that the data
Multi-Dipole Modeling in MEG 153
are of highest quality. No analysis method can redeem data of poor quality.
The first step is to decide what one really needs to find out, and choose the
paradigm accordingly. For example, accurate localization is certainly needed
for presurgical mapping of the central sulcus. The early somatosensory evoked
fields are well suited for this purpose, as there is little activity in other cortical
areas than the primary somatosensory cortex. However, if one studies differ-
ences in timing, e.g., during language tasks, extremely accurate localization is
not really the issue to push. Parametric variation of stimuli or tasks is essential
for functional localization (e.g., Tarkiainen et al., 1999) and very useful in
source modeling.
One has to learn to read the MEG signals. That is the sound basis for all
analysis. It is essential to know the data throughout. It may be a good idea to
use different analysis methods to look at the same data; it also makes one spend
more time scrutinizing the signals and thus become thoroughly familiar with
them. This makes it possible to fully understand the solutions one gets.
Finally, when a solution or model is ready, one should be able to recognize
it also in the original signals. Everything should make sense in the end. The
analysis tools must provide visual control of the model and what it explains.
A major advantage of the ECD approach is that there are no hidden
assumptions in mapping the field pattern to activation in the brain. The main
criticism that has been raised against ECD modeling is that the user makes (or
can make) subjective choices about the total number of active areas and the
subset of sensors to be included in source localization. However, fundamen-
tally this is no different from the subjective choices of thresholding one needs
to make when using distributed analysis methods, or in PET/fMRI analysis.
When ECD analysis is performed on the same MEG data set by different,
experienced researchers, the same major sources emerge; variability, if any,
appears in the inclusion of the weakest sources. The unparalleled asset of ECD
analysis is that, having full visual control of the process, the user knows exactly
why certain areas are found to be active and others not. It is advisable to carry
out an ECD analysis, at least in a cursory fashion, before using more automatic
analysis or visualization tools.
Basic sensory or motor processes often involve a relatively small number of
brain areas that can be readily distinguished by location and/or direction of cur-
rent flow and that show clear dipolar field patterns, at least at specific time inter-
vals. When studying cognitive functions, the overall activation tends to be
weaker, the time intervals of interest longer, and the effective signal-to-ratio
lower because of the large number of sources that are spatially close and show
considerable temporal overlap. Because of that, one often has to accept less per-
fect dipolarity of the field patterns than for functionally simpler tasks. Otherwise,
the general procedures and considerations are similar for all types of data sets.
When using the ECD approach, one obviously seeks to identify the dis-
tinct dipolar field patterns in the data. In principle, one could argue that such
an assumption ignores active areas that do not show dipolar signal distribution.
However, at the distance MEG signals are recorded (at least 3 cm from the
154 MEG: An Introduction to Methods
source currents) cortical activity appears dipolar as higher terms are reduced,
relatively more rapidly, with distance. Most importantly, when a set of ECDs,
determined from clearly dipolar field patterns, suffice to account for the data
recorded by all sensors without leaving any systematic group of signals unex-
plained—and that is normally the case—it is hard to argue that alternative
accounts of source structure would be critically needed.
Obviously, there are no miracle tools that would be more correct than
others—and this is, of course, the real inverse problem in MEG (or EEG)
analysis. User-independent tools should be particularly suitable for compari-
son across sites and users; nevertheless, one should not assume that such
methods are more correct than others. For clear, good-quality data, any anal-
ysis tool should work well and give very similar results. Whichever tool one
uses, one should be cautious in the interpretations—but that is true for all
imaging techniques.
References
Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993). Models of reading aloud:
dual-route and parallel-distributed-processing approaches. Psychological Review,
100, 589–608.
Cornelissen, P., Tarkiainen, A., Helenius, P., & Salmelin, R. (2003). Cortical effects
of shifting letter-position in letter-strings of varying length. Journal of Cognitive
Neuroscience 15, 731–746.
Dale, A. M., Liu, A. K., Fischl, B. R., Buckner, R. L., Belliveau, J. W., Lewine, J. D., et al.
(2000). Dynamic statistical parametric mapping: Combining fMRI and MEG for
high-resolution imaging of cortical activity. Neuron, 26, 55–67.
Forss, N., Hari, R., Salmelin, R., Ahonen, A., Hämäläinen, M., Kajola, M., et al. (1994).
Activation of the human posterior parietal cortex by median nerve stimulation.
Experimental Brain Research, 99, 309–315.
Halgren, E., Dhond, R. P., Christensen, N., Van Petten, C., Marinkovic, K., Lewine, J. D.,
et al. (2002). N400-like magnetoencephalography responses modulated by semantic
context, word frequency, and lexical class in sentences. Neuroimage, 17, 1101–1116.
Hari, R. & Salmelin, R. (1997). Human cortical oscillations: a neuromagnetic view
through the skull. Trends in Neurosciences 20, 44–49.
Helenius, P., Salmelin, R., Service, E., & Connolly, J. F. (1999). Semantic cortical
activation in dyslexic readers. Journal of Cognitive Neuroscience, 11, 535–550.
Hämäläinen, M., Hari, R., Ilmoniemi, R. J., Knuutila, J., & Lounasmaa, O. V. (1993).
Magnetoencephalography – theory, instrumentation, and applications to noninva-
sive studies of the working human brain. Reviews of Modern Physics, 65, 413–497.
Lehtelä, L., Salmelin, R., & Hari, R. (1997). Evidence for reactive magnetic 10-Hz
rhythm in the human auditory cortex. Neuroscience Letters, 222, 111–114.
Liljeström, M., Kujala, J., Jensen, O., & Salmelin, R (2005). Neuromagnetic locali-
zation of rhythmic activity in the human brain: a comparison of three methods.
Neuroimage, 25, 734–745.
Lin, F. H., Witzel T., Hämäläinen, M. S., Dale, A. M., Belliveau, J. W., & Stufflebeam,
S. M. (2004). Spectral spatiotemporal imaging of cortical oscillations and interac-
tions in the human brain. Neuroimage, 23, 582–595.
Multi-Dipole Modeling in MEG 155
Mäkelä, J., Ahonen, A., Hämäläinen, M., Hari, R., Ilmoniemi, R., Kajola, M.,
et al. (1993). Functional differences between auditory cortices of the two hemisphe-
res revealed by whole-head neuromagnetic recordings. Human Brain Mapping, 1,
48–56.
Pantev, C., Ross, B., Berg, P., Elbert, T., & Rockstroh, B. (1998). Study of the human
auditory cortices using a whole-head magnetometer: left vs. right hemisphere and
ipsilateral vs. contralateral stimulation. Audiology and Neurootology, 3, 183–190.
Pfurtscheller, G., & Lopes da Silva, F. H. (1999). Event-related EEG/MEG synchronization
and desynchronization: basic principles. Clinical Neurophysiology, 110, 1842–1857.
Saarinen, T., Laaksonen, H., Parviainen, T., & Salmelin, R. (2006). Motor Cortex
Dynamics in Visuomotor Production of Speech and Non-speech Mouth Move-
ments. Cerebral Cortex, 16, 212–222.
Salenius, S., Schnitzler, A., Salmelin, R., Jousmäki, V., & Hari, R. (1997). Modulation
of human cortical rolandic rhythms during natural sensorimotor tasks. Neuro-
image, 5, 221–228.
Salmelin, R. (2007). Clinical neurophysiology of language: the MEG approach. Clini-
cal Neurophysiology, 118, 237–254.
Salmelin, R., & Hari, R. (1994a). Characterization of spontaneous MEG rhythms in
healthy adults. Electroencephalography and clinical Neurophysiology, 91, 237–248.
Salmelin, R., & Hari, R. (1994b). Spatiotemporal characteristics of sensorimotor MEG
rhythms related to thumb movement. Neuroscience, 60, 537–550.
Salmelin, R., Hämäläinen, M., Kajola, M., & Hari, R. (1995). Functional segregation of
movement-related rhythmic activity in the human brain. Neuroimage, 2, 237–243.
Salmelin, R., & Sams, M. (2002). Motor cortex involvement during verbal versus non-
verbal lip and tongue movements. Human Brain Mapping, 16, 81–91.
Salmelin, R., Schnitzler, A., Parkkonen, L., Biermann, K., Helenius, P., Kiviniemi, K.,
et al. (1999). Native language, gender, and functional organization of the auditory
cortex. Proceedings of the National Academy of Sciences USA, 96, 10460–10465.
Salmelin, R., Schnitzler, A., Schmitz, F., & Freund, H.-J. (2000). Single word reading
in developmental stutterers and fluent speakers. Brain, 123, 1184–1202.
Schnitzler, A., Salmelin, R., Salenius, S., Jousmäki, V., & Hari, R. (1995). Tactile
information from the human hand reaches the ipsilateral primary sometosensory
cortex. Neuroscience Letters, 200, 25–28.
Tarkiainen, A., Helenius, P., Hansen, P. C., Cornelissen, P. L., & Salmelin, R. (1999).
Dynamics of letter string perception in the human occipitotemporal cortex. Brain,
122, 2119–2131.
Uusitalo, M. A., & Ilmoniemi, R. J. (1997). Signal-Space Projection method for sepa-
rating MEG and EEG signals into components. Medical & Biological Engineering &
Computing, 35, 135–140.
Uutela, K., Hämäläinen, M., & Somersalo, E. (1999). Visualization of magnetoence-
phalographic data using minimum current estimates. Neuroimage, 10, 173–180.
Vieth, J. B., Kober, H., & Grummich, P. (1996). Sources of spontaneous slow waves
associated with brain lesions, localized by using the MEG. Brain Topography, 8,
215–221.
Weekes, B. S. (1997). Differential effects of number of letters on word and nonword
latency. Quarterly Journal of Experimental Psychology, 50A, 439–456.
Wydell, TN., Vuorinen, T., Helenius, P., & Salmelin, R. (2003). Neural correlates of
letter-string length and lexicality during reading in a regular orthography. Journal
of Cognitive Neuroscience, 15, 1052–1062.
7
Estimating Distributed Representations of
Evoked Responses and Oscillatory Brain Activity
Ole Jensen and Christian Hesse
Introduction
Several techniques have been developed to construct source models accounting
for electrophysiological brain activity measured by EEG and MEG. Rather than
modeling the measured signal using only a small number of (discrete) dipole
sources, the approaches considered in this chapter essentially estimate the con-
tribution of all sources within the entire brain volume to the observed MEG
or EEG, and thereby provide a distributed representation of the underlying
neuronal activity. A major advantage is that such distributed representations
156
Estimating Distributed Representations of Evoked Responses 157
Theory
Basics of Source Modeling
The aim of biophysical source modeling is to account for the measured MEG/
EEG signal in terms of a set of underlying current sources located within the
brain volume. Equivalent current dipoles are a widely used model (Fig. 7–1)
for approximately describing the spatial and temporal characteristics (orien-
tation and magnitude) of the local current flow within a small volume of
brain tissue, caused by the synchronized activation (post-synaptic currents)
of many neurons (pyramidal cells). More formally, the N measured signals
b(t) = [b1 (t), b2 (t),…, dN(t)]T reflecting the time-varying magnetic field (or
electric potential) patterns at the scalp are assumed to have been generated by
the summed field patterns of a set of M (unobserved) current dipole sources
b (t ) = Gq (t ) + n (t ) (7–1)
b=Gq+n
min || q ||
Figure 7–1. The principle of current estimates. (b) The brain volume is
modeled by a three-dimensional grid. (b) The full current distribution is
approximated by the current dipoles, q, at each grid point. The aim is to
find the current distribution that produces a field that matches the field
measured by the sensors, b. The forward model G relates the current
distribution to the measured field. (b) For visualization, the magnitudes
of the current dipoles are projected to the brain surface and represented
by a color code.
min q (7–2)
A) B)
1
0.5
2
0.4
0.3
0.2
0.1
−4 −3 −2 −1 0 1 2 3 4
Figure 7–2. (A) The current distributions to the left and the right produce
very similar fields. However, when applying the minimum current constraint
the distributions to the right are chosen. (B) The a priori distributions of
current magnitudes when applying the L1- and the L2-norm estimates.
Reproduced from Uutela et al. (1999).
160 MEG: An Introduction to Methods
Okabe, 1995) whereas estimates obtained using the L1–norm are referred to
as Minimum Current Estimates (MCEs) (Matsuura & Okabe, 1995; Uutela
et al., 1999). The L2–norm implicitly assumes a Gaussian a priori current dis-
tribution, whereas the L1–norm assumes a Laplacian (double exponential)
distribution (Figure 7–2B). Consequently, the MNE results in more spatially
smeared current distributions compared to the MCE. Thus, when selecting
between the L1– and the L2–norm, an implicit choice is being made in terms
of spatial smoothness of the current distributions. From a theoretical point of
view there are no well-motivated reasons for choosing one approach over the
other. From a physiological perspective we do not have sufficient information
to judge which approach is more appropriate. Minimization procedures with
additional smoothness constraints have been implemented in the LORETA
method (Pascual-Marqui et al., 1994). As for MNE, the LORETA approach
results in quite smeared source estimates.
One practical problem that arises when calculating current estimates is
that superficial sources are overestimated at the expense of deeper sources.
This is a consequence of deep focal sources producing fields very similar to
extended superficial sources (Figure 7–3). Since the MEG sensors are more
sensitive to nearby sources, the current-estimate methods become biased
toward the superficial sources when applying the minimization constraint. In
order to reduce this bias, a weighted norm is often applied to reduce the con-
tribution of the superficial currents (Ioannides et al., 1990; Liu et al., 1998;
Uutela et al., 1999). It is important to keep in mind that when selecting a spe-
cific approach to reduce the depth bias, a choice is implicitly being made that
has consequences for the final current distribution. The minimum norm esti-
mate using the L2–norm can be derived directly by inverting Equation (7–1),
but this is not the case for the minimum current estimate, since it involves
the L1–norm. Uutela et al. (1999) used linear programming in order to solve
Figure 7–3. Deep focal and superficial distributed sources produce very
similar fields. Since magnetic fields decrease with distance, the deeper
sources must be stronger in magnitude compared to superficial sources in
order to produce the same field. Thus, the minimization constraint biases
the solution toward superficially distributed sources.
Estimating Distributed Representations of Evoked Responses 161
Basics of Beamforming
C −1Gi
wi = (7–4)
GTi C −1Gi
where C−1 is the inverse of the cross-covariance matrix of the measured signal
C = b(t)b(t)T/(T-1), where T is the number of time.
The expression for the spatial filter weights in Equation (7–4) is vari-
ously known as Capon’s beamformer, the linearly constrained minimum
variance (LCMV) beamformer, and the minimum mean square error (MMSE)
estimator (e.g., Haykin, 2002). It is the use of the cross-covariance matrix of
the data that makes this type of filter “adaptive” and accounts for its spa-
tially selective enhancement and interference-suppression characteristics.
When applied to frequency-domain data, the cross-covariance matrix C is
replaced by a matrix reflecting the cross spectrum (or cross-spectral den-
sity) at a particular frequency, or over a frequency band of interest. In the
context of MEG/EEG neuroimaging, frequency-domain beamforming is
also known as the dynamic imaging of coherent sources (DICS) method (e.g.,
Gross et al., 2001; Chapter 9). Alternatively, LCMV beamforming can also
be applied to bandpass-filtered data, which, in MEG neuroimaging, is also
referred to as synthetic aperture magnetometry (SAM) analysis (Robinson &
Vrba, 1999).
Beamforming relies on two key assumptions, namely: a) that the activa-
tion time courses of all sources are mutually uncorrelated; and b) that the
forward model describing the field patterns of the sources is correct, or at
least sufficiently accurate.2 The mathematical exposition required to explain
in detail why these assumptions are necessary—and what happens when
they are violated—is beyond the scope of this introductory chapter. However,
a relative intuitive account would be the following: The cross-covariance
Estimating Distributed Representations of Evoked Responses 163
the fields of neighboring dipoles close to the surface. Particularly in the pres-
ence of noise, this leads to larger localization errors. Primarily because of this
“spatial leakage,” beamformers tend to overestimate the signal power of deep-
lying sources, which makes it difficult to interpret the “raw” beamformer sig-
nal amplitude estimates. Two approaches have been proposed to circumvent
this problem. A normalization approach based on the neural activity index
(NAI), put forward by Van Veen (1997), scales (divides) the projected power
at each source location by the corresponding projected noise power, using a
suitably accurate model estimate of the signal noise.3 A practical alternative
for comparisons of source activity between two experimental conditions is to
simply use the ratio of source power between conditions. The latter approach
assumes an implicit signal model in which the noise characteristics (regard-
less of interpretation) are identical in both conditions, and is more widely
used in cognitive neuroscience (e.g., Nieuwenhuis et al., 2008).
Finally, it should be noted that beamforming, or adaptive spatial filtering
as such, is not a source localization technique, but rather a method for estimat-
ing (extracting) a signal at a known location. Nevertheless, the location of impor-
tant or dominant sources can be determined by appropriate postprocessing of
the distributed representation of source activity obtained by beamforming,
e.g., thresholding or detection of local peaks in activation/power.
Stimulus
onset
Averaging
Figure 7–4. Evoked responses are phase-locked to the stimulus and will
subsequently be presented in the averaged data. Induced responses are
not necessarily phase-locked, and are attenuated by averaging.
Reproduced from Tallon-Baudry and Bertand (1999).
It should be pointed out that just because there are differences in a given fre-
quency band when comparing two groups of subjects, it does not necessarily
mean that there is spontaneous oscillatory activity with a magnitude strong
enough for the sources to be modeled. In order to be able to assess whether
the sources of spontaneous oscillatory activity can be identified, it is recom-
mended to perform a spectral analysis of power at the sensor level. Such an
analysis serves several purposes. If a “clear” peak in the spectra can be identi-
fied, it suggests that oscillatory activity is present, rather than a broad band
signal. Second, spectral analysis will allow for identifying the frequency peak
and range. Given that the peak frequency of alpha oscillations varies among
subjects, the optimal approach when identifying oscillatory sources is to tar-
get the peak frequency in individual subjects (Posthuma et al., 2001). A com-
mon approach to spectral analysis of spontaneous oscillatory activity is to
apply Fourier transforms to the ongoing data. We will here describe in detail
how this is done, given that this insight is important for understanding how
to calculate MCEs/MNEs in the frequency domain. The power spectrum is
calculated from the modulus of the Fourier transform squared as a function
of frequency. Typically, the fast Fourier transform (FFT) algorithm is applied.
Using this approach directly for power estimation has one drawback: the
variance of the power spectra estimates does not decrease with increasing
166 MEG: An Introduction to Methods
data length. This problem can be solved using Welch’s method by dividing
the data into segments of equal length (“windowing”). The power spectra
calculated for each segment are then averaged, thus reducing the variance of
the spectral estimate with increasing number of segments (Welch, 1967;
Challis & Kitney, 1991). Frequency smoothing can be controlled by changing
the length of the data segments, i.e., the frequency resolution decreases as the
length of the data segments becomes shorter. Some frequency smoothing
might be advantageous, given the frequency fluctuations in cortical rhythms.
For instance, in a typical subject the alpha activity might fluctuate between
10 and 12 Hz. Practical experience has shown that adjusting the length of the
time window yielding a frequency resolution of 0.5–1 Hz results in sensible
power spectra for activity in the alpha and beta range. Another concern that
arises when segmenting data prior to spectral estimates, is spectral leakage
emerging due to edge effects of the segments. The spectral leakage can be
reduced by applying a windowing function, i.e., a taper, to each segment.
Often a Hanning taper is applied (Challis & Kitney, 1991). To compensate for
the data loss due to the taper at the edges of the segment, overlapping seg-
ments are applied—e.g., 50% overlapping segments are typically used for
Hanning tapers. Figure 7–5A shows an example where the power spectra have
been calculated using Welch’s method for a subject resetting with eyes closed.
Note that the dominant frequencies of both alpha and beta activity can be
identified in the spectra. While Welch’s method for power spectra estimation
is commonly used, multitaper techniques provide a better control of the fre-
quency smoothing, and are becoming increasingly popular (Percival, 1993;
Mitra & Pesaran, 1999).
×1000 (fT/cm)2
A) A B)
2
1
0 11 Hz 21 Hz
5 10 15 20 25
0.3
Frequency (Hz)
Strength (nAm)
A
C)
×1000 (fT/cm)2
2
B
1
0
5 10 15 20 25
Frequency (Hz)
Figure 7–5. Power spectra and MCEs calculated from MEG data recorded from a subject resetting with eyes closed. (A) The
power spectra were calculated using Welch’s method. The window length was 1.7 s. (B) The frequency domain MCEs were
calculated for the peak frequencies (11 and 21 Hz) identified in the power spectra. (C) The sources of the peak frequencies
were co-registered on the subject’s MRI.
Reproduced from Jensen and Vanni (2002).
168 MEG: An Introduction to Methods
FFT
bk(t) Bk(f) Bk(f’)
N samples
b1(t)
b2(t)
b3(t)
b306(t)
k=1 k=2 2 2
B1Re (f') B1Im(f') B Re(f') B Im(f')
1
+ + =
2M
q1Re (f') q1Im (f')
Figure 7–6. Calculating MCEs in the frequency domain. The Fourier trans-
form is calculated for the each time window (epoch) for the frequency
of interest. The current estimates for the real and imaginary parts of the
Fourier-transformed data are combined for each time window. Finally,
the current estimates are averaged.
Reproduced from Jensen and Vanni (2002).
methods for power spectra estimation, applying a Hanning taper to each win-
dow will reduce the spectral leakage. Windowing the data poses one problem:
the complex current estimates for each time window will have different
phases. This problem is solved by calculating the current estimates for the real
and imaginary parts of the Fourier-transformed data. The real and imaginary
parts of the current estimates are then combined, leaving only the magnitude
of the current estimate (Jensen & Vanni, 2002). These magnitudes are then
averaged over time windows. As a result, the phase information and thus the
source orientation are lost. While this method has been implemented for
MCEs (Jensen & Vanni, 2002), it should be mentioned that the same approach
can be used for MNE and LORETA. The implications for the estimated source
distributions are the same as those for event-related fields: the MNE source
distributions will be more spatially smeared as compared to those calculated
by MCE. Figure 7–5 shows an example where the sources of spontaneous
oscillatory activity in alpha and beta band have been estimated. First, the power
spectra were calculated in order to identify the frequency peaks at ~11 and
Estimating Distributed Representations of Evoked Responses 169
~21 Hz. The current distributions for those frequencies were then estimated
and subsequently mapped onto the subject’s anatomical MRI.
−t 2
w (t , f ) = A exp( )exp (2iπ f 0t ) (7–5)
2σ t2
1
where A = . The time-frequency representation (TFR) of power P(t, f0)
σ t 2π
for a given signal at time (t) and frequency (f ) is given by the squared norm
of the convolution of a Morlet wavelet to the signal s(t):
A) B) 30
f = 40 Hz :
25
Frequency (Hz)
20
f = 20 Hz : 15
10
5
f = 10 Hz :
0
−0.5 0 0.5 1 1.5
Time (s)
100 ms
C)
−27
MRC15 / 77 × 10
30 3
25 2
20 1
15 0
10 −1
5 −1
0 −3
−0.5 0 0.5 1 1.5
When localizing induced oscillatory activity, the aim is to calculate the time
course of oscillatory sources. This is different from localizing the sources of
event-related fields, since the single trial signals cannot be assumed to be
phase-locked to the stimuli. Thus, the current estimates must be calculated
for each trial and then averaged in an appropriate manner. In principle it is
possible to extend the method of Jensen and Vanni (2002) in order to identify
sources of induced oscillatory activity, by applying a sliding time window to
each trial. The power values of the current estimates for each time window,
with respect to the same delay in relation to the stimuli, are then averaged.
Lin et al. (2004) have proposed such an approach based on the complex wave-
let transform. Like the Fourier transform, the complex wavelet transform
yields real and imaginary representations for which the current estimates can
be calculated. The current estimate’s real and imaginary parts are then summed,
averaged over trials, and represented as a function of time (Figure 7–8).
Lin et al. applied this method to MNEs, but it can also be used in conjunction
with MCEs. As always, when characterizing induced activity it is convenient
to study changes in power of induced oscillatory activity with respect to a
baseline period, since this attenuates noise due to slow drifts in signal power.
Lin et al. (2004) proposed to use an F–test to compare stimulus power to
baseline power in the current estimates, thus effectively producing a signal-
to-noise estimate of the induced activity. Only a few applied neuroscience
studies so far have been published using MNEs/MCEs to identify induced
oscillatory activity. However, given the growing interest in oscillatory activity,
more studies on this subject are bound to emerge in the near future.
Applications
Applications of Current Estimates to Event-Related Fields
A) phase response
imag.
real
1
PLV
+ +
T
extract phase
reference signal
baseline
baseline power
2
+ 1
+
T
2
F test
response
2
imag.
real
power
+ + 1
T
2
inverse
noise covariance
fMRI prior
CWT(f )
MEG sensor data
pre-stim
basline post-stim response
trial i
trigger
B)
x 10−8
40
2.5
35
Frequency (Hz)
2
30
1.5
25
20 1
15 0.5
10
−200 −100 0 100 200 300 400 500
Time (msec)
17 Hz noise-norm. power
Figure 7–8. Wavelets and MNE applied to study the temporal develop-
ment of induced oscillatory activity. (A) CWT denotes complex wavelet
transformations applied to the data. The MNEs are calculated for the real
and imaginary parts of the wavelet transformations and summed. The
current estimates are then averaged over trials and an F-test is applied to
normalize the data with respect to noise. (B) Early activity (40–100 ms) in
the beta band (~ 17 Hz) was identified using the wavelet approach and
MNE. The beta activity was induced by median nerve stimulation. The
resulting MNE is represented on the subject’s inflated brain. As expected,
the sources are localized around sensorimotor cortex.
Reproduced from Lin et al. (2004).
Estimating Distributed Representations of Evoked Responses 173
6 deg
12 deg
0.5
Scale / nAm
Fixation Trial Intertrial White=Left visual field stimulation
Black=Right visual field stimulation
Passive viewing
1–4 s
Request for response 3s
Stimulus
0
1.5 s 1.5 s
Time (s)
C)
Left visual field Right visual field
Sum amplitude (nAm)
3
Trial
Passive
0 0 400
0 400
Time (ms)
Figure 7–9. MCEs used to investigate how foveal attention modulates MEG responses to peripheral stimuli. (A) When the
luminance of the foveal fixation square changed, subjects were supposed to respond by button-press. The ERFs were calcu-
lated from the MEG data time-locked to peripheral squares flashing in the right or left visual fields. The ERFs were calculated
during passive viewing as well, and the intertrial intervals were used as control conditions. (B) Grand average of the MCEs
in the 100-160 ms time period. The white circles indicate the regions of interest where the attention effect was identified.
(C) The time course of the regions of interest shown for stimulation in the left and right visual fields. (D) The regions of inter-
est mapped onto the segmented brain of a single subject.
Reproduced from Vanni & Uutela (2000).
Estimating Distributed Representations of Evoked Responses 175
The second example is an MEG study in which the MNE rather than
MCE approach has been applied to investigate the N400m effect. The N400
component has been intensively studied using EEG, and refers to a negative
detection in the ERPs occurring ~400 ms post-stimulus. The N400 compo-
nent increases in magnitude in response to semantic violations (i.e., a word in
the sentence that does not make sense)—this is referred to as the N400 effect
(Kutas & Hillyard, 1980). Using MEG, the dominant sources of the N400
effect have been identified in the left and right superior temporal cortex by
dipole modeling (Helenius et al., 1998). MNE has also been applied in order
to identify the sources of the N400 effect (Halgren et al., 2002). The study was
designed as a standard N400 paradigm in which 240 sentences were presented
visually. Half of the sentences ended with a word that was semantically incon-
gruent with respect to the sentence context; the other half of the sentences
had congruent endings. Figure 7–10A shows a typical ERF from one sensor
over the left temporal region. Note the stronger N400m component for the
incongruent sentences. Field distributions are shown in Figure 7–10B after
subtracting the two conditions.4 The field distribution over the left hemisphere
is well described by a dipole in the superior temporal cortex. Additionally, the
MNE approach was applied to account for the N400 effect. The MNEs were
calculated in time-steps of 5 ms and converted to a dynamic statistical
parametric map (dSPM). These maps represent the MNE results in terms of
signal-to-noise ratios rather than baseline-subtracted current distributions
A B
50 fT/cm
250 ms
C Incongruous
- Congruous
Figure 7–10. The MNE applied to the N400 effect. (A) The ERFs for congru-
ent (black) and incongruent (red) sentence-endings for a single subject.
(B) A dipole model accounting for the subtracted ERFs for a single subject.
(C) The MNEs of the subtracted conditions represented as dSPMs. The
source distributions in each subject were mapped to inflated brain
representations, and morphed prior to grand averaging.
Reproduced from Halgren et al. (2002).
176 MEG: An Introduction to Methods
(for details see Liu et al., 1998; and Dale et al., 2000). The noise used in dSPM
is estimated from the measured data. As shown in Dale et al. (2000), the
MNEs become more focal after noise normalization and thresholding. Source
distributions were mapped to the segmented brain surface for the two condi-
tions subtracted. The brain surfaces were then morphed according to gyri and
sulci and a grand average over 8 subjects was performed (Figure 7–10C). At
about 250 ms, the first activity emerges in Wernicke’s area. It then spreads
and, after 300 ms, prefrontal activity can be observed, which eventually
includes Broca’s areas. It should also be mentioned that the threshold applied
to the dSPMs is somewhat arbitrary. In conclusion, the MNE approach
applied to the N400 effect allows for visualizing the spatiotemporal develop-
ment of event-related activity. Compared to the dipole model, MNE does
produce source distributions that are fairly smeared.
Power (fT/cm)2/Hz
14.7 17.6
20.0 ? nAm
?
0.12
0 0 0 0 0
100 300 50 200
14.1 13.0
22.3
Power (fT/cm)2/Hz
22.9
13.0
16.4
19.4 C)
22.3
0 0 0 0
10 20 30 40 10 20 30 40 10 20 30 40 10 20 30 40
Frequency (Hz)
pre-BNZ
post-BNZ
Figure 7–11. Sources accounting for the increase in beta oscillations with benzodiazepine. (A) The power spectra from sensors
over the central band before and after the administration of benzodiazepine. Data from 8 subjects. (B) The grand average
of the MCEs calculated in the frequency domain. (C) The sources of the beta activity mapped onto the MRI. Example from a
single subject.
Reproduced from Jensen et al. (2005).
178 MEG: An Introduction to Methods
the individual subjects. As seen in the example in Figure 7–11C, the beta
sources were located in the sensorimotor cortex around the central sulcus. In
conclusion, this study shows that the increase in beta-band oscillations with
benzodiazepine can be accounted for by sources in sensorimotor cortex. Prior
to performing the MCE analysis, it was essential to first calculate the power
spectra in each subject in order to identify the peak frequency of the beta
oscillations.
(b)
Frequency [Hz]
120
80
40
−0.5 0 1.0 −0.5 0 1.0 −0.5 0 1.0
Time [s] Time [s] Time [s]
0.1 sec 0.2 sec 0.3 sec 0.4 sec 0.5 sec
−20 20
Δ power [Z value]
(a)
(b) 0.20
Power [(fT/cm)2/Hz]
misses
hits
Rα
0.10
0 5 10 15 20 25
Frequency (Hz)
Posterior (Rα)
(c) 1.2
Normalized Hit Rates
1.1
1.0
p = 0.024
p = 0.001
0.9
0.8
1 2 3 4
Power [Quartiles]
based on constructing spatial filters for each grid point of the discretized brain
volume. These filters can also be used to extract the activity from individual
trials. This is done in Figure 7–13c, in which the signal trial, alpha power, was
derived for the parieto-occipital region. The trials were then sorted in four
bins according to alpha power, and the hit rates were calculated for each bin.
These findings confirm that hit rate decreases with an increased alpha power
produced in the parieto-occipital sulcus.
Estimating Distributed Representations of Evoked Responses 181
Practical Concerns
There are both commercial and free software packages available for calculat-
ing current estimates and beamformer. Below, find Table 7–1 summarizing
the various software packages. It should be mentioned that each package has
different features, making if difficult to directly compare the different meth-
ods. Nevertheless, when choosing a software package users are encouraged to
read publications related to the experiment(s) in which the actual the soft-
ware has been applied. This will provide insight into whether the software is
suitable to answer a given question.
Summary
In this chapter we have described several approaches for constructing distrib-
uted representations of source activity. A common advantage of the different
estimation methods is that the resulting distributed representations can be
morphed onto a standard brain, and subsequently averaged across subjects.
Open source
Neuromag
VSM-CTF
MNE (L2)
MCE (L1)
Elektra/
LCMV/
Loreta
DICS
SAM
Free
4-D
Fieldtrip X X X X X X X
Loreta X X
MNE- software X X X X X
NUTMEG X X X
SAM (VSM/CTF) X X
MCE (Elek/Neu) X X
BESA X X X X X
Curry X X X X X X
182 MEG: An Introduction to Methods
Notes
1 It is important to point out that in the context of beamforming, the terms
“spatial” and “location” do not in the first instance refer to the 3-dimen-
sional space of the brain volume and the points within. What is meant is the
N-dimensional measurement space defined by the sensor array, in which
the vector representations of the scalp field patterns of all sources within the
brain volume happen to define a point.
2 Here, the term “accuracy” does not exclusively refer to the quality of the
electromagnetic volume conductor model used to compute the columns of
the leadfield matrix G. The field patterns of spatially extended sources (e.g.,
locally or bilaterally synchronized neuronal activity), which effectively com-
prise dipole sources at several grid locations (and orientations) with essen-
tially the same—and hence highly correlated—activation time courses, are
not accurately described by the fields of source at individual grid locations.
Thus, spatially extended sources, in fact, violate both of the fundamental
assumptions underlying the beamforming approach.
3 This is not a trivial problem, since it requires not only a concrete and precise
definition of what the noise term refers to in Equation (7–1), and what are
its statistical properties, but also necessitates the use of ancillary methods for
estimating the (brain) signal and noise-subspace-component contributions
to the cross-covariance matrix. Poor choices or estimates of the noise model
invariably distort the amplitude estimates for neuronal sources.
4 Alternatively, the noise-normalized current distributions for the conditions
could have been calculated and then subtracted.
Estimating Distributed Representations of Evoked Responses 183
References
Bauer, M., Oostenveld, R., Peeters, M., & Fries, P. (2006). Tactile spatial attention
enhances gamma-band activity in somatosensory cortex and reduces low-frequency
activity in parieto-occipital areas. J Neurosci, 26, 490–501.
Brookes, M. J., Stevenson, C. M., Barnes, G. R., Hillebrand, A., Simpson, M. I.,
Francis, S. T., et al. (2007). Beamformer reconstruction of correlated sources using
a modified source model. Neuroimage, 34, 1454–1465.
Capon, J. (1969). High-resolution frequency-wavenumber spectrum analysis. Procee-
dings of the IEEE, 57, 1408–1418.
Challis, R. E., & Kitney, R. I. (1991). Biomedical signal processing (in four parts). Part
2. The frequency transforms and their inter-relationships. Med Biol Eng Comput,
29, 1–17.
Dale, A., & Sereno, M. (1993). Improved localization of cortical activity by combining
EEG and MEG with MRI cortical surface reconstruction: a linear approach. J Cogn
Neurosci, 5, 162–176.
Dale, A. M., Liu, A. K., Fischl, B. R., Buckner, R. L., Belliveau, J. W., Lewine, J. D.,
et al. (2000). Dynamic statistical parametric mapping: combining fMRI and MEG
for high-resolution imaging of cortical activity. Neuron, 26, 55–67.
Gaetz, W., & Cheyne, D. (2006). Localization of sensorimotor cortical rhythms indu-
ced by tactile stimulation using spatially filtered MEG. Neuroimage, 30, 899–908.
Gomez, J. F., & Thatcher, R. W. (2001). Frequency domain equivalence between
potentials and currents using LORETA. Int J Neurosci, 107, 161–171.
Gross, J., Kujala, J., Hämäläinen, M., Timmermann, L., Schnitzler, A., & Salmelin, R.
(2001). Dynamic imaging of coherent sources: Studying neural interactions in the
human brain. Proc Natl Acad Sci USA, 98, 694–699.
Halgren, E., Dhond, R P., Christensen, N., Van Petten, C., Marinkovic, K., Lewine, J. D.,
et al. ( 2002 ). N400-like magnetoencephalography responses modulated by
semantic context, word frequency, and lexical class in sentences. Neuroimage, 17,
1101–1116.
Hämäläinen, M. S., & Ilmoniemi, R. J. (1994). Interpreting magnetic fields of the
brain: minimum norm estimates. Med Biol Eng Comput, 32, 35–42.
Hari, R., & Salmelin, R. (1997). Human cortical oscillations: a neuromagnetic view
through the skull. Trends Neurosci, 20, 44–49.
Haykin, S. (2002). Adaptive filter theory. (4th ed.). Upper Saddle River, New Jersey:
Prentice Hall.
Helenius, P., Salmelin, R., Service, E., & Connolly, J. F. (1998). Distinct time courses
of word and context comprehension in the left temporal cortex. Brain, 121 (Pt 6),
1133–1142.
Hillebrand, A., & Barnes. G. R. (2005). Beamformer analysis of MEG data. Int Rev
Neurobiol, 68, 149–171.
Hillebrand, A., Singh, K. D., Holliday, I. E., Furlong, P. L., & Barnes, G. R. (2005). A new
approach to neuroimaging with magnetoencephalography. Hum Brain Mapp, 25,
199–211.
Ioannides, A., Bolton, J., & Clarke, C. (1990), Continuous probabilistic solutions to
the biomagnetic inverse problem. Inverse Problems, 6, 523–542.
Jensen, O., & Vanni, S. (2002). A new method to identify multiple sources of oscillatory
activity from magnetoencephalographic data. Neuroimage. 15. 568–574.
184 MEG: An Introduction to Methods
Jensen, O., Goel, P., Kopell, N., Pohja, M., Hari, R., & Ermentrout, B. (2005). On the
human sensorimotor-cortex beta rhythm: sources and modeling. Neuroimage, 26,
347–355.
Jokisch, D., & Jensen, O. (2007). Modulation of gamma and alpha activity during
a working memory task engaging the dorsal or ventral stream. J Neurosci, 27,
3244–3251.
Kalcher, J., & Pfurtscheller, G. (1995). Discrimination between phase-locked and
non-phase-locked event-related EEG activity. Electroencephalogr Clin Neurophysiol,
94, 381–384.
Kujala, J., Pammer, K., Cornelissen, P., Roebroeck, A., Formisano, E., & Salmelin, R.
(2007). Phase coupling in a cerebro-cerebellar network at 8-13 Hz during reading.
Cereb Cortex, 17, 1476–1485.
Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: brain potentials reflect
semantic incongruity. Science, 207, 203–205.
Liljestrom, M., Kujala, J., Jensen, O., & Salmelin, R (2005). Neuromagnetic locali-
zation of rhythmic activity in the human brain: a comparison of three methods.
Neuroimage, 25, 734–745.
Lin, F. H., Witzel, T., Hämäläinen, M. S., Dale, A. M., Belliveau, J. W., & Stuffle-
beam, S. M. (2004). Spectral spatiotemporal imaging of cortical oscillations and
interactions in the human brain. Neuroimage, 23, 582–595.
Liu, A. K., Belliveau, J. W., & Dale, A. M. (1998). Spatiotemporal imaging of human
brain activity using functional MRI constrained magnetoencephalography data:
Monte Carlo simulations. Proc Natl Acad Sci USA, 95, 8945–8950.
Lutkenhoner, B. (1992). Frequency-domain localization of intracerebral dipolar
sources. Electroencephalogr Clin Neurophysiol, 82, 112–118.
Matsuura, K. & Okabe, Y. (1995). Selective minimum-norm solution of the bioma-
gnetic inverse problem. IEEE Trans Biomed Eng, 42, 608–615.
Medendorp, W. P., Kramer, G. F., Jensen, O., Oostenveld, R., Schoffelen, J. M., & Fries, P.
(2006). Oscillatory activity in human parietal and occipital cortex shows hemisphe-
ric lateralization and memory effects in a delayed double-step saccade task. Cereb
Cortex. 17, 2364–2374.
Mitra, P. P., & Pesaran, B. (1999). Analysis of dynamic brain imaging data. Biophys J,
76, 691–708.
Niedermeyer, E., & Lopes da Silva, F. H. (1999). Electroencephalography: basic principles,
clinical applications, and related fields. (4th ed.). Baltimore: Williams & Wilkins.
Nieuwenhuis, I. L., Takashima, A., Oostenveld, R., Fernandez, G., & Jensen, O. (2008).
Visual areas become less engaged in associative recall following memory stabilization.
Neuroimage, 40, 1319–1327.
Osipova, D., Takashima, A., Oostenveld, R., Fernandez, G., Maris, E., & Jensen, O.
(2006). Theta and gamma oscillations predict encoding and retrieval of declarative
memory. J Neurosci, 26, 7523–7531.
Pascual-Marqui, R. D., Michel, C. M., & Lehmann, D. (1994). Low resolution electro-
magnetic tomography: a new method for localizing electrical activity in the brain.
Int J Psychophysiol, 18, 49–65.
Percival, D. B., & Walden, A. T. (1993). Spectral analysis for physical applications:
multitaper and conventional univariate techniques. Cambridge, UK: Cambridge
University Press.
Posthuma, D., Neale, M. C., Boomsma, D. I., & de Geus, E. J. (2001). Are smarter brains
running faster? Heritability of alpha peak frequency, IQ, and their interrelation.
Behav Genet, 31, 567–579.
Estimating Distributed Representations of Evoked Responses 185
Introduction
Independently, different brain imaging methods provide compromised
spatial and temporal resolutions. For example, anatomical MRI provides
highly accurate images of the individual brain anatomy, but does not convey
information about the dynamically changing patterns of brain activity. In
functional imaging, fMRI is temporally limited by the slow time course of the
hemodynamic response, but can provide a spatial sampling on a millimeter
scale (Belliveau et al., 1992; Kwong et al., 1992). EEG and MEG, in turn, pro-
vide a temporal resolution of milliseconds, but the localization of sources is
186
Anatomically and Functionally Constrained Minimum-Norm Estimates 187
To this end, the MCE implementation described in Uutela, et al. (1999) used
the current-source orientations provided by MNE.
The fusion of electromagnetic and hemodynamic data is still in its
infancy. In the presently available modeling methods, a distributed current
estimate confined to the cortical gray matter is usually employed, with a
stronger a priori weighting at locations with significant fMRI activity (Dale
et al., 2000; Liu et al., 1998). More elaborate methods that attempt to model
the two data sets jointly under a common framework are also emerging.
Furthermore, basic studies that aim at understanding the relationship between
the hemodynamic and electromagnetic signals are ongoing, and will eventu-
ally result in a physiologically motivated rather than a partly heuristic model
of the coupling between the hemodynamic and electromagnetic data.
In this chapter, we will first outline the overall MEG data-processing
workflow, with emphasis on source estimation and incorporation of anatom-
ical information. Thereafter, we provide an overview of analytical methods
needed in the computation of the minimum-norm solutions, including appli-
cation of minimum-norm solutions in the computation of time-frequency
representations in the source domain. Next, we will discuss a specific work-
flow to compute the cortically constrained, distributed source estimates,
including practical approaches to acquiring and processing the MRI and
MEG data. Finally, we will discuss a few representative studies where the
presented methods have been employed.
Theoretical Background
This section provides technical details of the analytical methods involved in
the computation of cortically constrained distributed source estimates. Our
discussion includes both the l2 minimum-norm estimates (MNE) and the
more focal l1 minimum-current estimates (MCE). We also apply the compu-
tational methods to frequency-domain analyses, in particular to the calcula-
tion of time-frequency representations and phase-locking values. While this
section contains some essential mathematical details of the algorithms, it is
not necessary for the reader to go into the depths of this section to understand
the workflow and the section titled Practical Implementation of Cortically
Constrained Estimates, below.
Y = AX + N, (8–1)
( ) ( )
−1 −1
X MNE = RA T ARA T + λ 2 C Y = λ −2 RA T λ −2 ARA T + C Y = WY, (8–2)
F } (8–3)
where Θ is the 3n-by-n matrix containing the unit vectors pointing to the
directions of the currents.
If the direction cosines of the kth dipole are ckx, cky, and ckz, the kth column
of Θ reads
Anatomically and Functionally Constrained Minimum-Norm Estimates 193
where
Y = C −1/ 2 Y
(8–7)
A = C −1/ 2 A
are the spatially whitened data and gain matrix, respectively.
The noise-covariance matrix of the whitened data is an identity matrix, as
indicated by the comparison of Equations (8–2) and (8–6). The whitening pro-
cedure also naturally leads to the choice λ2 = ζ2 tr(ÃRÃT)/m, where ζ2 denotes
the inverse of the SNR of the whitened data, to bring the regularization param-
eter to the correct scale even in cases where the measurements have different
units of measure—which is the case when planar gradiometer and magnetom-
eter data, or MEG and EEG data, are combined in a single estimate.
The sensitivity of MEG sensors is not uniform across cortical locations
(Hillebrand & Barnes, 2002). In fact, it follows generally from Maxwell’s
equations that the lead fields of both MEG sensors and EEG electrodes have a
maximum at the border of the source area (Heller & van Hulsteyn, 1992).
Because of the minimum-norm penalty, both MNE and MCE solutions are
biased to superficial locations, to which the sensors are most sensitive. It is
possible to compensate for this tendency by modifying our diagonal source-
covariance matrix R by scaling the entries by a function increasing monotoni-
cally with source depth, e.g., a norm of the corresponding column of the gain
matrix A (Fuchs et al., 1999; Ioannides et al., 1990; Lin, Witzel, et al., 2006).
Noise-Normalization
In the above, Equations (8–2) and (8–6) provide the best-fitting values of the
current amplitudes or, in Bayesian view, the maximum a posteriori (MAP)
estimate. To make the resulting maps conceptually similar with those calcu-
lated in other widely used functional imaging modalities (fMRI and PET),
Dale et al. (2000) proposed that current values should be converted into
dynamic statistical parametric maps. To this end, we need to consider the
variances of the currents
(
w k2 = WCWT ) = (WW
kk
) T
kk
. (8–8)
Note that under the null hypothesis, Fkp is F-distributed, with three
degrees of freedom for the numerator. The degrees of freedom for the denomi-
nator is typically large, again depending on the number of time samples used
to calculate the noise-covariance matrix.
As discussed in Dale et al. (2000), the noise-normalized estimates resulting
from the transformations given in Equations (8–8) and (8–9) have a smaller
depth bias than the MNEs obtained without applying depth weighting.
Furthermore, the point-spread function, i.e., the image of a point current
source, is more uniform in space in the noise-normalized estimate than in the
MNE.
Another variation of the noise-normalized MNE is the sLORETA
(Pascual-Marqui, 2002):
(
w k2 = W(C + λ −2 ARA T )WT ) = (W (I + λ
kk
−2 T )W
ARA T )
kk
(8–11)
From this expression it can be clearly seen that the sLORETA noise-
normalization factor differs from that given in Equation (8–8) by the addition
of the term λ−2ARAT to the noise-covariance matrix C. If the a priori informa-
tion incorporated in λ−2R is correct, C + λ−2ARAT = Cd, the data covariance
matrix. It has been shown that, in the absence of noise, sLORETA yields an
unbiased estimate of the location of the activity (Pascual-Marqui, 2002). In
our experience, under realistic noise conditions the difference between
sLORETA and dSPM is less dramatic than claimed, as also indicated by our
recent study on the depth biases of MNE and the two noise-normalized
estimates (see Lin, Witzel, et al. 2006).
It is also important to realize that the roles of the MNE and the noise-
normalized estimates are different. The former gives an estimate of the cur-
rent amplitudes in physical units [A/m], [A/m2] while the noise-normalized
distributions are test statistics to be employed as significance measures.
Optimally, dSPM or sLORETA and MNE distributions should be used in
combination so that the significance map is used to delineate the areas of
activity with high signal-to-noise ratio and, subsequently, MNE is consulted
for the true current amplitudes at that particular region.
Anatomically and Functionally Constrained Minimum-Norm Estimates 195
⎧n ⎫
X MCE
p = argmin ⎨∑ w X i , p ⎬
Xi , p ⎩ i =1 i
⎭ (8–12)
subject to Yrp = Br X p ,
MCE
where X MCE
p is the solution at time point p, wi are the weights of dipole
sources, while Ỹrp and Br are derived from the measurement data and the
forward solution for fixed-orientation sources to implement regularization as
follows.
As before, let Θ be the n-by-3 matrix containing the source orientations
and compute the singular-value decomposition
A Θ = UΠVT . (8–13)
Then
Br = U Tr A Θ,
(8–14)
Y = U T Y
r r
Nk
y k = ∑ w kp B(rkp ) ⋅ eˆkp , (8–15)
p =1
where wkp are weight factors specific for sensor k, rkp are the integration points,
êkp are the unit normal vectors describing the orientation of the coil at each of
the integration points, and $\vec B(\vec r_{kp})$ is the magnetic field gener-
ated by the current dipole of interest at $\vec r_{kp}$.
Anatomically and Functionally Constrained Minimum-Norm Estimates 197
fMRI Constraints
Time-frequency Analyses
Given the TFR on cortical surfaces, the following quantities can be calcu-
lated at each source location respectively: (1) averaged power, which is the
modulus squared of ψ(t, f) computed from the average of all trials; (2) induced
power, which is the average of the squared moduli of ψ(t, f) across trials; and
(3) Phase-Locking Value (PLV), which is a measure of phase difference con-
sistency across trials with respect to a reference signal: θ (t ) = e iϑk (t ) , where
ϑk(t) is the phase difference at time t in trial k (Lachaux et al., 1999) and ⋅
denotes the modulus of the average over trials. The reference signal can be a
waveform time-locked to the onset of the experimental stimulus during
stimulus-cortical PLV calculation, or a waveform on the other cortical source
location for cortico-cortical PLV calculations. As shown in Figure 8–1, fMRI
information can further be utilized to improve source-localization accuracy
of oscillatory activity (Lin et al., 2007; 2004).
real imag.
time
1
K
Extract phase
Reference
signal
Inverse
fMRI prior
Wavelet transform
MEG data
PLV
Trigger Trial k
Figure 8–1. A schematic diagram illustrating the process of using raw MEG
data to calculate the phase locking value (PLV) on the cortical surface
The yellow and red boxes on the MEG data indicate the baseline period
and the post-stimulus interval, respectively. Note that an fMRI prior can
be incorporated into the calculation of the linear inverse operator. For
details, see text.
of the necessary MRI and MEG data and on crucial computational steps
involved.
The use of cortically constrained source estimates has some implications for
recommended MEG data acquisition practices. Most importantly, reliable
estimation of the noise-covariance matrix requires that continuous raw data
are available, as will be discussed in Estimation of the Noise Covariance Matrix,
below. Some MEG systems provide a choice between truly continuous and
epoch-based data acquisition modes. We strongly recommend that the con-
tinuous acquisition mode is employed. This approach makes off-line digital
filtering straightforward, and allows the use of high-pass filters with low
corner frequencies.
We also recommend that a generous bandpass is used in the actual data
acquisition, since the data are invariably filtered and re-averaged off line, even
200 MEG: An Introduction to Methods
if on-line averages are usually computed for quality control and initial explor-
atory analyses. We usually employ a highpass at 0.1 Hz or lower in the data
acquisition. Depending on the system and environmental noise conditions,
a DC recording may even be feasible. Since the computation of the noise-
covariance matrix is a nonlinear operation, and thus does not commute with
filtering, the off-line averages and noise-covariance matrix estimate should be
computed with an identical bandpass. If the MEG system provides a choice of
storing the data with or without noise compensation applied, we have found
that it is useful to consistently save uncompensated data, and apply software
noise compensation during the analysis as appropriate.
For quality control purposes, optimization of the noise-rejection algorithms,
and for the computation of a noise-covariance matrix for spontaneous data
analyses, about a 5-minute continuous record of empty room data without
subject should be collected before each measurement session, or at least once
every day.
As described above, a feasible anatomical constraint for MEG and EEG source
localization is to restrict the source locations to the cortical mantle, extracted
from the individual subject’s MRI. With modern segmentation methods, a
representation for the geometry of the cortex can be automatically generated
from high-resolution 3D MRI data sets. We routinely employ the FreeSurfer
(Dale et al., 1999; Fischl et al., 2001; 1999) and BrainSuite (Shattuck & Leahy,
2002) software packages to build the triangular cortical surface mesh from
T1-weighted anatomical 3D-volume MRI data. For example, on the Siemens
MRI scanners, we employ the MPRAGE sequence and acquire two identical
data sets to enhance the SNR by averaging. For reconstruction of the inner
skull, outer skull, and scalp surfaces needed for the MEG and EEG forward
solutions, and surface-based alignment of the MEG and MRI coordinate
frames, we often acquire an additional 3D multi-echo FLASH sequence.
However, for MEG-only modeling the MPRAGE data are usually sufficient,
because the outer skull surface is not needed.
The principal surfaces generated by FreeSurfer and BrainSuite are the
pial surface and the gray-white matter boundary; we use the latter to generate
the cortically constrained source space. In addition to the folded surface, both
FreeSurfer and BrainSuite also compute inflated and flattened representa-
tions of the cortex, which expose the parts of the cortex embedded in the
sulci. These representations are thus particularly useful for visualizing MEG
data, which are mainly sensitive to fissural activity. FreeSurfer also provides
an automatic parcellation of the cortex, which can be useful in inquiring
source waveforms in specific regions of interest. An additional benefit of the
surface-based analysis is that cortical surfaces can be aligned across individuals
for computation of group statistics using a morphing procedure in a spherical
coordinate system (Fischl et al., 1999).
Anatomically and Functionally Constrained Minimum-Norm Estimates 201
In order to employ MRI and MEG data together, it is necessary to bring the
two data sets into a common coordinate frame. For this purpose, 3 – 5 small
head-position indicator (HPI) coils are usually attached on the head surface,
and their locations as well as additional head surface points are digitized prior
to the MEG acquisition in a coordinate frame defined by fiducial landmarks.
During the MEG measurement, current is fed to the coils either intermit-
tently (see, e.g., Fuchs et al., 1995) or continuously (Uutela et al., 2001), to
compute their locations with respect to the sensor array. If a continuous
head-position measurement is used, it is possible to take the head movements
202 MEG: An Introduction to Methods
10
1
0 50 100 150 200 250
Eigenvalue index
Examples
This section contains a few examples of experiments analyzed with the methods
and software described above.
MNE Software
MGH MIC UMN
dSPM estimates
Brainstorm/Brainsuite Software
MGH MIC UMN
dSPM Estimates
Figure 8–3. Field maps and source localizations for median nerve stimula-
tion at 20 ms after the onset of the stimulus. The top row shows field
maps for data collected on the same subject using MEG machines from
three different manufacturers, after projection into a common reference
frame. The second and third rows show comparisons of dSPM estimates,
obtained using two different software packages (MNE/FreeSurfer and
BrainStorm/BrainSuite). The estimates of activity across MEG systems and
software packages within the same subject are remarkably consistent.
suggested by animal models (Rauschecker & Tian, 2000; Tian et al., 2001).
This example demonstrates the potential of using anatomically and functionally
constrained source estimates to study the fine details of cortical processing of
sensory signals; see Figure 8–4.
During the experiments in Ahveninen et al. (2006), subjects were pre-
sented with pairs of Finnish vowels /æ/ and /ø/. Each vowel was simulated
from either straight ahead or 45 degrees to the right. The sound pairs were
identical, phonetically discordant (but spatially identical), or spatially discordant
(but phonetically identical). Cortically constrained MEG/fMRI minimum-
norm estimates of responses to “probes” preceded by identical, phonetically
different, or spatially different “adaptors” were compared. The result suggested
a double dissociation in response adaptation to sound pairs with phonetic vs.
208 MEG: An Introduction to Methods
Inflated left
hemisphere
Am x 10−10
1.8
Superior
Temporal Lobe
0.4
MEG/fMRI estimates at N1 peak
100 ms after Adaptor 100 ms after Probe
Posterior
Location
change
Anterior
Phoneme
change
PP
HG
PT
Identical
sound
STS STG
pair
a 0.08
(ms-1)2 / Hz
0.06
b
0.04
0.02 VM
R
0
02 5 10 15 20 25 30
Frequency (Hz) 0.3
c 0.5 0.2
VM
0.4 0.1
R
0.3 0
0.2 e
0.1
0
0 2 4 6 810 15 20 25 30
Frequency (Hz)
d 0.06
VM 0.4
R
(nAm)2 / Hz
0.04
0.2
0.02
0
0
02 5 10 15 20 25 30
Frequency (Hz)
Figure 8–5. Coherence between brain activity and hand speed. (a) Trackball
speed (TBS) power spectrum during visuomotor (VM – blue) and rest (R –
green) conditions. (b) Cortical map of task-related Z-transformed coherence
with TBS (VM vs. R) in the 2–5Hz range (P < 0.001, corrected). The white dot
indicates the location of maximum coherence difference (Montreal Neuro-
logical Institute coordinates: (−42, −17, 67 mm), hand area M1). (c) M1–TBS
coherence spectrum during VM and R, with a peak at 4 Hz. (d) M1 power
spectrum. Compared with R (green), VM (blue) has more power in 3–5 Hz
(P < 0.05, corrected), followed by the well known power suppression of
10–20-Hz oscillations. (e) Cortical map of difference in brain–TBS phase-
locking at 4 Hz (+/–1Hz) between VM and R (P < 0.001, corrected). The
white dot indicates the location of maximum phase-locking difference.
Using the MNE software, the Martinos Center Clinical MEG Service at
MGH has begun to evaluate the utility of cortically constrained distributed-
current estimates in the analysis of MEG and EEG data acquired from epileptic
patients (Knake et al., 2006; Shiraishi et al., 2005). With this technique, a
‘movie’ is generated that shows estimates of the brain activity over the entire
time course of an interictal event (see Figure 8–6). This allows a physician to
determine precisely were in the brain the epileptic spike originates, and focus
the treatment more specifically on the diseased tissue.
Anatomically and Functionally Constrained Minimum-Norm Estimates 211
Interictal Discharge
Anterio- Lt Rt
Left-Temporal 7292 ms
Anterio-Midline
Anterio- 7347 ms
Right-Temporal
Posterio-
Left-Temporal
Posterio-Midline 7419 ms
Posterio-
Right-Temporal
7293 ms 7635 ms
Fp1-F7
F7-T3
T3-T5
T5-O1 7635 ms
Fp2-F8
F8-T4
T4-T6
T6-O2
Fp1-F3
F3-C3
C3-P3
P3-O1
Fp2-F4
F4-C4
C4-P4
P4-O2
Fz-Cz 0.2 mV
Cz-Pz
ECG 1 sec
Figure 8–6. Interictal epileptiform discharges in MEG and EEG, with sources
estimated using equivalent current dipoles (ECDs) and dynamic statistical
parametric mapping (dSPM). Top left: Interictal epileptiform discharges
on magnetoencephalography were observed over the right temporal and
frontal regions widely. Bottom left: Right temporal dominant but right
hemispheric spikes or polyspikes could be detected frequently on EEG.
First row, middle and right: The equivalent current dipoles calculated by
the beginning of the polyspike burst located at the right temporal oper-
culum, some also in the lower operculum in the right frontal lobe. Second
to fifth row, middle and right: A dynamic statistical parametric mapping
showed the wide activity over the whole temporal lobe propagated to
the ipsilateral frontal and parietal lobes. The threshold of displayed activ-
ity is p < 10−4, and the full yellow area indicates p < 10−9.
Concluding Remarks
Within the last decade, anatomical MRI has become an integral and indis-
pensable component of MEG source analysis. MRI data are used not only as
an anatomical map in the visualization of results, but also as a source of geo-
metrical information for both forward and inverse modeling. In distributed
source-modeling approaches, the cortical geometry information is particularly
212 MEG: An Introduction to Methods
Acknowledgments
This work was supported by The Center for Functional Neuromaging
Technologies (NIH grant P41 RR14075), grants R01-EB002010 and 1R01-
EB009048-01 from the National Institute of Biomedical Imaging and
BioEngineering, Department of Energy Award Number DE-FG02-99ER62764
to The MIND Institute in Albuquerque, New Mexico, and by Los Alamos
National Security, LLC, for the National Nuclear Security Administration of
the U.S. Department of Energy under contract DE-AC52-06NA25396.
References
Ahveninen, J., Jaaskelainen, I. P., Raij, T., Bonmassar, G., Devore, S., Hämäläinen, M.,
et al. (2006). Task-modulated “what” and “where” pathways in human auditory
cortex. Proc Natl Acad Sci U S A, 103(39), 14608–14613.
Baillet, S., Mosher, J. C., & Leahy, R. M. (2001). Electromagnetic Brain Mapping. IEEE
Signal Processing Magazine, 18(6), 14 – 30.
Belliveau, J. W., Kwong, K. K., Kennedy, D. N., Baker, J. R., Stern, C. E., Benson, R., et
al. (1992). Magnetic resonance imaging mapping of brain function. Human visual
cortex. Invest Radiol, 27(2), S59–65.
Bertsekas, D. P. (2000). Dynamic programming and optimal control (2nd ed.).
Belmont, Mass: Athena Scientific.
Besl, P., & MacKay, N. (1992). A Method for Registration of 3-D Shapes. IEEE Trans.
Pat. Anal. and Mach. Intel., 14(2), 239–256.
Dale, A., & Sereno, M. (1993). Improved localization of cortical activity by combining
EEG and MEG with MRI cortical surface reconstruction: A linear approach. J. Cog.
Neurosci, 5, 162–176.
Dale, A. M., Fischl, B., & Sereno, M. I. (1999). Cortical surface-based analysis. I.
Segmentation and surface reconstruction. Neuroimage, 9(2), 179–194.
Dale, A. M., Liu, A. K., Fischl, B. R., Buckner, R. L., Belliveau, J. W., Lewine, J. D.,
et al. (2000). Dynamic statistical parametric mapping: combining fMRI and MEG
for high-resolution imaging of cortical activity. Neuron, 26(1), 55–67.
Daunizeau, J., Grova, C., Marrelec, G., Mattout, J., Jbabdi, S., Pélégrini-Issac, M., et al.
(2007). Symmetrical event-related EEG/fMRI information fusion in a variational
Bayesian framework. Neuroimage, 36(1), 69–87.
Anatomically and Functionally Constrained Minimum-Norm Estimates 213
David, O., Kiebel, S. J., Harrison, L. M., Mattout, J., Kilner, J. M., & Friston, K. J.
(2006). Dynamic causal modeling of evoked responses in EEG and MEG. Neuro-
image, 30(4), 1255–1272.
Devor, A., Dunn, A. K., Andermann, M. L., Ulbert, I., Boas, D. A., & Dale, A. M.
(2003). Coupling of total hemoglobin concentration, oxygenation, and neural acti-
vity in rat somatosensory cortex. Neuron, 39(2), 353–359.
Fischl, B., Liu, A., & Dale, A. M. (2001). Automated manifold surgery: constructing
geometrically accurate and topologically correct models of the human cerebral
cortex. IEEE Trans Med Imaging, 20(1), 70–80.
Fischl, B., Salat, D. H., van der Kouwe, A. J., Makris, N., Segonne, F., Quinn, B. T.,
et al. (2004). Sequence-independent segmentation of magnetic resonance images.
Neuroimage, 23, 1, S69–84.
Fischl, B., Sereno, M. I., & Dale, A. M. (1999). Cortical surface-based analysis. II:
Inflation, flattening, and a surface-based coordinate system. Neuroimage, 9(2),
195–207.
Fischl, B., Sereno, M. I., Tootell, R. B., & Dale, A. M. (1999). High-resolution inter-
subject averaging and a coordinate system for the cortical surface. Hum Brain
Mapp, 8(4), 272–284.
Fuchs, M., Wagner, M., Kohler, T., & Wischmann, H. A. (1999). Linear and nonlinear
current density reconstructions. J Clin Neurophysiol, 16(3), 267–295.
Fuchs, M., Wischmann, H. A., Wagner, M., & Kruger, J. (1995). Coordinate system
matching for neuromagnetic and morphological reconstruction overlay. IEEE
Trans Biomed Eng, 42(4), 416–420.
Hämäläinen, M., Hari, R., Ilmoniemi, R., Knuutila, J., & Lounasmaa, O. (1993).
Magnetoencephalography-theory, instrumentation, and application to noninva-
sive studies of the working human brain. Review of Modern Physics, 65, 413–497.
Hämäläinen, M., & Ilmoniemi, R. (1984). Interpreting measured magnetic fields of the
brain: estimates of curent distributions. Helsinki, Finland: Helsinki University of
Technology.
Heller, L., & van Hulsteyn, D. B. (1992). Brain stimulation using electromagnetic
sources: theoretical aspects. Biophys J, 63(1), 129–138.
Helmholtz, H. (1853). Ueber einige Gesetze der Vertheilung elektrischer Strome in
korperlichen Leitern, mit Anwendung auf die thierisch-elektrischen Versuche.
Ann. Phys. Chem., 89, 211–233, 353–377.
Hillebrand, A., & Barnes, G. R. (2002). A quantitative assessment of the sensitivity of
whole-head MEG to activity in the adult human cortex. Neuroimage, 16(3 Pt 1),
638–650.
Huang, M. X., Dale, A. M., Song, T., Halgren, E., Harrington, D. L., Podgorny, I.,
et al. (2006). Vector-based spatial-temporal minimum L1-norm solution for MEG.
Neuroimage, 31(3), 1025–1037.
Ioannides, A. A., Bolton, J. P., & Clarke, C. J. S. (1990). Continuous probabilistic
solutions to the biomagnetic inverse problem. Inverse Problem, 6(4), 523–542.
Jerbi, K., Lachaux, J. P., N’Diaye, K., Pantazis, D., Leahy, R. M., Garnero, L., et al.
(2007). Coherent neural representation of hand speed in humans revealed by MEG
imaging. Proc Natl Acad Sci U S A, 104(18), 7676–7681.
Knake, S., Halgren, E., Shiraishi, H., Hara, K., Hamer, H. M., Grant, P. E., et al. (2006).
The value of multichannel MEG and EEG in the presurgical evaluation of 70 epi-
lepsy patients. Epilepsy Res, 69(1), 80–86.
214 MEG: An Introduction to Methods
Kwong, K. K., Belliveau, J. W., Chesler, D. A., Goldberg, I. E., Weisskoff, R. M.,
Poncelet, B. P., et al. (1992). Dynamic magnetic resonance imaging of human
brain activity during primary sensory stimulation. Proc Natl Acad Sci U S A, 89(12),
5675–5679.
Lachaux, J. P., Rodriguez, E., Martinerie, J., & Varela, F. J. (1999). Measuring phase
synchrony in brain signals. Hum Brain Mapp, 8(4), 194–208.
Limpiti, T., Van Veen, B. D., & Wakai, R. T. (2006). Cortical patch basis model for
spatially extended neural activity. IEEE Trans Biomed Eng, 53(9), 1740–1754.
Lin, F.-H., Raij, T., Ahveninen, J., Ahlfors, S., Leuthold, A. C., Pantazis, D., et al.
(2007). Imaging of oscillatory cortical activity using combined MEG and fMRI.
Elsevier Internation Congress Series, 1300, 19–22.
Lin, F. H., Belliveau, J. W., Dale, A. M., & Hämäläinen, M. S. (2006). Distributed
current estimates using cortical orientation constraints. Hum Brain Mapp, 27(1),
1–13.
Lin, F. H., Witzel, T., Ahlfors, S. P., Stufflebeam, S. M., Belliveau, J. W., &
Hämäläinen, M. S. (2006). Assessing and improving the spatial accuracy in MEG
source localization by depth-weighted minimum-norm estimates. Neuroimage,
31(1), 160–171.
Lin, F. H., Witzel, T., Hämäläinen, M. S., Dale, A. M., Belliveau, J. W., & Stufflebeam,
S. M. (2004). Spectral spatiotemporal imaging of cortical oscillations and interac-
tions in the human brain. Neuroimage, 23(2), 582–595.
Liu, A. K., Belliveau, J. W., & Dale, A. M. (1998). Spatiotemporal imaging of human
brain activity using functional MRI constrained magnetoencephalography data:
Monte Carlo simulations. Proc Natl Acad Sci U S A, 95(15), 8945–8950.
Matsuura, K., & Okabe, Y. (1995). Selective minimum-norm solution of the bioma-
gnetic inverse problem. IEEE Trans Biomed Eng, 42(6), 608–615.
Moon, T. K., & Stirling, W. C. (2000). Mathematical methods and algorithms for
signal processing [xxxvi, 937] Upper Saddle River, NJ: Prentice Hall.
Mosher, J. C., Baillet, S., & Leahy, R. M. (2003). Equivalence of Linear Approaches
in Bioelectromagnetic Inverse Solutions, 2003 IEEE Workshop on Statistical Signal
Processing, St. Louis, Missouri, Sep 28 – Oct 1, 2003.
Okada, Y. C., Wu, J., & Kyuhou, S. (1997). Genesis of MEG signals in a mammalian
CNS structure. Electroencephalogr Clin Neurophysiol, 103(4), 474–485.
Ou, W., Hamalainen, M. S., & Golland, P. (2009). A distributed spatio-temporal EEG/
MEG inverse solver. Neuroimage, 44(3), 932-946.
Pascual-Marqui, R. D. (2002). Standardized low-resolution brain electromagnetic
tomography (sLORETA): technical details. Methods Find Exp Clin Pharmacol, 24
Suppl D, 5–12.
Rauschecker, J. P., & Tian, B. (2000). Mechanisms and streams for processing
of “what” and “where” in auditory cortex. Proc Natl Acad Sci U S A, 97(22), 11800–
11806.
Riera, J., Aubert, E., Iwata, K., Kawashima, R., Wan, X., & Ozaki, T. (2005). Fusing
EEG and fMRI based on a bottom-up model: inferring activation and effective con-
nectivity in neural masses. Philos Trans R Soc Lond B Biol Sci, 360(1457), 1025–1041.
Riera, J. J., Wan, X., Jimenez, J. C., & Kawashima, R. (2006). Nonlinear local electro-
vascular coupling. I: A theoretical model. Hum Brain Mapp, 27(11), 896–914.
Shattuck, D. W., & Leahy, R. M. (2002). BrainSuite: an automated cortical surface
identification tool. Med Image Anal, 6(2), 129–142.
Anatomically and Functionally Constrained Minimum-Norm Estimates 215
Shiraishi, H., Stufflebeam, S. M., Knake, S., Ahlfors, S. P., Sudo, A., Asahina, N., et al.
(2005). Dynamic statistical parametric mapping for analyzing the magnetoence-
phalographic epileptiform activity in patients with epilepsy. J Child Neurol, 20(4),
363–369.
Taulu, S., Simola, J., & Kajola, M. (2005). Applications of the Signal Space Separation
Method. IEEE Trans. Biomed. Eng., 53(9), 3359–3372.
Tesche, C. D., Uusitalo, M. A., Ilmoniemi, R. J., Huotilainen, M., Kajola, M., & Salo-
nen, O. (1995). Signal-space projections of MEG data characterize both distributed
and well-localized neuronal sources. Electroencephalogr Clin Neurophysiol, 95(3),
189–200.
Tian, B., Reser, D., Durham, A., Kustov, A., & Rauschecker, J. P. (2001). Functional
specialization in rhesus monkey auditory cortex. Science, 292(5515), 290–293.
Uusitalo, M. A., & Ilmoniemi, R. J. (1997). Signal-space projection method for
separating MEG or EEG into components. Med Biol Eng Comput, 35(2), 135–140.
Uutela, K., Hämäläinen, M., & Somersalo, E. (1999). Visualization of magnetoence-
phalographic data using minimum current estimates. Neuroimage, 10(2), 173–180.
Uutela, K., Taulu, S., & Hämäläinen, M. (2001). Detecting and correcting for head
movements in neuromagnetic measurements. Neuroimage, 14(6), 1424–1431.
Wagner, M., Köhler, T., Fuchs, M., & Kastner, J. (2000). An extended source model
for current density reconstructions. Paper presented at the Proceedings of the 12th
International Conference on Biomagnetism.
Weisend, M. P., Hanlon, F. M., Montano, R., Ahlfors, S. P., Leuthold, A. C., Mosher, J. C.,
et al. (2007). Paving the way for cross-site pooling of magnetoencephalography
(MEG) data. Elsevier International Congress Series, 1300, 615–618.
9
Noninvasive Functional Tomographic
Connectivity Analysis with
Magnetoencephalography
Joachim Gross, Jan Kujala, Riitta Salmelin, and Alfons Schnitzler
• The excellent temporal resolution of MEG can be used for the non-
invasive investigation of long-range interactions between brain areas.
This analysis requires statistically validated interaction measures and
localization techniques
• We describe several measures for the quantification of neural interac-
tions, their implementation and limitations
• Localization techniques are introduced and discussed with respect to
their applicability for MEG connectivity studies
• We describe the identification of significant connectivity effects by
using surrogate data and permutation techniques
• As an example, we present in detail the different analysis steps of one
particular connectivity study
Introduction
One possible mechanism for neural communication has attracted consider-
able interest in recent years. Neural synchronization, i.e., temporally precise
interactions between neural assemblies, may indicate functionally relevant
interactions between these assemblies (Singer, 1999; Bressler & Kelso, 2001;
Engel, Fries et al., 2001; Varela, Lachaux et al., 2001; Fries, 2005; Schnitzler &
Gross, 2005). Investigating long-range interactions requires simultaneous
216
Noninvasive Functional Tomographic Connectivity Analysis with MEG 217
Measures of Interactions
A large number of methods exist for the characterization of interdependen-
cies between two or more time series. The aim of the characterization is
a detailed description of the connection in terms of the time course and
frequency of interaction, the type of interaction (e.g., linear or nonlinear) and
possibly the direction of interaction. In this section we describe some depen-
dency measures that are most frequently used in the analysis of functional
interactions based on electrophysiological recordings. We focus on measures
in the frequency domain.
One way to classify functional connectivity measures is the distinction
between parametric and nonparametric, and between linear and nonlinear
measures. Nonparametric techniques estimate dependency measures directly
from the data. They usually employ the fast Fourier transform (FFT), the
wavelet transform, or the Hilbert transform. In contrast, parametric tech-
niques fit a model to the data. The estimation of interactions between two
218 MEG: An Introduction to Methods
time series uses the parameters of the model, and thus relies on the accuracy
of the estimated model parameters.
Linear dependency measures assume that the output of the system under
investigation is linearly related to the input (scaling the input by a factor A
leads to an output scaled by the same factor). Although nonlinear interactions
have been demonstrated in electrophysiological recordings (e.g., Breakspear,
2002; Breakspear & Terry, 2002; Stam, Breakspear et al., 2003) linear mea-
sures are still widely used. Their advantage is a robust implementation
and fast computation, but they suffer from their insensitivity to nonlinear
dependencies in the data. In contrast, nonlinear measures can reveal some
nonlinear dependencies, but often rely on additional parameters that must be
specified by the user.
Phase Synchronization
A) Coherence
0.2
0.1
0
0 10 20 30 40
B) Filtering Frequency (Hz)
C) Phase computation
210 330
240 300
270
dynamics of one signal is influenced by the phase dynamics of the other signal.
The DI is normalized such that -1 and 1 represent an unidirectional effect
(to and from the first time series, respectively) and 0 indicates a symmetric
bidirectional effect.
Statistical Considerations
data, except the property that is studied. We can illustrate the flexible use of
surrogate data for coherence spectra. Artificial peaks in coherence spectra due
to stable oscillatory signals (e.g., line noise) can be identified by shifting one
time series relative to the other, e.g., by 1–2 seconds (Figure 9–3A,B).
Coherences from physiological processes are most likely destroyed by the
time-shift operation, whereas the artificial coherence peak should not
change if the underlying oscillation is stable (Figure 9–3B). To destroy syn-
chronization and oscillations in the time series, both time series can be inde-
pendently permuted (i.e., the order of the data points is randomly changed;
see Figure 9–3C). Applying the same permutation to both time series destroys
oscillations but preserves non-oscillatory dependencies between the time
series (Figure 9–3D). This particular approach is interesting for the validation
of cerebro-cerebral coherence spectra. Depending on the quality of the data,
the signal-to-noise ratio (SNR), and the possible presence of artifacts, the
coherence spectrum may show a substantial offset (an artificial shift of the
spectrum across all frequencies to higher coherence values). The offset is
accounted for by the surrogate data, and allows the identification of signifi-
cant coherence despite the offset. Another relevant type of surrogate data is
computed by randomly changing the phase of a signal (for each frequency)
without changing the amplitude (Figure 9–3E). The surrogate data has the
same power spectrum as the original data (Figure 9–3F) but any coherence
between the data is destroyed.
How can we use the methods introduced above to identify a network of
interacting brain areas? Three strategies may be employed. First, selection of
individual anatomical brain areas based on a priori information (e.g., from
other studies). Further analysis is needed to establish interactions between these
areas. Second, areas are selected based on their activity and subsequently tested
for interactions. Third, areas may be selected directly, based on their functional
connectivity to other areas. As the first strategy does not require any computa-
tion, we describe the second and third strategies in the following paragraphs.
0.8 0.8
Coherence
Coherence
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Frequency (Hz) Frequency (Hz)
C 1 D 1
0.8 0.8
Coherence
0.6 0.6
Coherence
0.4 0.4
0.2 0.2
0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Frequency (Hz) Frequency (Hz)
1
E F
0.8
0
Coherence
0.6
Power (dB)
−5
0.4
0.2 −10
0 −15
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Frequency (Hz) Frequency (Hz)
Figure 9–3. Surrogate data for coherence. Two time signals were simu-
lated with a broadband component around 10 Hz and a 50-Hz sinusoid,
representing line noise. (A) The original coherence spectrum. (B) Coher-
ence computed with one time series shifted by 1 s. The broadband
nonstationary component (representing physiological coherence) has
vanished, while the 50-Hz component is not affected. (C) Coherence
between the two time series, after the order of data points has been ran-
domly changed independently for each time series (independent permu-
tation). Physiological and artificial coherence is destroyed. (D) Coherence
between the time series after the order of data points has been changed
in exactly the same way for both time series (same permutation). Again,
physiological and artificial coherence is destroyed. (E) Coherence spec-
trum after random phases have been added to one time series. (F) Power
spectrum of the original time series (solid line) and the same time series
after phase randomization (dashed line) indicating that the phase ran-
domization does not affect the power spectrum.
224
Noninvasive Functional Tomographic Connectivity Analysis with MEG 225
Minimum-norm Techniques
Spatial Filter
Dipole Models
Dynamic Imaging of Coherent Sources (Gross, Kujala et al., 2001) allows the
tomographic mapping of both power and coherence in the entire brain, using
spatial filtering (Robinson & Vrba 1997; Sekihara, Nagarajan et al. 2001a;
Sekihara, Nagarajan et al., 2002a; Van Veen, van Drongelen et al., 1997) in the
frequency domain (Figure 9–4).
The output of the spatial filter is a linear combination of the adequately
weighted channels of the data matrix. Spatial filters can be designed according
to very different objectives depending on the aim of the analysis. For source
localization, the LCMV (linearly constrained minimum variance) beam-
former is often used. Here, the set of coefficients or weights is computed as
the solution of a constraint-minimization problem: the minimization of out-
put power subject to the constraint that activity from the region of interest is
passed with unit gain. The coefficients depend on the solution of the forward
problem for the region of interest, and the covariance matrix of the data. The
linear combination of the weighted channels acts like a spatial filter that leaves
Noninvasive Functional Tomographic Connectivity Analysis with MEG 227
10 45
Coherence (%)
Power (a.u.)
8
35
6
4 25
2
15
10 20 30 10 20 30
Frequency (Hz) Frequency (Hz)
Figure 9–4. Power and coherence mapping with DICS. Uniformly sized
volume elements (voxels) are defined to cover the entire brain based on
magnetic resonance images (MRI). The sensor information is used to com-
pute the spatial distribution of power (A) or coherence (B) in a predefined
frequency band (bottom row).
the signal from the region of interest unchanged and attenuates, as much as
possible, signals originating elsewhere.
The filter has to be computed for each region of interest. Indeed, to create
a tomographic map of power, the spatial filter must be computed for each
voxel on a regular 3-dimensional grid covering the entire brain. Since syn-
chronization is often frequency-dependent, DICS has been developed as a
frequency-domain implementation of a spatial filter that allows the tomo-
graphic mapping of power and coherence in a predefined frequency band.
We employ the cross-spectral density matrix as the basic representation
of the oscillatory components and their dependencies between MEG and
possibly additional signals. For continuous data, the complex cross-spectral
density P for signals x(t), y(t) is computed using Welch’s method of spectral
density estimation (Welch, 1967; see also Figure 9–1). For trial-based data,
P should be computed using wavelets or a moving FFT window. One element
Pi,j of the cross-spectral matrix consists of the cross-spectral densities of
signals i and j. Therefore, P contains the cross-spectral densities of all combi-
nations of signals.
228 MEG: An Introduction to Methods
Two measures are derived from this matrix. First, the power spectrum of
the signal i is represented by the diagonal element Pi,i. It allows identification
of frequency bands containing most of the power, or those showing task-
dependent changes. Second, coherence is the magnitude-squared cross spec-
trum divided by the power spectra of both time series.
To obtain a tomographic map of power or coherence, the spatial filter
coefficients are computed sequentially for each voxel. Based on the matrix
P and the coefficients, the cross spectrum can be computed between any two
voxels. If the same two voxels are chosen, the resulting map will show the
spatial distribution of power. Alternatively, coherence between a reference
voxel and all other voxels in the brain can be mapped based on the cross spec-
trum of the reference voxel and all other voxels, and power of both voxels.
Except DICS, all the methods introduced in the previous section identify
regions of interest based on their activation (maybe in a particular frequency
band). DICS also supports the computation of maps representing oscillatory
power—but, in addition, it can directly map coherence to a reference area
(i.e., the third strategy listed in the section on statistical considerations).
The functioning of a spatial filter can be illustrated in a simplified scenario
(Gross, Timmermann et al., 2003). We assume that a single brain area is
active. The frequency-dependent power estimate at any given point can be
computed as follows:
−1
P (r ,f ) = ⎡⎣ LT (r )C (f )−1 L(r )⎤⎦ (9–1)
where S1 represents the first singular value of C and S2 the second singular
value. (I-Ms) acts as a projection to the noise space.
Now the mechanism of the spatial filter becomes apparent (see Figure 9–5).
For this illustration we arbitrarily assume S1= 10 and S2= 1. At the true source
location (marked by +) the corresponding leadfield is in the signal space. For
the first addend in equation 9–2 we get 1/10 (reciprocal value of S1). In the
second addend the projection to the noise space is zero (since L(r) is entirely
in the signal space). Consequently, we obtain a value of P=10. For a point
close to the true source location (marked by *) the corresponding leadfield is
not entirely in the source space but has a component in the noise space. Thus,
the first addend becomes smaller (e.g., 0.9/10) and the second addend takes a
Noninvasive Functional Tomographic Connectivity Analysis with MEG 229
10
8
Power (a.u.)
6
0
−1 −0.5 0.0 0.5 1
x (cm)
Figure 9–5. Illustration of spatial filter. At each location along the x–axis
the power of a point source located at 0 is displayed as computed from a
spatial filter directed to this position. The + sign marks the true location
for which the solution of the forward problem is entirely in the signal
space. The * marks a point for which the solution of the forward problem
has a component in noise-space leading to a reduced power estimate.
nonzero value (e.g., 0.1). The resulting value is P=5.3 (Figure 9–5). As we
move away from the true source location, the estimated power decreases. The
steepness of the decrease depends on the signal-to-noise ratio (SNR, here the
ratio S1/S2) and the change of the leadfields. This underlines the importance
of the SNR for the spatial resolution of beamformer tomographic maps.
Efforts to increase the SNR during measurement and analysis (e.g., by using
an optimal signal representation) would be rewarded by an increased spatial
resolution.
et al., 2000; Van Veen, van Drongelen et al., 1997; Vrba & Robinson, 2001).
Preferably, methods should be used that account for the multiple-comparison
problem. Performing a statistical test for each voxel (or surface node) results
in an increase of the possible numbers of false positive results. Consequently,
the statistical procedure should account for the number of tests. The simplest
method is the Bonferroni correction that can be performed by multiplying
the resulting p-values by the number of tests. For spatially correlated maps
(as in our case) with many elements, this method is not effective and too
conservative. Two approaches may be used that are more efficient.
First, random field theory has successfully been applied for the corre-
sponding problem of statistically analyzing fMRI (functional magnetic reso-
nance imaging) or PET (positron emission tomography) data (Worsley &
Friston, 1995). It has recently been adapted for distributed source representa-
tions from MEG/EEG recordings (e.g., Barnes & Hillebrand, 2003; Carbonell,
Galan et al., 2004; Pantazis, Nichols et al., 2005); see also Kiebel, Tallon-
Baudry et al. (2005). The approach from Carbonell, Galan et al. (2004) offers
a unified statistical framework for the identification of significant compo-
nents in topographic and tomographic representations of the data. Thus, the
method can be applied to the data recorded by the sensors, and the data
linearly mapped into the brain to create maps of brain activity. Barnes and
coworkers (Barnes & Hillebrand, 2003) employed random field theory to
specifically address the multiple-comparison problem on tomographic maps
created by spatial filtering. A limitation of random field theory for MEG is
dependence on the assumptions of sufficient smoothness, gaussianity, and
stationarity of the underlying data.
Second, permutation methods may be used to extract significant regions
of interest while correcting for multiple tests. Similarly to random field the-
ory, permutation tests have been applied to functional magnetic resonance
data (Nichols & Holmes, 2002). Permutation methods rely on very few
assumptions, which are illustrated in the following example. Let us assume
that MEG data was continuously recorded during two experimental condi-
tions, A (e.g., rest) and B (e.g., continuous finger movements). We want to
know which areas in the brain show a significantly different activation between
the two conditions. Specifically, the null hypothesis is that a change of the
experimental condition has no effect on the activation of brain areas. Any of
the described localization procedures can be used to compute a distributed
source representation for condition A and B. Statistics can then be computed
that characterize the difference between A and B. As preparation for the per-
mutation approach, each data set is split into a number of segments of equal
length.2 Before carrying out the localization, the data segments from A and B
are randomly exchanged. Thus, the first and the second localization is now
performed on a random set of data segments from both conditions.
Subsequently, the same test statistics (e.g., relative difference) is evaluated on
both functional maps. The random exchange of segments, localization, and
evaluation of the test statistics, is repeated a large number of times. For each
232 MEG: An Introduction to Methods
repetition the maximum of the test statistics is identified. This results in the
distribution of the maximum statistics. The critical threshold corresponding
to a given significance level α (e.g., 0.05) can be identified as the (α*N)+1
largest element of the distribution where N is the number of repetitions.
It can be shown that the use of the maximum statistics provides a strong
protection against Type 1 error, i.e., rejecting the null hypothesis when it is
true (Nichols & Holmes, 2002).
Permutation methods make no assumptions on the distribution of the
data (and can thus be applied to activation or connectivity maps) but they
require the exchangeability of the segments.
Using permutation methods with the maximum statistics on real data
may lead to problems if the null distribution is too variable across voxels.
If the simple difference between two conditions is used, the maximum statis-
tics may be dominated by some voxels with a large variability in the mean
difference. This leads to a reduced sensitivity for other voxels. When that is
the case, other statistics (such as t-statistics) that lead to more homogeneous
distributions across voxels should be used (Nichols & Holmes, 2002; Pantazis,
Nichols et al., 2005).
Pantazis and coworkers performed a comparison of random field theory
and permutation methods for functional maps obtained from MEG data
(Pantazis, Nichols et al., 2005). For simulated data both methods showed
valid results, although the method based on random field theory demon-
strated conservative performance and a dependence on the smoothness of the
data. Differences in real data that were observed between the methods may
have been caused by violations of the distribution assumption of random
field theory.
Permutation statistics for MEG data have been used by several groups in
various ways (e.g., Chau, McIntosh et al., 2004; Greenblatt & Pflieger, 2004;
Gross, Schmitz et al., 2004; Singh, Barnes et al., 2003) that demonstrate the
flexibility of permutation methods. They can be applied to test for significant
differences between two conditions for single subjects or for a group of sub-
jects, or to test for significant differences between two groups of subjects for
one condition.
Permutation methods are implemented in SnPM (www.sph.umich.edu/
ni-stat/SnPM/; Nichols & Holmes, 2002). This particular implementation has
been used for the analysis of MEG tomographic maps (Gross, Schmitz et al.
2004; Singh, Barnes et al., 2003). Another noncommercial Matlab toolbox
that implements permutation methods is Fieldtrip (for web page, see above).
An interesting approach that should be mentioned in this context is the
application of partial least squares (PLS) analysis on tomographic maps (www.
rotman-baycrest.on.ca/index.php?section=84) (McIntosh & Lobaugh, 2004).
PLS is a multivariate method that performs a singular value decomposition of
the covariance of sets of variables. The method is very flexible, since the vari-
ables may contain MEG signals, spectra, behavioral results like reaction times,
tomographic maps, or other data from different conditions and subjects.
Noninvasive Functional Tomographic Connectivity Analysis with MEG 233
Paradigm
A 15
Power (a.u.)
10
0 5 10
B
15
Power (a.u.)
10
0 5 10
C
0.3
Coherence
0.2
0.1
0 5 10
Frequency (Hz)
235
236 MEG: An Introduction to Methods
Peripheral Signals
First, power and coherence spectra were computed for muscle activity and
finger velocity for all subjects. As expected from the previous studies, regular
changes in finger movements were seen at about 8 Hz in power spectra of
muscle activity and finger velocity, and also in coherence between those sig-
nals (Figure 9–7). The frequency of these regular changes varied across sub-
jects in the range 6–9 Hz. For each individual subject, the exact
frequency of maximum power in the finger velocity was determined for the
subsequent analysis. The findings from this analysis step confirmed the previ-
ously published results (e.g., Vallbo & Wessberg, 1993).
Cerebro-muscular Coherence
The aim of the next analysis step was the identification of possible cerebro-
muscular coherence. To this end, coherence between all MEG signals and the
muscle activity was computed for each subject. Figure 9–8 shows the results
B a b c
Figure 9–9. Localization results for a single subject. (A). DICS was used to
compute a functional map of cerebro-muscular coherence. For each voxel
in the brain, coherence of activity in this voxel to muscle activity is com-
puted, and coherence at 8 Hz is used for color coding. Maximum coherence
is evident in the contralateral sensorimotor cortex. (B). Coherence between
the contralateral primary motor cortex and all other voxels in the brain
was computed with DICS. Local maxima in the coherence map were found
in ipsilateral cerebellum (a), contralateral thalamus (b) and contralateral
premotor cortex (c).
238 MEG: An Introduction to Methods
Cerebro-cerebral Coherence
The significance of these results does not lie in the identification of the
cerebello-thalamo-cortical loop (which has been described before) but in the
functional characterization of the loop in the current paradigm. The study
showed for the first time an oscillatory interaction at 6–9 Hz within this loop,
and its relation to finger velocity changes at the same frequency. The oscilla-
tory interplay within the network may well implement discrete movement
control (Gross, Timmermann et al., 2002).
In this particular study, network identification as a critical part of the
functional connectivity analysis was performed using group statistics of coher-
ence maps. A similar approach has been used in other studies (Pollok, Gross
et al., 2004; Sudmeyer, Pollok et al.,. 2004; Timmermann, Gross et al., 2003).
Notes
1 Please note that this list is not complete.
2 For localizations in the frequency domain, Fourier-transformed data
segments can be used.
References
Amor, F., Rudrauf, D., et al. (2005). Imaging brain synchrony at high spatio-temporal
resolution: application to MEG signals during absence seizures. Signal Processing
85, 2101–2111.
Andrew, C., & Pfurtscheller, G. (1996). Event-related coherence as a tool for studying
dynamic interaction of brain regions. Electroencephalogr Clin Neurophysiol 98(2),
144–148.
Astolfi, L., & Cincotti, F., et al. (2005). Assessing cortical functional connectivity by
linear inverse estimation and directed transfer function: simulations and applica-
tion to real data. Clin Neurophysiol 116(4), 920–932.
Baccala, L. A., & Sameshima, K. (2001). Partial directed coherence: a new concept in
neural structure determination. Biol Cybern, 84(6), 463–474.
Barnes, G. R., & Hillebrand, A. (2003). Statistical flattening of MEG beamformer
images. Hum Brain Mapp 18(1), 1–12.
Bezruchko, B., Ponomarenko, V., et al. (2003). Characterizing direction of coupling
from experimental observations. Chaos, 13(1), 179–184.
Breakspear, M. (2002). Nonlinear phase desynchronization in human electroence-
phalographic data. Hum Brain Mapp, 15(3), 175–198.
Breakspear, M., & Terry, J. R. (2002). Nonlinear interdependence in neural systems:
motivation, theory, and relevance. Int J Neurosci 112(10), 1263–1284.
Noninvasive Functional Tomographic Connectivity Analysis with MEG 241
Bressler, S. L., & Kelso, J. A. (2001). Cortical coordination dynamics and cognition.
Trends Cogn Sci 5(1), 26–36.
Brovelli, A., & Ding, M., et al. (2004). Beta oscillations in a large-scale sensorimotor
cortical network: directional influences revealed by Granger causality. Proc Natl
Acad Sci USA 101(26), 9849–9854.
Buchel, C., & Friston, K. J. (1997). Modulation of connectivity in visual pathways by
attention: cortical interactions evaluated with structural equation modelling and
fMRI. Cereb Cortex 7(8), 768–778.
Carbonell, F., Galan, L., et al. (2004). Random field-union intersection tests for EEG/
MEG imaging. Neuroimage, 22(1), 268–276.
Chau, W., McIntosh, A. R., et al. (2004). Improving permutation test power for group
analysis of spatially filtered MEG data. Neuroimage, 23(3), 983–996.
Chen, Y., Bressler, S. L., et al. (2006). Frequency decomposition of conditional Gran-
ger causality and application to multivariate neural field potential data. J Neurosci
Methods 150(2), 228–237.
Dahlhaus, R. (2000). Graphical interaction models for multivariate time series.
Metrika, 51, 157–172.
Dahlhaus, R., Eichler, M., et al. (1997). Identification of synaptic connections in
neural ensembles by graphical models. J Neurosci Methods, 77(1), 93–107.
Dale, A., Halgren, E., et al. (2000). Spatiotemporal cortical activation patterns during
semantic processing of novel and repeated words as revealed by combined fMRI
and MEG. Neuron, 26, 55–67.
David, O., Cosmelli, D., et al. (2004). Evaluation of different measures of functional
connectivity using a neural mass model. Neuroimage, 21(2), 659–673.
David, O., Cosmelli, D., et al. (2003). A multitrial analysis for revealing significant
corticocortical networks in magnetoencephalography and electroencephalography.
Neuroimage, 20(1), 186–201.
David, O., Garnero, L., et al. (2002). Estimation of neural dynamics from MEG/
EEG cortical current density maps: application to the reconstruction of large-scale
cortical synchrony. IEEE Trans Biomed Eng, 49(9), 975–987.
David, O., Harrison, L., et al. (2005). Modelling event-related responses in the brain.
Neuroimage, 25(3), 756–770.
Ding, M., Bressler, S. L., et al. (2000). Short-window spectral analysis of cortical event–
related potentials by adaptive multivariate autoregressive modeling: data prepro-
cessing, model validation, and variability assessment. Biol Cybern 83(1), 35–45.
Eichler, M., Dahlhaus, R., et al. (2003). Partial correlation analysis for the identification
of synaptic connections. Biol Cybern 89(4), 289–302.
Engel, A. K., Fries, P., et al. (2001). Dynamic predictions: oscillations and synchrony
in top-down processing. Nat Rev Neurosci 2(10), 704–716.
Fries, P. (2005). A mechanism for cognitive dynamics: neuronal communication
through neuronal coherence. Trends Cogn Sci, 9(10), 474–480.
Friston, K. J., Harrison, L., et al. (2003). Dynamic causal modelling. Neuroimage,
19(4), 1273–1302.
Grasman, R. P., Huizenga, H. M., et al. (2004). Frequency domain simultaneous
source and source coherence estimation with an application to MEG. IEEE Trans
Biomed Eng 51(1), 45–55.
Greenblatt, R. E. & Pflieger, M. E. (2004). Randomization-based hypothesis testing
from event-related data. Brain Topogr, 16(4), 225–232.
Gross, J., Kujala, J., et al. (2001). Dynamic imaging of coherent sources: Studying
neural interactions in the human brain. Proc Natl Acad Sci U S A 98(2), 694–699.
242 MEG: An Introduction to Methods
Gross, J., Schmitz, F., et al. (2004). Modulation of long-range neural synchrony
reflects temporal limitations of visual attention in humans. Proc Natl Acad Sci
U S A, 101(35), 13050–13055.
Gross, J., Tass, P., et al. (2000). Cortico-muscular synchronization during isometric
muscle contraction in humans as revealed by magnetoencephalography. J Physiol
(Lond) 527(3), 623–631.
Gross, J., L. Timmermann, et al. (2002). “The neural basis of intermittent motor
control in humans.” Proc Natl Acad Sci U S A 99(4): 2299–2302.
Gross, J., Timmermann, L., et al. (2003). Properties of MEG tomographic maps
obtained with spatial filtering. Neuroimage 19(4), 1329–1336.
Hadjipapas, A., Hillebrand, A., et al. (2005). Assessing interactions of linear and
nonlinear neuronal sources using MEG beamformers: a proof of concept. Clin
Neurophysiol 116(6), 1300–1313.
Halliday, D. M., Rosenberg, J. R., et al. (1995). A framework for the analysis of mixed
time series/point process data–theory and application to the study of physiological
tremor, single motor unit discharges and electromyograms. Prog Biophys Mol Biol
64(2–3), 237–278.
Hoechstetter, K., Bornfleth, H., et al. (2004). BESA source coherence: a new method
to study cortical oscillatory coupling. Brain Topogr, 16(4), 233–238.
Hurtado, J. M., Rubchinsky, L. L., et al. (2004). Statistical method for detection of
phase-locking episodes in neural oscillations. J Neurophysiol, 91(4), 1883–1898.
Ioannides, A. A., Bolton, J. P. R., et al. (1990). Continous probabilistic solutions to the
biomagnetic inverse problem. Inverse Problems, 6, 523–542.
Ioannides, A. A., Liu, L. C., et al. (2000). Coupling of regional activations in a human
brain during an object and face affect recognition task. Hum Brain Mapp 11(2),
77–92.
Ioannides, A. A., Poghosyan, V., et al. (2004). Real-time neural activity and connecti-
vity in healthy individuals and schizophrenia patients. Neuroimage, 23(2), 473–482.
Kaminski, M., Ding, M., et al. (2001). Evaluating causal relations in neural systems:
granger causality, directed transfer function and statistical assessment of signi-
ficance. Biol Cybern, 85(2), 145–157.
Kaminski, M. & Liang, H. (2005). Causal influence: advances in neurosignal analysis.
Crit Rev Biomed Eng, 33(4), 347–430.
Kiebel, S. J., Tallon-Baudry, C., et al. (2005). Parametric analysis of oscillatory activity
as measured with EEG/MEG. Hum Brain Mapp, 26(3), 170–177.
Lachaux, J. P., Rodriguez, E., et al. (1999). Measuring phase synchrony in brain signals.
Hum Brain Mapp 8(4), 194–208.
Le Van Quyen, M., Foucher, J., et al. (2001). Comparison of Hilbert transform and
wavelet methods for the analysis of neuronal synchrony. J Neurosci Methods 111(2),
83–98.
Lin, F. H., Witzel, T., et al. (2004). Spectral spatiotemporal imaging of cortical oscilla-
tions and interactions in the human brain. Neuroimage, 23(2), 582–595.
Logothetis, N. K. (2003). MR imaging in the non-human primate: studies of function
and of dynamic connectivity. Curr Opin Neurobiol, 13(5), 630–642.
Makeig, S., Debener, S., et al. (2004). Mining event-related brain dynamics. Trends
Cogn Sci, 8(5), 204–210.
Makeig, S., Delorme, A., et al. (2004). Electroencephalographic brain dynamics
following manually responded visual targets. PLoS Biol, 2(6), e176.
McIntosh, A. R. & Lobaugh, N. J. (2004). Partial least squares analysis of neuroima-
ging data: applications and advances. Neuroimage, 23 Suppl 1, S250–63.
Noninvasive Functional Tomographic Connectivity Analysis with MEG 243
Mechelli, A., Price, C. J., et al. (2003). A dynamic causal modeling study on category
effects: bottom-up or top-down mediation? J Cogn Neurosci, 15(7), 925–934.
Mitra, P. P. & Pesaran, B. (1999). Analysis of dynamic brain imaging data. Biophys J,
76(2), 691–708.
Moller, E., Schack, B., et al. (2001). Instantaneous multivariate EEG coherence ana-
lysis by means of adaptive high-dimensional autoregressive models. J Neurosci
Methods, 105(2), 143–158.
Nichols, T. E. & Holmes, A. P. (2002). Nonparametric permutation tests for functional
neuroimaging: a primer with examples. Human Brain Mapping 15(1), 1–25.
Palus, M. & Hoyer, D. (1998). Detecting nonlinearity and phase synchronization with
surrogate data. IEEE Eng Med Biol Mag, 17(6), 40–45.
Pantazis, D., Nichols, T. E., et al. (2005). A comparison of random field theory and
permutation methods for the statistical analysis of MEG data. Neuroimage, 25(2),
383–394.
Passingham, R. E., Stephan, K. E., et al. (2002). The anatomical basis of functional
localization in the cortex. Nat Rev Neurosci, 3(8), 606–616.
Pfurtscheller, G. & Andrew, C. (1999). Event-related changes of band power and
coherence: Methodology and interpretation. J Clin Neurophysiol, 16(6), 512–519.
Pfurtscheller, G. & Lopes da Silva, F. H. (1999). Event-related EEG/MEG synchro-
nization and desynchronization: basic principles. Clin Neurophysiol, 110(11),
1842–1857.
Pollok, B., Gross, J., et al. (2004). The Cerebral Oscillatory Network of Voluntary
Tremor. J Physiol, 554, 871–878.
Rennie, C. J., Robinson, P. A., et al. (2002). Unified neurophysical model of EEG
spectra and evoked potentials. Biol Cybern 86(6), 457–471.
Robinson, S. E. & Vrba, J. (1997). Functional neuroimaging by synthetic aperture
magnetometry (SAM). In: T. Yoshimoto, M. Kotani, S. Kuriki, H. Karibe &
B. Nakasato. Recent Advances in Biomagnetism. Sendai: Tohoku University Press,
pp. 302–305.
Rodriguez, E., George, N., et al. (1999). Perception’s shadow: long-distance
synchronization of human brain activity. Nature, 397(6718), 430–433.
Rosenblum, M. G., Cimponeriu, L., et al. (2002). Identification of coupling direction:
application to cardiorespiratory interaction. Phys Rev E Stat Nonlin Soft Matter
Phys, 65(4 Pt 1), 041909.
Rosenblum, M. G. & Kurths, J. (1998). Analysing Synchronization Phenomena from
Bivariate Data by Means of the Hilbert Transform. In: Kantz, H., Kurths, J. & Mayer-
Kress, G. (pp. 91–99). Nonlinear Analysis of Physiological Data. Berlin: Springer.
Rosenblum, M. G. & Pikovsky, A. S. (2001). Detecting direction of coupling in inter-
acting oscillators. Phys Rev E Stat Phys Plasmas Fluids Relat Interdiscip Topics,
64(4-2), 1–4.
Rudrauf, D., Douiri, A., et al. (2006). Frequency flows and the time-frequency dyna-
mics of multivariate phase synchronization in brain signals. Neuroimage, 31(1),
209–227.
Sameshima, K. & Baccala, L. A. (1999). Using partial directed coherence to describe
neuronal ensemble interactions. J Neurosci Methods, 94(1), 93–103.
Schack, B., Rappelsberger, P., et al. (1999). Adaptive phase estimation and its appli-
cation in EEG analysis of word processing. Journal of Neuroscience Methods 93(1),
49–59.
Schnitzler, A. & Gross, J. (2005). Normal and pathological oscillatory communication
in the brain. Nat Rev Neurosci, 6(4), 285–296.
244 MEG: An Introduction to Methods
Schreiber, T. & Schmitz, A. (2000). Surrogate time series. Physica D, 142, 346–382.
Sekihara, K., Nagarajan, S. S., et al. (2001). Reconstructing spatio-temporal activities
of neural sources using an MEG vector beamformer technique. IEEE Trans Biomed
Eng 48(7), 760–771.
Sekihara, K., Nagarajan, S. S., et al. (2002a). Application of an MEG eigenspace beam-
former to reconstructing spatio-temporal activities of neural sources. Hum Brain
Mapp 15(4), 199–215.
Sekihara, K., Nagarajan, S. S., et al. (2002b). Performance of an MEG adaptive-beam-
former technique in the presence of correlated neural activities: effects on signal
intensity and time-course estimates. IEEE Trans Biomed Eng 49(12 Pt 2), 1534–46.
Singer, W. (1999). Neuronal synchrony: a versatile code for the definition of relations?
Neuron 24(1), 49–65, 111–125.
Singh, K. D., Barnes, G. R., et al. (2003). Group imaging of task-related changes in
cortical synchronisation using nonparametric permutation testing. Neuroimage
19(4), 1589–1601.
Sporns, O., Tononi, G., et al. (2005). The human connectome: a structural description
of the human brain. PLoS Comput Biol 1(4), e42.
Stam, C. J., Breakspear, M., et al. (2003). Nonlinear synchronization in EEG and
whole-head MEG recordings of healthy subjects. Hum Brain Mapp, 19(2), 63–78.
Sudmeyer, M., Pollok, B., et al. (2004). Postural tremor in Wilson’s disease: A magne-
toencephalographic study. Mov Disord, 19(12), 1476–1482.
Tass, P., Rosenblum, M. G., et al. (1998). Detection of n:m Phase Locking from Noisy
Data: Application to Magnetoencephalography. Phys Rev Lett 81(15), 3291–3294.
Tass, P. A., Fieseler, T., et al. (2003). Synchronization tomography: a method for
three-dimensional localization of phase synchronized neuronal populations in the
human brain using magnetoencephalography. Phys Rev Lett, 90(8), 088101.
Thomson, D. J. (1982). Spectrum estimation and harmonic analysis. Proc IEEE 70(9),
1055–1096.
Timmermann, L., Gross, J., et al. (2003). The cerebral oscillatory network of parkin-
sonian resting tremor. Brain 126(Pt 1), 199–212.
Vallbo, A. B., & Wessberg, J. (1993). Organization of motor output in slow finger
movements in man. J Physiol (Lond), 469, 673–691.
Van Veen, B. D., van Drongelen,W., et al. (1997). Localization of brain electrical acti-
vity via linearly constrained minimum variance spatial filtering. IEEE Trans Biomed
Eng 44(9), 867–880.
Varela, F., Lachaux, J. P., et al. (2001). The brainweb: phase synchronization and
large-scale integration. Nat Rev Neurosci 2(4), 229–239.
Vrba, J. & Robinson, S. E. (2001). Signal processing in magnetoencephalography.
Methods 25(2), 249–271.
Welch, P. D. (1967). The use of Fast Fourier Transform for the estimation of power
spectra: A method based on time averaging over short, modified periodograms.
IEEE Trans Audio Electroacout, AU-15, 70–73.
Wessberg, J., & Kakuda, N. (1999). Single motor unit activity in relation to pulsatile
motor output in human finger movements. J Physiol (Lond) 517(Pt 1), 273–285.
Wessberg, J., & Vallbo, A. B. (1995). Coding of pulsatile motor output by human mus-
cle afferents during slow finger movements. J Physiol (Lond) 485(Pt 1), 271–282.
Wessberg, J., & Vallbo, A. B. (1996). Pulsatile motor output in human finger move-
ments is not dependent on the stretch reflex. J Physiol (Lond), 493(Pt 3), 895–908.
Worsley, K. J., & Friston, K. J. (1995). Analysis of fMRI time-series revisited–again.
Neuroimage 2(3), 173–181.
10
Statistical Inference in MEG Distributed
Source Imaging
Dimitrios Pantazis and Richard M. Leahy
• This chapter reviews the statistical tools available for MEG analysis
• We describe how statistical maps of brain activation are created on the
cortex using the General Linear Modeling approach
• We review methods to threshold these maps and establish statistical
significance while controlling for multiple comparisons
Introduction
As more whole-head systems become available, magnetoencephalography
(MEG) is increasingly being used in clinical and cognitive neuroscience to
image human brain function. With the use of novel experimental paradigms,
researchers are using MEG to explore many aspects of the workings of the
human brain. To assure an objective scientific interpretation of these studies,
it is important that experimental findings be accompanied by appropriate
statistical analysis that effectively controls for false positives.
This chapter reviews the statistical tools available for the analysis of dis-
tributed activation maps, defined either on the 2D cortical surface or throughout
the 3D brain volume. Statistical analysis of MEG data bears a great resemblance
to the analysis of fMRI or PET activation maps; therefore much of the meth-
odology can be borrowed or adapted from the functional neuroimaging
245
246 MEG: An Introduction to Methods
approaches that are used for analyzing, respectively, evoked and induced
components in the MEG data (Figure 10–1).
The fact that we often want to identify and localize experimental
effects—not only over space, as traditionally done in fMRI with the notion of
voxels, but also in time and possibly frequency—introduces challenges that
differentiate MEG analysis from that of PET and fMRI. The high dimension-
ality of the data (space × time × frequency × experimental design) presents
challenges in terms of high computational costs, but also possibilities in terms
of greater flexibility in the design of the linear models.
Another important difference relative to fMRI is that MEG offers only
limited spatial resolution. Distributed cortical imaging involves the reconstruc-
tion of thousands of elemental current sources from a few hundred measure-
ments. The problem is highly underdetermined and requires regularization to
produce a stable solution. The resulting images are typically of low resolution,
so that reconstructions of focal sources are blurred with extended point-spread
functions (PSF). The shape of the PSF will depend on the reconstruction space,
cortical or volumetric, and whether the orientations of the sources are con-
strained to be normal to the cortical surface. Unlike in fMRI, the PSFs for MEG
are highly asymmetric and can extend over multiple gyri or sulci. As a result,
even after thresholding to control for false positives, one can still observe false
positives at locations within the point spread of truly active regions and, there-
fore, care must be taken in interpreting these results. Figure 10–8, shown later
in this chapter, illustrates this issue; the reconstructed statistical map has much
greater spatial extent than the single simulated cortical patch, and subsequent
thresholding procedures identify significant activity in broad cortical areas.
Figure 10–1. MEG brain activity in response to a task consists of two compo-
nents: evoked responses that are phase locked to the stimuli, and induced
responses that are not. Averaging the MEG time-series over epochs pre-
serves the evoked components, but suppresses the induced components
(left). Averaging the power time-frequency decompositions of the time-
series preserves both evoked and induced components (right).
248 MEG: An Introduction to Methods
This issue is closely related to the MEG resolution kernel and the actual
hypothesis being tested. In MEG, the channel measurements m are linearly
related with the brain activation s as: m = Gs + n, where G is the lead field
matrix that depends on the shape and conductivity of the head as well as
the position of the sensors, and n is the channel noise. We obtain an estimate
of the MEG sources ŝ with some linear inverse method: ŝ = Wm = WGs + Wn,
where W is the inverse operator and R = WG is the resolution kernel.
In distributed cortical imaging, where we estimate more sources than the
available channel measurements, R is not the identity matrix and our solution
is biased. For a particular cortical location i, when we test the hypothesis
ŝ i = 0, we are effectively testing (WG)is = 0, where (.)i represents the ith row of
a matrix. Our true null hypothesis is, therefore, not that the cortical activity at
location i is zero, but rather that the whole brain activity, linearly weighted by
the resolution kernel at location i, is zero. Waldorp et al. (2006) therefore
suggests performing hypothesis testing on multivariate models where spa-
tially smooth regions of interest lead to more interpretable hypotheses than
univariate models.
Analysis in fMRI is typically performed in the 3D volumetric space, while
in MEG the 2D cortical surface is often chosen as the source space. Cortically
constrained maps can complicate the analysis in several ways. For example,
isotropic smoothing on the cortical surface when applying random field
methods requires the use of the Laplace-Beltrami operator (Chung, 2001). In
group analysis the data should be brought into a common coordinate system,
which requires cortical surface alignment rather than volumetric registration
(Joshi et al., 2007b; Fischl et al., 1999), and the resulting areas of activation
should be reported with respect to cortical anatomy rather than the standard
Talairach coordinates. Orientation-free MEG reconstructions produce vector
rather than scalar fields (3 elemental dipoles at each location), which can also
complicate analysis.
In addition to producing a nonuniform PSF, the MEG inverse operator
also introduces a highly non-stationary spatial covariance structure in recon-
structed images. Contributions to the covariance can include trial-to-trial
variations in induced and evoked responses, as well as physiological and envi-
ronmental noise. Furthermore, the covariance can also vary substantially
over the course of an experiment so that we can often not assume temporal
stationarity. In comparison, variations in fMRI data can often reasonably be
approximated as spatially and temporally stationary. As a result, statistical
inference for MEG with random field theory requires the use of special
formulas that correct for non-stationarity (Worsley et al., 1999).
use a general linear model to fit the observations at each location, and finally,
generate a contrast of the estimated parameters and normalize with its vari-
ance to create a map of pivotal statistics (t-maps, F-maps etc). This methodol-
ogy is a standard approach in fMRI and PET data analysis, and together with
subsequent statistical inference, is generally referred to as Statistical Parametric
Mapping (SPM) (Friston et al., 1995).
Observations
Cortical Regions Si
Frequency
5 Hz
10 Hz
40 Hz
F2
Y = Xb + ε
1 1
Contrast of
1
1
1
1 estimated parameters CT bˆ
T= =
1 1
σˆ 2CT (XT X )−1C
1 1 variance
1 1 estimate
1 1
1 1
averaged evoked responses due to the central limit theorem. However, power
time-frequency decompositions of single trial data have a χ2-distribution.
Fortunately, Kiebel et al. (2005) has shown that, under most circumstances,
one can appeal to the central limit theorem, or transform the MEG power
estimates with a log or square-root transform, to make the error terms normal,
and thus GLM theory is still appropriate.
Under the GLM framework, the MEG observations Y are predicted from
the parameters b:
Y = Xb + e (10–1)
where ε is the modeling error. X is the design matrix whose elements model
an experimental paradigm and consist of qualitative (0s or 1s) and/or quanti-
tative variables.
To provide intuition on using the GLM theory for MEG data modeling,
consider the following example. In a MEG visual attention study, we acquire
multi-trial data for two conditions: subject attends to the right (condition 1),
or to the left (condition 2). By combining an inverse method with time-
frequency analysis of individual trials, we produce dynamic images of brain
activity in the α-frequency band. The α-power observations for a single voxel,
y11, …, y1N for condition 1 and y21, …, y2M for condition 2, are fitted into a one-
way ANOVA GLM model with two predictor variables, b1 and b2, for the two
conditions:
yij = bi + e ij (10–2)
where i = {0, 1} denotes the condition, j the trial repetition for each condition,
and εij is the model error.
The same ANOVA model can be written in matrix notation. If the
observations yij are arranged on a single observation vector Y, and the rows of
the design matrix X have 0s and 1s to indicate the condition for each MEG
observation, the ANOVA model becomes Y = Xb + ε, explicitly written as:
⎡ y11 ⎤ ⎡1 0⎤ ⎡ e11 ⎤
⎢ y ⎥ ⎢1 0⎥ ⎢e ⎥
⎢ 12 ⎥ ⎢ ⎥ ⎢ 12 ⎥
⎢
⎥ ⎢
⎥ ⎢
⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ y1N ⎥ = ⎢1 0⎥ ⎡ b1 ⎤ + ⎢ e1N ⎥
⎢ y 21 ⎥ ⎢0 1⎥ ⎢⎣b2 ⎥⎦ ⎢ e 21 ⎥ (10–3)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ y 22 ⎥ ⎢0 1⎥ ⎢ e 22 ⎥
⎢
⎥ ⎢
⎥ ⎢
⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢⎣ y 2 M ⎥⎦ ⎢⎣0 1⎥⎦ ⎢⎣e 2 M ⎥⎦
⎡ y1. ⎤
bˆ = ( X T X )−1 X T Y = ⎢ ⎥ (10–4)
⎣ y 2. ⎦
where the bar denotes the mean over the dotted subscript. The estimated
error variance σ̂ 2 has N + M − 2 degrees of freedom, because two of them
where used to estimate the model predictors. The error and error variance are
estimated as:
εˆ T εˆ Σ ( y − y )2 + Σi ( y 2i − y 2. )2
σˆ 2 = = i 1i 1. (10–6)
trace{P} N + M −2
where P is a projection operator onto the left null space of X. We want to test
whether there is a difference between the two conditions, or equivalently
whether the difference b1 − b2 is significantly different from zero. The statistic
T
T ˆ ⎡1⎤ ˆ
of interest is therefore the contrast of the two parameters, c b = ⎢ ⎥ b = b1 − b2 ,
⎣ −1⎦
which is then normalized with an estimate of its standard deviation. The
resulting statistic T is a two-sample t-test between the two conditions.
c T bˆ y1. − y 2.
T= = (10–7)
−1
σˆ c ( X X ) c
2 T T
σˆ (1/ N + 1/ M )
2
Types of GLMs
There are three ways to organize the MEG data into GLM observations Y: a
mass-univariate approach, a multivariate approach, and a general univariate
formulation (Kiebel and Friston, 2004). The mass-univariate approach con-
siders the data at each location in isolation. Therefore, a separate but identical
GLM is fitted at each spatial-temporal-spectral location and analyzed using
an ANOVA or ANCOVA approach. The data correlations in the respective
dimensions are ignored at this stage, and accommodated at the inference
stage through adjusting the P-values associated with the statistical maps.
For example, even though the activation of nearby voxels is correlated, the
Statistical Inference in MEG Distributed Source Imaging 253
After selection of a GLM approach, the MEG observations are fitted to the
models and a contrast (or linear combination) of the parameters is computed
(Figure 10–2e). This contrast statistic captures the effect of interest—for
example, the difference between two experimental conditions—or the corre-
lation of a response variable with brain activation. It is then preferable to
normalize the statistics into known parametric distributions (pivotal statis-
tics). This allows the application of random field theory, which, as we will see,
requires a Gaussian distribution or one derived from Gaussian data (e.g., at or
F statistic). The normalization also helps when using nonparametric permu-
tation methods, because it makes the variance at all voxels homogeneous
under the null hypothesis, which should produce approximately uniform
specificity; i.e. false positives are equally likely at all locations.
We conclude by showing that the GLM framework is parsimonious in
MEG analysis. Consider, for example, the simple case where the MEG data
are used to create dSPM maps (Dale et al., 2000), i.e. minimum-norm inverse
maps normalized with an estimate of the noise standard deviation at each
location. This corresponds to the simplest case of GLM analysis following a
mass-univariate approach: the one-way ANOVA model y itj = bit + e itj is fitted
to the data separately at each spatial location i and temporal location t, where
j is the trial repetition index and bit is the main effect (brain response) (Pantazis
et al., 2005b). We use superscripts for i and t to denote that the same model is
fit separately in each spatial-temporal location. The estimated contrast of
interest is the parameter itself, which is equal to the trial average according to
the minimum-norm solution: c T bˆit = [1]bˆit = bˆit = y.it , where the bar indicates
an average over the dotted subscript. Since the error terms e j are assumed to
it
and J is the total number of trials. This is equivalent to the noise normaliza-
tion performed in dSPM for orientation-constrained linear inverses. For the
unconstrained case, the dSPM output is an F-map. The sLORETA solution
(Pascual-Marqui, 2002) is similar to dSPM, but with a different normalization
coefficient. In this case we normalize by the standard deviation computed
from the data covariance, rather than the noise-only covariance. Under the
null hypothesis of noise-only (or equivalently, a zero experimental effect),
sLORETA and dSPM are identical. Similarly, the beam-former neural activity
index (Veen et al., 1997) corresponds to a t-map for the orientation-
constrained case, or an F-map for the orientation-free case, which can be
again cast in a GLM framework.
Multisubject Studies
where Yk are the MEG observations and bk are the model parameters of sub-
ject k. While each design matrix Xk can have a different number of rows (for
256 MEG: An Introduction to Methods
example, different number of MEG trials per subject), all the design matrices
must have the same number of columns, with each column expressing the
same effect among subjects. The subject parameters are estimated using a
generalized least-squares solution, which normally requires the estimation of
the error covariance matrix Ck:
The second stage model takes only one contrast c T bˆk from each subject
and fits it to the group GLM:
⎡ c T bˆ1 ⎤
⎢ ⎥
⎢ c T bˆ2 ⎥
⎢ ⎥ = X g bg + e g (10–10)
⎢
⎥
⎢c T bˆ ⎥
⎣ K⎦
where Xg and bg are the group design matrix and group level parameters
respectively. The summary statistics model error εg has two variance compo-
nents, the intrasubject and intersubject variance. Under the assumption of
homogeneous intrasubject variance (i.e. c T bˆk , has the same variance for all
subjects), the intrasubject variance is a scaled identity matrix. Similarly, under
the assumption of independent subjects, the intersubject variance is a scaled
identity matrix. Therefore, the covariance of εg is also a scaled identity matrix,
and the generalized least-squares solution of the second stage model (which
normally requires estimation of the error covariance matrix) becomes
equivalent to an ordinary least-squares solution that does not require the
covariance matrix:
This statistic map can then be thresholded for significant activity at the group
level, using any of the methods described later in this chapter.
Consider the following example. In Pantazis et al. (2007), the first-stage
model consisted of fitting a univariate model for each subject in the alpha
band, for each of several cortically defined regions of interest and time bands.
A contrast statistic was then estimated that captured an attention effect:
ipsilateral minus contralateral alpha power in each spatiotemporal band. The
contrast for all subjects was then fitted to a second-stage GLM (Equation 10–10),
whose design matrix Xg is a column of 1s. This simply leads to averaging the
responses from all subjects (Equation 10–11), since the assumption of homo-
geneous intrasubject variance allows application of the simple summary
statistics approach described above. Finally, the FDR approach, as described
below, was used to threshold the resulting statistic map. Other examples of
multisubject MEG studies can be found in Singh et al. (2003), and Kiebel and
Friston (2004).
have one or more false positives, or type I errors, at any location in the brain.
We now investigate how the FWER is related to the maximum statistic.
Maximum Statistic
The FWER is directly related to the maximum value in the statistical image;
one or more voxels Ti will exceed the threshold uα under the null hypothesis
H0 only if the maximum exceeds that threshold:
for control of FWER. As we see below, P-values are also used when controlling
the FDR.
Bonferroni Correction
≤ ∑ P(Ti ≥ uα }| H o ) (Bonferroniinequality)
b
i
(10–13)
= ∑ αb
i
= ∑ α /V = α
i
Figure 10–6. Random field theory uses the topological features of a sta-
tistical map to convert the voxels into RESELS, a dimensionless quantity
that represents the image with interpretable units of smoothness. Under
the null hypothesis of no experimental effects, it is the degree of spatial
correlation in the noise in the statistical maps that determines the RESEL
count.
262 MEG: An Introduction to Methods
Permutation Methods
assumed to have a symmetric distribution around zero under the null hypoth-
esis. Therefore, randomly multiplying it by 1 or −1 does not change its distri-
bution under the null hypothesis. With K subjects (and thus K contrasts), a
total of 2K permutation samples can be created, which can then each be fitted
to the second level GLM, to estimate permuted group parameters that are
used in turn to estimate the empirical distribution of the group-averaged
map. Such an approach was followed, for example, by Singh et al. (2003) and
264 MEG: An Introduction to Methods
pre-stim post-stim
j
Original Date
t : Time index
i : Voxel index
i
j : Trial index
Trial Average
t
Threshold
Permutation Samples
control false positives in the channel domain (Blair & Karnisky, 1993, 1994;
Karnisky et al., 1994; Maris, 2004; Achim, 2001; Galan et al., 1997) and in the
source domain (Park et al., 2002; Pantazis et al., 2003, 2005b; Chau et al.,
2004; Singh et al., 2003; Sekihara et al., 2005). These methods have been
applied in multiple MEG studies (Kaiser et al., 2000; Lutzenberger et al., 2002;
Cheyne et al., 2006; Bayless et al., 2006; Itier et al., 2006; Pantazis et al., 2005a,c,
2007). Reviews on the application of permuation tests in MEG are available in
Maris and Oostenveld (2007); and in Maris et al. (2007).
Various thresholding methods are illustrated in Figure 10–8. The permu-
tation method was based on that described in Pantazis et al. (2005b) and
produced a threshold of 3.99, which controls FWER over the whole cortex at
a 5% level.
In contrast to the above methods that control the FWER, FDR controls the
expected proportion of errors among the rejected hypotheses (Benjamini &
Hochberg, 1995; Genovese et al., 2002). For example, if we set an α = 5% FDR
threshold, then on average we should expect 5% of our suprathreshold voxels
to be false positives.
The standard FDR method proposed by Benjamini and Hochberg (1995)
is conservative, as it controls the FDR at a V0 α level, where V is the total num-
V
ber of voxels and V0 is the number of voxels where the null hypothesis is true:
⎛ false positives ⎞ V0
E(FDR) = E ⎜ ⎟⎠ ≤ V α (10–15)
⎝ suprathreshold voxels
Figure 10–8. Simulated MEG source on the left hemisphere and recon-
structed dSPM statistical map, thresholded using several methods to con-
trol false positives. Uncorrected thresholding (P-value = 0.05) and FDR
(P-value = 0.0065) produced many false positives. Permutations and ran-
dom field theory gave thresholds t = 3.99 and t = 4.12, respectively. The
Bonferroni approach was very conservative (P-value 0.05/7501 = 6.6610−6)
and did not identify the source.
266 MEG: An Introduction to Methods
When the true brain activation extends over broad areas, V0 may become
small and the FDR procedure too conservative. Thus, a number of adaptive
procedures improve on the original FDR approach by first estimating V0, and
then using this estimate to tighten the threshold (Benjamini et al., 2006).
The FDR methods adapt to the properties of the data; when a large number
of voxels are truly active, the threshold will adjust to allow for more false posi-
tives; when no truly activated voxels exist, FDR controls the FWER, but in a
weak sense (see False Positive Measures earlier in the chapter). They are more
powerful than Bonferroni, random field, and permutation control of FWER,
and for this reason may become popular for thresholding MEG maps.
The FDR approach requires the estimation of the marginal distribution
at each location in the statistical map, or, equivalently, conversion of the sta-
tistic value at each location into an equivalent P-value. We can do this para-
metrically or non-parametrically, as described for the Bonferroni approach.
Once the maps are converted to P-values, implementation of the stan-
dard FDR method is relatively straightforward (Genovese et al., 2002). If V is
the total number of voxels being tested, the procedure is as follows:
(1) Order the voxel P-values from smallest to largest:
i α
P(i ) ≤ (10–17)
V c (V )
(3) Declare all voxels corresponding to the P-values P(1),…, P(r) active.
where c (V ) = ∑ i =11/ i if no assumptions on the joint distribution of the
V
P-values across voxels is made, and c(V) = 1 if the P-values in different voxels
are independent or they have positive dependence (Benjamini & Yekutieli,
2001).
The procedure is demonstrated graphically in Figure 10–9 for the FDR
procedure applied to the statistical map in Figure 10–8. The estimated thresh-
old for 7501 cortical voxels at the 5% level was 0.065, and produced an
extended region of suprathreshold voxels. Unfortunately, the large extent of
the significantly active region determined using FDR, is a result of the limited
spatial resolution of MEG; many voxels surrounding the true simulated
source exhibit significant activity in the statistical map, and FDR is sensitive
enough to identify them. Conversely, the more conservative thresholds from
FWER control tend to reduce the size of activated regions, as also shown in
Figure 10–8. Other examples of the application of FDR in MEG maps include
Edwards et al. (2005); Jacobs et al. (2006); Pantazis et al. (2007); and Jacques
and Rossion (2007).
Statistical Inference in MEG Distributed Source Imaging 267
0.04 × 10−3
Sorted P-values
7
0.05
i
0.02 7501
6.5
0 6
960 980 1000
0 1000 2000
Figure 10–9. Graphical representation of the FDR procedure. The blue line
represents the sorted P-values of the statistical map in Figure 10–8, and
the red line is i V = i 7501. The largest P-value below the line is 0.0065 and
α 0.05
Discussion
We have presented a GLM framework to produce statistical maps from MEG
distributed cortical imaging, and subsequently threshold them while control-
ling for false positives. Choosing a thresholding method depends on the data
available: Bonferroni is simple and efficient for a small number of tests with
minimal dependence; random fields are robust when their parametric
assumptions are satisfied, and strong correlation exists in the data; permuta-
tion tests are very general, adapt to the underlying data correlations, can use
any statistic, and are more powerful than random fields for data with low
degrees of freedom (e.g., studies with a few subjects); FDR is more powerful,
works well with sparse signals, and is recommended when we can afford a few
false positives. However, as we have shown in our simulations, FDR can pro-
duce large regions of significant activation as a result of the limited resolution
of MEG inverses.
A number of important statistical issues are beyond the scope of this
chapter, such as conjunction analysis (Nichols et al., 2005), i.e., the identifica-
tion of brain areas that are simultaneously active in multiple tasks; extraction
of confidence intervals for distributed solutions using the bootstrap (DiNocera
and Ferlazzo, 2000; Gross et al., 2003; Darvas et al., 2005); and thresholding
using cluster-size tests (Hayasaka et al., 2004). Also, the theory was developed
for distributed inverse methods in MEG. However, discrete solutions are also
popular in MEG, especially with well localized activation when a few equiva-
lent current dipoles can represent most cortical activity. In this case, alternative
approaches can be used to establish significance; for example, the bootstrap
resampling approach, or Monte Carlo simulations to find localization accu-
racy and confidence intervals for current dipoles (Darvas et al., 2005; Braun
et al., 1997).
268 MEG: An Introduction to Methods
References
Achim, A. (2001). Statistical detection of between-group differences in event-related
potentials. Clinical Neurophysiology, 112, 1023–1034.
Adler, R. J. (ed.) (1981). The Geometry of Random Fields. New York: Wiley.
Barnes, G. R., & Hillbrand, A. (2003). Statistical flattening of MEG beamformer
images. Human Brain Mapping, 18, 1–12.
Bayless, S. J., Gaetz, W. C., Cheyne, D. O., & Taylor, M. J. (2006). Spatiotemporal
analysis of feedback processing during a card sorting task using spatially filtered
MEG. Neuroscience Letters, 410, 31–36.
Beckmann, C. F., Jenkinson, M., & Smith, S. M. (2003). General multilevel linear
modeling for group analysis in fMRI. Neuroimage, 20(2), 1052–1063.
Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical
and powerful approach to multiple testing. Journal of the Royal Statistical Society,
57, 289–300.
Benjamini, Y., Krieger, A. M., & Yekutieli, D. (2006). Adaptive linear step-up procedures
that control the false discovery rate. Biometrica, 93(3), 491–507.
Benjamini, Y., & Yekutieli, D. (2001). The control of the false discovery rate in
multiple testing under dependency. The Annals of Statistics, 29(4), 1165–1188.
Blair, R. C., & Karnisky, W. (1993). An alternative method for significance testing of
waveform difference potentials. Psychophysiology, 30, 518–524.
Blair, R. C., & Karnisky, W (1994). Distribution-free statistical analyses of surface
and volumetric maps. In: R. W. Thatcher, M. Hallett, T. Zeffiro, E. R. Jony, and
M. Huerta, eds.)Functional Neuroimaging (pp. 19–28). New York: Academic Press.
Braun, C., Kaiser, S., Kinces, W., & Elbert, T. (1997). Confidence interval of single
dipole locations based on EEG data. Brain Topography, 10(1), 31–39.
Brooks, M. J., Gibson, A. M., Hall, S. D., Furlong, P. L., Barnes, G. R., Hillebrand, A.,
et al. (2004). A general linear model for MEG beamformer imaging. Neuroimage,
23, 936–946.
Bruns, A. (2004). Fourier-, hilbert and wavelet-based signal analysis: are they really
different approaches? Journal of Neuroscience Methods, 137, 321–332.
Statistical Inference in MEG Distributed Source Imaging 269
Carbonell, F., Galan, L., Valdes, P., Worsley, K., Biscay, R. J., Diaz-Comas, L.,
et al. (2004). Random field - union intersection tests for EEG/MEG imaging.
NeuroImage, 22, 268–276.
Chau, W., McIntosh, A. R., Robinson, S. E., Schulz, M., & Pantev, C. (2004). Impro-
ving permutation test power for group analysis of spatially filtered MEG data.
NeuroImage, 23, 983–996.
Cheyne, D., Bakhtazad, L., & Gaetz, W. (2006). Spatiotemporal mapping of cortical
activity accompanying voluntary movements using an event-related beamforming
approach. Human Brain Mapping, 27, 213–229.
Christensen, G. E., & Johnson, H. J. (2001). Consistent image registration. IEEE
Transactions on Medical Imaging, 20(7), 568–582.
Chung, M. K. (2001). Statistical morphometry in neuroanatomy. Ph.D. thesis, McGill
University, Montreal.
Dale, A. M., Liu, A. K., Fischi, R. B., Buckner, R. L., Belliveau, J. W., Lewine, J. D.,
et al. (2000). Dynamic statistical parametric mapping: Combining fMRI and MEG
for high-resolution imaging of cortical activity. Neuron, 26, 55–67.
Darvas, F., Rautiainen, M., Pantazis, D., Baillet, S., Benali, H., Mosher, J. C., et al. (2005).
Investigations of dipole localization accuracy in MEG using the bootstrap. Neuro-
Image, 25, 355–368.
DiNocera, F., & Ferlazzo, F. (2000). Resampling approach to statistical inference:
bootstrapping from event-related potentials data. Behavioral Research Methods,
Instruments and Computers, 32(1), 111–119.
Edgington, E. S. (ed.). (1995). Randomization Tests (3rd ed.). New York: Marcel Dekker.
Edwards, E., Soltani, M., Deouell, L. Y., Berger, M. S., & Knight, R. T. (2005). High
gamma activity in response to deviant auditory stimuli recorded directly from
human cortex. Journal of Neurophysiology, 94, 4269–4280.
Fischl, B., Serano, M., Tootell, R., & Dale, A. M. (1999). High-resolution intersubject
averaging and a coordinate system for the cortical surface. Human Brain Mapping,
8(4), 272–284.
Friston, K. J., Holmes, A. P., Worsley, K. J., Poline, J. B., Frith, C., &
Frackowiak, R. S. J. (1995). Statistical parametric maps in functional imaging:
A general linear approach. Human Brain Mapping, 2, 189–210.
Friston, K. J., Penny, W., Phillips, C., Kiebel, S., Hinton, G., & Ashburner, J. (2002).
Classical and bayesian inference in neuroimaging: Theory. NeuroImage, 16, 465–483.
Friston, K. J., Stephan, K. M., Heather, J. D., Frith, C. D., Ioannides, A. A., Liu, L. C.,
et al. (1996). A multivariate analysis of evoked responses in EEG and MEG data.
Neuroimage, 3, 167–174.
Fuchs, M., Wagner, M., Kohler, T., & Wischmann, H. (1999). Linear and non-
linear current density reconstructions. Journal of Clinical Neurophysiology, 16,
267–295.
Galan, L., Biscay, R., Rodriguez, J. L., Abalo, M. C. P., & Rodriguez, R. (1997). Testing
topographic differences between event related brain potentials by using non-parametric
combinations of permutation tests. Electroencephalography and clinical Neurophy-
siology, 102, 240–247.
Genovese, C., Lazar, N., & Nichols, T. (2002). Thresholding of statistical maps in
functional neuroimaging using the false discovery rate. NeuroImage, 15, 870–878.
Gross, J., Kujala, J., Hämäläinen, M., Timmermann, L., Schnitzler, A., & Salmelin, R.
(2001). Dynamic imaging of coherent sources: Studying neural interactions in the
human brain. Proceedings of National Academy of Sciences, 98(2), 694–699.
270 MEG: An Introduction to Methods
Gross, J., & Timmermann, L., Kujala, J., Salmelin, R., & Schnitzler, A. (2003).
Properties of MEG tomographic maps obtained with spatial filtering. Neuroimage,
19, 1329–1336.
Hämäläinen, M. S., & Ilmoniemi, R. J. (1984). Interpreting magnetic fields of the
brain: minimum norm estimates. Medical and Biological Engineering and Computing,
32(1), 35–42.
Hayasaka, S., & Nichols, T. E. (2003). Validating cluster size inference: Random field
and permutation methods 20 (4), 2343–2356.
Hayasaka, S., & Nichols, T. E. (2004). Combining voxel intensity and cluster extent
with permutation test framework. Neuroimage, 23(1), 54–63.
Hayasaka, S., Phan, K. L., Libarzon, I., Worsley, K. J., & Nichols, T. E. (2004).
Nonstationary cluster-size inference with random field and permutation methods.
NeuroImage, 22, 676–687.
Hellier, P., Ashburner, J., Corouge, I., Barillot, C., & Friston, K. J. (2002). Inter-subject
registration of functional and anatomical data using textSPM. In: Medical Image
Computing and Computer-Assisted Intervention - MICCAI 2002, pp. 590–597.
Hochberg, Y., & Tamhane, A. C. (eds.). (1987). Multiple Comparison Procedures.
New York: Wiley.
Holmes, A., & Friston, K. J. (1998). Generalisability, random effects and population
inference. NeuroImage, 7, S754.
Hui, H., Leahy, R. M. (2006). Linearly constrained MEG beamformers for MVAR
modeling of cortical interactions. In: Biomedical Imaging: From Macro to Nano,
2006. 3rd IEEE International Symposium, pp. 237–240.
Itier, R. J., Herdman, A. T., George, N., Cheyne, D., & Taylor, M. J. (2006). Inversion
and contrast-reversal effects on face processing. Brain Research, 1115, 108–120.
Jacobs, J., Hwang, G., Curran, T., & Kahana, M. J. (2006). EEG oscillations and
recognition memory: Theta correlates of memory retrieval and decision making.
NeuroImage, 32, 978–987.
Jacques, C., & Rossion, B. (2007). Early eloctrophysiological responses to multiple
face orientations correlate with individual discrimination performance in humans.
NeuroImage, 36, 863–876.
Jerbi, K., Lachaux, J. P., N’Diaye, K., Pantazis, D., Leahy, R. M., & Garnero, L. (2007).
Coherent neural representation of hand speed in humans revealed by MEG imaging.
PNAS, 104(18), 7676–7681.
Joshi, A. A., Shattuck, D. W., Thompson, P. M., Leahy, R. M. (2007a). A finite element
method for elastic parameterization and alignment of cortical surfaces using sulcal
constraints. In: Biomedical Imaging: From Macro to Nano, 2007. 4th IEEE Interna-
tional Symposium, pp. 640–643.
Joshi, A. A., Shattuck, D. W., Thompson, P. M., & Leahy, R. M. (2007b). Surface cons-
trained volumetric brain registration using harmonic mappings. IEEE Transactions
on Medical Imaging, 26(12), 1657–1669.
Kaiser, J., Lutzenberger, W., Preissl, H., Mosshammer, D., & Birbaumer, N. (2000).
Statistical probability mapping reveals high-frequency magnetoencephalographic
activity in supplementary motor area during self-paced finger movements.
Neuroscience Letters, 283(1), 81–84.
Karnisky, W., Blair, R. C., & Snider, A. D. (1994). An exact statistical method for
comparing topographic maps with any number of subjects and electrods. Brain
Topography, 6, 203–210.
Statistical Inference in MEG Distributed Source Imaging 271
Kiebel, S. (2003). The general linear model. In: Frackowiak, R., Friston, K., Frith, C.,
Dolan, R., Friston, K., Price, C., Zeki, S., Ashburner, J., Penny, W. (eds.), Human
Brain Function, (2nd ed.). San Diego: Academic Press.
Kiebel, S. J., & Friston, K. J. (2004). Statistical parametric mapping for event-related
potentials: I. generic considerations. Neuroimage, 22, 492–502.
Kiebel, S. J., Tallon-Baudry, C., & Friston, K. J. (2005). Parametric analysis of oscilla-
tory activity as measured with EEG/MEG. Human Brain Mapping, 26, 170–177.
Kilner, J. M., Kiebel, S. J., & Friston, K. J. (2005). Applications of random field theory
to electrophysiology. Neuroscience Letters, 374, 174–178.
Lutzenberger, W., Ripper, B., Busse, L., Birbaumer, N., & Kaiser, J. (2002). Dynamics
of gamma-band activity during an audiospatial working memory task in humans.
Journal of Neuroscience, 22(13), 5630–5638.
Maris, E., (2004). Randomization tests for ERP topographies and whole spatiotemporal
data matrices. Psychophysiology, 41, 142–151.
Maris, E., & Oostenveld, R. (2007). Nonparametric statistical testing of EEG- and
MEG-data. Journal of Neuroscience Methods, 164, 177–190.
Maris, E., Schoffelen, J.-M., & Fries, P. (2007). Nonparametric statistical testing of
coherence differences. Journal of Neuroscience Methods, 163, 161–175.
Mosher, J. C., & Leahy, R. M. (1998). Recursive MUSIC: A framework for EEG and MEG
source localization. IEEE Transactions of Biomedical Engineering, 45 (11), 1342–1354.
Mumford, J. A., & Nichols, T. (2006). Modeling and inference of multisubject fmri
data. IEEE Engineering in Medicine and Biology Magazine, March/April, 42–51.
Nichols, T., Brett, M., Andersson, J., Wager, T., & Poline, J.-B. (2005). Valid conjunc-
tion inference with the minimum statistic. NeuroImage, 25, 653–660.
Nichols, T. E., & Hayasaka, S. (2003). Controlling the familywise error rate in functio-
nal neuroimaging: A comparative review. Statistical Methods in Medical Research,
12 (5), 419–446.
Nichols, T. E., & Holmes, A. P. (2001). Nonparametric permutation tests for functio-
nal neuroimaging: A primer with examples. Human Brain Mapping, 15, 1–25.
Pantazis, D., Merrifield, W., Darvas, F., Sutherling, W., & Leahy, R. M. (2005a).
Hemispheric language dominance using MEG cortical imaging and non-parametric
statistical analysis. WSEAS Transactions on Biology and Biomedicine, 2(3), 318325.
Pantazis, D., Nichols, T. E., Baillet, S., & Leahy, R. M. (2003). Spatiotemporal localiza-
tion of significant activation in MEG using permutation tests. In: Taylor, C., Noble, J. A.
(eds.), Proc. 18th Conf. Information Processing in Medical Imaging, pp. 512–523.
Pantazis, D., Nichols, T. E., Baillet, S., & Leahy, R. M. (2005b). A comparison of ran-
dom field theory and permutation methods for the statistical analysis of MEG data.
Neuroimage, 25(2), 383–394.
Pantazis, D., Simpson, G., Weber, D., Dale, C., Nichols, T., & Leahy, R. (2007). Explo-
ring human visual attention in an meg study of a spatial cueing paradigm using a
novel ancova design. In: 2007 IEEE International Symposium on Biomedical Ima-
ging: From Nano to Macro.
Pantazis, D., Weber, D., Dale, C., Nichols, T., G.V. Simpson, & Leahy, R. (2005c). Imag-
ing of oscillatory behavior in event-related MEG studies. In: Bouman, C., Miller, E.
(Eds.), Proceedings of SPIE, Computational Imaging III. Vol. 5674. (pp. 55–63).
Park, H., Kwon, J., Youn, T., Pae, J., Kim, J., Kim, M., & Ha, K. (2002). Statistical para-
metric mapping of LORETA using high density EEG and individual MRI: application
to mismatch negativities in schizophrenia. Human Brain Mapping, 17, 168–178.
272 MEG: An Introduction to Methods
Introduction
Combining Techniques: The Enticements
Since any measurement technique has its own strengths and limitations,
combining different experimental approaches to better probe a scientific
question is a commonplace idea. In the human brain-imaging domain, it is
273
274 MEG: An Introduction to Methods
crude approximations, complex data may not easily fit into the current model
interpretation (here, “model” is taken in a loose sense, as the researcher’s
formal or informal representation of the brain mechanism under investiga-
tion), and therefore be too big a step forward. Multimodal experiments add
complexity, and are therefore even more likely to lead to large deviations
from current models’ predictions. This phenomenon is accentuated when
theoretical models are still under construction or poorly conceptualized,
which is a common situation in neuroimaging.
For those reasons, multimodal studies should probably be planned even
more carefully than others, and based on a solid theoretical framework for
sound interpretation.
“Fusion” and “combining” do not reflect the same operations in the litera-
ture. In their article, “How can PET/fMRI and EEG/MEG can be combined?”,
Horwitz and Poeppel (2002) distinguish three levels of combination. Here,
we briefly summarize and comment on them.
Generative models. The most ambitious and possibly most fruitful direction is
to establish generative models for which parameters are estimated from data
of different nature or provenance. Only a few examples of this can be found
in the literature. One was proposed by Horwitz et al. (1999) who considered
the possibility of constructing a large-scale, biologically realistic neural
network model that can perform a specific cognitive task. The model con-
struction would allow the simulation of fMRI/PET and MEG/EEG data that
could be compared to experimentally observed values. Here, different data
types with different spatiotemporal properties are not directly compared, but
are informing a common neural model. Note that this model should include
not only current knowledge about the neurophysiology and the measurement
process, but also how the full model can achieve the cognitive operation under
study. Pushing the idea further, one can use the data of different imaging or
non-imaging techniques to estimate the parameters of the model.
Assumptions
could play a key role, for instance, in bridging the gap between modalities such
as fMRI and MEG, using conjoint recordings in both cases. However, it is
certainly not impossible that a different dynamic can be found within the
same network of regions between the two acquisitions, which should permit
the use of a common localization, but not of a common dynamical description.
In summary, non-simultaneous acquisitions rely on the strong assump-
tion of stationary signals in space and in time, and this needs to be further
validated.
In the following, we review the techniques used to analyze jointly
non-simultaneous signals (MEG-fMRI, non-simultaneous EEG-fMRI).
Distributed Models
If dipolar source reconstruction is the most widely used technique, some
authors have employed other localization techniques, such as the distributed
source model, to compare the results from different modalities. For example,
Moradi et al. (2003) have compared the localization obtained in a retinotopic
experiment using fMRI with those obtained using Magnetic Field Tomography
(MFT). They found small discrepancies between the two experiments, on the
order of 3 to 5 mm.
Since the two data sets share a common spatial dimension, it is, in principle,
simple to look for the instances where the spatial information is the most
coherent. This necessitates the reconstruction of distributed sources at a reso-
lution comparable to the fMRI, or degrading the fMRI spatial information to
the point that it can be compared to MEG/EEG. Note that using the fMRI
information to reconstruct the MEG/EEG source at this stage would bias
results. Once the two data sets have exactly the same number of “voxels,” it is
a straightforward matter to compute their cross-correlation. The result is a
matrix (fMRI_time X MEG/EEG_time) in which high values indicate the
times for which both electrical and metabolic maps are most correlated.
The cross-correlation can then be decomposed by standard singular-value
decomposition techniques, and corresponding BOLD pattern and MEG/EEG
timing extracted.
To quantify this equivalence in an experiment combining MEG and fMRI
to study the sensory cortex representation of digit and lip, Schultz et al. (2004)
describe the reconstruction of MEG activity with the SAM beamformer tech-
nique (SAM: Synthetic Aperture Magnetometry) and obtained an image with
space-dimension equivalent to fMRI. The SAM images are constructed for dif-
ferent frequency bands, and a voxel-per-voxel multiplication is performed
between the unthresholded fMRI and SAM images. The results are summarized
through standard singular-value decomposition. The authors report some areas
within SI that are not found by either fMRI nor MEG. While the idea is interest-
ing, the interpretation of the voxel-per-voxel multiplication of SAM and fMRI
images is not clear, raises intensity and spatial normalization issues, and will
cancel out signals that are not present in only one of the two modalities.
A similar approach considers the space defined by the scalp as the shared
common space. While this would strongly diminish the fMRI resolution, it
avoids the difficult source reconstruction step. Assuming—again—that the
BOLD activity will coincide with sources having significant electrical energy,
a projection onto the scalp (or electrode) space would make the two data sets
directly comparable. The MEG/EEG data would then be transformed to retain
only the energy, and the same dimension-reduction techniques can be further
used to decompose this cross-correlation and extract the conjoint time/space
information.
However, because of the different nature of the data (as reviewed in the
introduction), this approach raises several questions: (1) What fMRI infor-
mation should be used? Spatial localization, or amount of activation?
Information resulting from group or individual analysis? (2) Which types of
constraints should be implemented—hard constraints (assuming complete
equivalence between MEG/EEG and fMRI localizations), or enabling extra
sources or small displacements (soft constraints)? (3) What is the influence of
those constraints in the results?
As these different concerns depend on the EEG/MEG inverse method, we
will address these questions separately for dipolar and distributed-source
reconstructions. Those methods are described in detail in other chapters of
this book, so we will review the literature only with respect to combining
multimodal information. See also the review by Halchenko et al. (2005) on
this topic.
They directly include the percentage increase of the fMRI signal in the diagonal
elements of the matrix, and introduce in the non-diagonal elements the cor-
relation of the fMRI signals between the two corresponding areas. Simulations
performed with this method showed that in the case of concordant activa-
tions, results are superior to those of a standard inverse problem with no
constraints—and, crucially, solutions are equivalent when active sources do
not correspond to fMRI spots.
Recently, Ahlfors et al. (2004) formulated the problem as a geometrical prob-
lem. This is a classic interpretation of the linear model, as the least-square esti-
mate can be viewed as the orthogonal projection of the data onto a space spanned
by the columns of the model. With this view, the bias introduced by fMRI
data can be seen as a non-orthogonal projection, defined by fMRI activity.
To conclude the discussion of methods with hard or soft constraints: the
work of Im et al. (2005) again suggests a word of caution. They report, both
in a simulation and in a lexical judgment experiment, cases where modifying
the variance of the source in order to bias results towards fMRI data, may also
weaken or eliminate actual MEG sources. They also suggest a technique to use
in checking whether those sources are likely artifacts, or true sources. The
difficulty still lies in the lack of solid ground information on the actual local
coupling between MEG sources and fMRI/BOLD activity.
Finally, Daunizeau et al. (2006) have recently presented an interesting
framework to test the relevance of fMRI priors in the inverse problem, using
a Bayesian formalism. They constructed different priors under two different
hypotheses. Under the first one, H0, there is no correspondence between
fMRI and MEG/EEG sources (the priors are simply defined as a independent
zero-mean Gaussian with constant variance). Under the second hypothesis,
H1, a link is assumed between the fMRI activation map and the source inten-
sities, and the corresponding prior is modified by weighting the source vari-
ance by a factor that is proportional to a function of the fMRI activation.
These two hypotheses are then compared, by computing the ratio between
the posterior probability of the model H0 and H1, given the MEG/EEG
data M: log(P(H1/M)/P(H0/M)). If this ratio is positive, then the fMRI-
constrained inverse problem solution should be favored.
if this can be achieved in the foreseeable future. It should be noted that more
and more researchers are attempting to, at least partially, address the prob-
lem. Horwitz and colleagues in particular (Horwitz & Tagamets 1999; Husain
et al., 2004; Riera et al., 2005; see Horwitz & Glabus, 2005, for a review) have
pursued these ideas in the context of cognitive systems interactions. Recently,
Friston and colleagues (Friston et al., 2003) described a model to study effec-
tive connectivity that expresses activity at the neuronal level, but estimates the
parameters using fMRI data and the Balloon model (Buxton et al., 1998).
In Babajani et al. (2005), an integrated model for MEG and fMRI is
described in which the neural activity is related to the postsynaptic potentials
(PSPs). In each voxel, the neural activity is modeled by a MEG ECD, and as
input of extended balloon model in fMRI. The model shows that it is possible
to detect fMRI activity but no MEG activity, and vice versa. The model could
be used in the future to evaluate and compare different conjoint analysis
methods of MEG and fMRI.
The major advantage of this approach is that one could work with a
model able to reproduce the mechanism at the origin of signals—therefore,
predictions, validation or refutations should be much easier to perform. In a
sense, this is the grail that any neuroscientist is looking for. Unfortunately—
and this constitutes a major issue in this research—there is simply no such
model available at the level where it would be needed. While the balloon
model sounds like a good approximation of the mechanistic aspects of the
vascular properties of brain tissue, a full model that would truly permit the
reproduction of even simple tasks are few, and are hardly convincing. In other
words, while this direction of research seems to be the most appealing, it
might not be fruitful until much more is known about how a simulated neu-
ral network can mimic brain processes and performances. At the moment, it
is only too likely that those models are too restrictive, incomplete, or even
wrong, and therefore will not help the fusion of data from several modalities.
Nevertheless, it is worth noting the works that propose dynamic recurrent
network models relating neuronal electrophysiological data to fMRI or PET
(Corchs & Deco, 2002; Tagamets and Horwitz, 1998), as well as models
that relate neuronal data to MEG/EEG signals (Arezzo & Vaughan, 1988,
David et al., 2006).
First, for individual subject or group analyses, simultaneous EEG and fMRI
acquisition does not have to rely on the assumption of reproducible physiological
or cognitive state between two scanning sessions. Since it is not possible in
general to verify this assumption, the advantage is fundamental. At the level
286 MEG: An Introduction to Methods
of a group study, the additional assumptions are those necessary for fMRI
alone, i.e., that there exists a common spatial and functional space in which
the different subjects can be averaged.
Second, there are a number of research questions that can only be
addressed by simultaneous acquisitions. For example, in the study of the
vigilance state, to which alpha waves are associated (Goldman et al., 2002),
simultaneous recordings are mandatory. Another clear example of this is
shown in the sleep studies (Portas et al., 2000; Czisch et al., 2002; Maquet
et al., 2003 & 2005) and the epileptic activation (Lemieux et al., 2001; Salek-
Haddadi, 2002) for which the studied phenomenon is better defined through
external measurements. In general, it is possible that simultaneous recordings
are necessary for a correct interpretation of the fMRI data.
Third, joint recordings also allow one to investigate the link between the
two kinds of measures, and may therefore set the basis and the limit for their
combination in nonsimultaneous experiments. Here we describe further how
the EEG-fMRI data covariation can be investigated locally.
Acquisition Schemes
Summary technique
Extract relevant
temporal information
Convolve with
hemodynamic
response
EEG/MEG
design matrix
Construct Linear modal
statistical
map Sampling every RT
can then be used for detecting regions. It is also simple to interpret, as the
maps produced can be easily thought of as correlation maps. The method is
flexible and has successfully been used in a number of previous works. The
works of Laufs et al. (2003), Goldman et al. (2002), and others who have
found regions where hemodynamic functions correlate most with increased
alpha- or beta-frequency power, are based on this method. More recently,
Foucher et al. (2003) show that the gamma-band power is related to BOLD
activity in an oddball detection task. In Debener et al. (2005), EEG single-trial
electric activity is computed as the minimum value in a temporal window,
defined using the grand average of independent component analysis results.
These amplitudes are convolved with the canonical hemodynamic response,
and used as regressors for the fMRI analysis. This trial-by-trial EEG measure
predicted the fMRI activity in the rostral cingulate zone, a brain region
thought to play a key role in processing of response errors. Using the same
principle, Parkes et al. (2006) investigate the post-movement beta rebound
(an increased in beta-frequency power following movement) and use a time-
frequency analysis of the EEG signal to construct regressors for the fMRI
analysis—and, with the EEG, detect an additional region in the post-central
sulcus for this task. Interestingly, signals of no interest can be treated simi-
larly, to remove fMRI artifacts such as the cardiac beat (Liston et al., 2005).
These are but very few of the examples that can be found in the literature, and
it is expected that many more studies will use this method to relate the two
types of information.
290 MEG: An Introduction to Methods
With the regression technique, a map is constructed. This permits the descrip-
tion of local phenomena and, to some extent, their interpretation. However,
the actual local correlation of the electrical signal with the BOLD signal is
never computed, and cannot be until a local estimation of the EEG signal
is estimated through an inverse solution. In Lahaye et al. (2004), this local
correlation is computed and the locations of high correlations are compared
to the position of the highest BOLD signal and EEG signals. This is a challeng-
ing task, since the number of electrodes used for simultaneous acquisitions,
and the associated signal-to-noise ratio issues, make the source reconstruction
particularly difficult. In this work, a parcellation of the surface of the brain was
computed to reduce the number of possible sources to 1000 per hemisphere,
and the local hemodynamic responses were computed on those parcels
(Ciuciu et al., 2003). The correlation could then be computed using a local
regression analysis technique, by convolving the EEG source-reconstructed
signal of interest by the local hemodynamic function. Because the experimental
paradigm was designed as event-related, it was also possible to correlate each
event BOLD-magnitude to the event EEG-summary measure. Preliminary
results show that the location of the highest correlation does not coincide
closely either with the location of the reconstructed sources with the strongest
energy, nor to the mean maximum BOLD value across events. This confirms
other findings that indirectly compared the positions of the MEG/EEG recon-
structed sources and the BOLD activity.
This local coupling can then be used to constrain the inverse problem in
an iterative scheme. Lahaye et al. (2004) have proposed to weight differently
the source covariance matrix, using the coupling coefficient derived locally on
Combining Neuroimaging Techniques: The Future 291
a trial by trial basis. This allows a new estimate of an optimized spatial filter in
a beamformer formalism. fMRI priors on localization are introduced only
where a significant coupling is demonstrated.
One difficulty with the standard regression analysis described above is that
the EEG data to be convolved and regressed must be specified, and may not
be clearly known from the literature. To solve for this, Martinez-Montes
(2004) proposes a multivariate linear analysis using a technique originally
developed in the domain of chemometrics. Its basic idea is to extend the
Partial Least Square technique for data with higher dimensions than are pos-
sible with PLS. fMRI data has dimension time and voxels, while the EEG data
has dimension time (shared dimension), frequency, and electrodes. The
method proposed will find the weights of a BOLD image and the weights of
electrodes and frequencies, such that the correlation of the two time dimen-
sions is maximized. In other words, the technique will find a BOLD image on
the one hand, a frequency spectrum, and the topographic map on the other
hand. In this way, the linear combination of the BOLD time series weighted
by the found BOLD image will maximally correlate with the linear combina-
tion of the EEG time series weighted by the topographic map and frequency
spectrum. The method is interesting as a means to quickly summarize the
important features of the covariation of such large data. It does require that a
constant model of the hemodynamic response is chosen a priori, and that the
EEG data are preprocessed to remove muscle and motion artifacts. When
applied to the data used by Goldman et al. (2002), the results show that the
time dimensions are well correlated in the alpha band, less in the theta, and
not in the gamma. Spatially, the alpha and gamma spatial representation
is found to be in the occipital, while the theta has a pattern less clearly
interpretable.
Conclusion
The combination of neuroimaging techniques is an intense research domain,
which will grow in the future due to the increasing number of noninvasive
imaging techniques and their technological advances. If this chapter is mainly
concerned with the fusion between electromagnetic and hemodynamic
signals for brain mapping, other imaging modalities or information can be
considered (diffusion MRI, optical imaging, etc.) but the corresponding
literature is still meager on the use of those with MEG/EEG data.
Could a MEG with good spatial localization power replace conjoint
EEG-fMRI recording?
With the progress of inverse resolution techniques in MEG, one can ask
the question whether sufficient spatial and temporal resolution can be reached
292 MEG: An Introduction to Methods
References
Ahlfors, S. P., Simpson, G. V., Dale, A. M., Belliveau, J. W., Liu, A. K., Korvenoja, A.,
et al. (1999). Spatiotemporal activity of a cortical network for processing visual
motion revealed by MEG and fMRI. J Neurophysiol, 82(5), 2545–2555.
Ahlfors, S. P., & Simpson, G. V. (2004). Geometrical interpretation of fMRI-guided
MEG/EEG inverse estimates. Neuroimage, 22(1), 323–332.
Allen, P. J., Josephs, O., & Turner, R. (2002). A method for removing imaging artifact
from continuous EEG recorded during functional MRI. Neuroimage, 12(2), 230–239.
Arezzo, J. C., & Vaughan, H. G. Jr. (1982). The contribution of afferent fiber tracts to
the somatosensory evoked potential. Ann N Y Acad Sci, 388, 679–682.
Arthurs, O. J., Williams, E. J., Carpenter, T. A., Pickard, J. D., & Boniface, S. J. (2000).
Linear coupling between functional magnetic resonance imaging and evoked
potential amplitude in human somatosensory cortex. Neuroscience, 101, 803–806.
Babajani A, Nekooei MH, Soltanian-Zadeh H. Integrated MEG and fMRI model:
synthesis and analysis. Brain Topogr. 2005 Winter;18(2):101–13. Epub 2005 Dec 5.
PubMed PMID: 16341578.
Babiloni, F., Babiloni, C., Carducci, F., Romani, G. L., Rossini, P. M., Angelone, L. M.,
et al. (2003). Multimodal integration of high-resolution EEG and functional magne-
tic resonance imaging data: a simulation study. Neuroimage, 19(1), 1–15.
Babiloni, F., Cincotti, F., Babiloni, C., Carducci, F., Mattia, D., Astolfi, L., et al. (2005).
Estimation of the cortical functional connectivity with the multimodal integration
of high-resolution EEG and fMRI data by directed transfer function. Neuroimage,
1, 24(1), 118–131.
Babiloni, F., Mattia, D., Babiloni, C., Astolfi, L., Salinari, S., Basilisco, A., et al. (2004).
Multimodal integration of EEG, MEG and fMRI data for the solution of the neuro-
image puzzle. Magn Reson Imaging, 22(10), 1471–1476.
Bagshaw, A. P., Hawco, C., Benar, C. G., Kobayashi, E., Aghakhani, Y., Dubeau, F.,
et al. (2005). Analysis of the EEG-fMRI response to prolonged bursts of interictal
epileptiform activity. Neuroimage, 24(4), 1099–1112.
Bagshaw, A. P., Kobayashi, E., Dubeau, F., Pike, G. B., & Gotman, J. (2006). Corre-
spondence between EEG-fMRI and EEG dipole localisation of interictal discharges
in focal epilepsy. Neuroimage, 30, 417–425.
Ball, T., Schreiber, A., Feige, B., Wagner, M., Lucking, C. H., & Kristeva-Feige, R.
(1999). The role of higher-order motor areas in voluntary movement as revealed
by high-resolution EEG and fMRI. Neuroimage, 10(6), 682–694.
Belin, P., Zatorre, R. J., Lafaille, P., Ahad, P., & Pike, B. (2000). Voice-selective areas
in human auditory cortex. Nature, 403(6767), 309–312.
Benar, C., Aghakhani, Y., Wang, Y., Izenberg, A., Al-Asmi, A., Dubeau, F., et al. (2003).
Quality of EEG in simultaneous EEG-fMRI for epilepsy. Clin Neurophysiol, 114(3),
569–580.
294 MEG: An Introduction to Methods
Benar, C. G., Grova, C., Kobayashi, E., Bagshaw, A. P., Aghakhani, Y., Dubeau, F., et al.
(2006). EEG fMRI of epileptic spikes: Concordance with EEG source localization
and intracranial EEG. Neuroimage, 30, 1161–1170
Bonmassar, G., Schwartz, D. P., Liu, A. K., Kwong, K. K., Dale, A. M., & Belliveau, J. W.
(2001). Spatiotemporal brain imaging of visual-evoked activity using interleaved
EEG and fMRI recordings. Neuroimage, 13(6–1),1035–1043.
Brunetti, M., Belardinelli, P., Caulo, M., Del Gratta, C., Della Penna, S., Ferretti, A.,
et al. (2005). Human brain activation during passive listening to sounds from
different locations: an fMRI and MEG study. Hum Brain Mapp, 26(4), 251–261.
Buxton, R. B., Wong, E. C., & Frank, L. R. (1998). Dynamics of blood flow and
oxygenation changes during brain activation: the balloon model. Magn Reson Med,
39(6), 855–864.
Ciuciu, P., Poline, J.-B., Marrelec, G., Idier, J., Pallier, Ch., & Benali, H. (2003). Unsu-
pervised robust non-parametric estimation of the hemodynamic response function
for any fMRI experiment. IEEE Trans Med Imag, 22(10), 1235–1251.
Corchs, S., & Deco, G. (2002). Large-scale neural model for visual attention: integra-
tion of experimental single-cell and fMRI data. Cereb Cortex, 12(4), 339–48.
Czisch, M., Wetter, T. C., Kaufmann, C., Pollmacher, T., Holsboer, F., & Auer, D. P.
(2002). Altered processing of acoustic stimuli during sleep: reduced auditory
activation and visual deactivation detected by a combined fMRI/EEG study.
Neuroimage, 16(1), 251–258.
Dale, A. M., Liu, A. K., Fischl, B. R., Buckner, R. L., Belliveau, J. W., Lewine, J. D.,
et al. (2000) Dynamic statistical parametric mapping: combining fMRI and MEG
for high-resolution imaging of cortical activity. Neuron, 26(1), 55–67.
Daunizeau, J., Mattout, J., Clonda, D., Goulard, B., Benali, H., & Lina, J. M. (2006)
Bayesian spatio-temporal approach for EEG source reconstruction: conciliating
ECD and distributed models. IEEE transactions on bio-medical engineering 53,
503–516.
David, O., Kiebel, S. J., Harrison, L. M., Mattout, J., Kilner, J. M., & Friston, K. J.
(2006). Dynamic causal modeling of evoked responses in EEG and MEG. Neuro-
image, 30, 1255–1272.
Debener, S., Ullsperger, M., Siegel, M., Fiehler, K., von Cramon, D. Y., Engel, A. K.
(2005) Trial-by-trial coupling of concurrent electroencephalogram and functional
magnetic resonance imaging identifies the dynamics of performance monitoring.
J Neurosci, 25, 11730–11737.
Dehaene, S., Naccache, L., Le Clec, H. G., Koechlin, E., Mueller, M., Dehaene-
Lambertz, G., et al. (1998). Imaging unconscious semantic priming. Nature, 395,
597–600.
Del Gratta, C., Della Penna, S., Ferretti, A., Franciotti, R., Pizzella, V., Tartaro, A.,
et al. (2002). Topographic organization of the human primary and secondary
somatosensory cortices: comparison of fMRI and MEG findings. Neuroimage,
17(3), 1373–1383.
Devor, A., Dunn, A. K., Andermann, M. L., Ulbert, I., Boas, D. A., & Dale, A. M.
(2003). Coupling of total hemoglobin concentration, oxygenation, and neural
activity in rat somatosensory cortex. Neuron, 39(2), 353–359.
Downing, P., Liu, J., & Kanwisher, N. (2001). Testing cognitive models of visual
attention with fMRI and MEG. Neuropsychologia, 39(12), 1329–1342.
Combining Neuroimaging Techniques: The Future 295
Duong, T, Q., Kim, D. S., Ugurbil, K., & Kim S. G.(2000). Spatiotemporal dynamics
of the BOLD fMRI signals: toward mapping submillimeter cortical columns using
the early negative response. Magn Reson Med, 44(2), 231–242.
Foucher, J. R., Otzenberger, H., & Gounot, D. (2003). The BOLD response and the
gamma oscillations respond differently than evoked potentials: an interleaved
EEG-fMRI study. BMC Neurosci, 4, 22.
Foucher, J. R., Otzenberger, H., & Gounot, D. Where arousal meets attention: a
simultaneous fMRI and EEG recording study. Neuroimage, 22(2), 688–697.
Friston, K. J., Harrison, L., & Penny, W. (2003). Dynamic causal modelling.
Neuroimage, 19(4), 1273–1302.
Fujimaki, N., Hayakawa, T., Nielsen, M., Knosche, T. R., & Miyauchi, S. (2002). An
fMRI-constrained MEG source analysis with procedures for dividing and grouping
activation. Neuroimage, 17(1), 324–343.
Garreffa, G., Bianciardi, M., Hagberg, G. E., Macaluso, E., Marciani, M. G., Maraviglia, B.,
et al. (2004). Simultaneous EEG-fMRI acquisition: how far is it from being a stan-
dardized technique? Magn Reson Imaging, 22(10), 1445–1455.
Goldman, R. I., Stern, J. M., Engel, J. Jr., & Cohen, M. S. (2002). Simultaneous EEG
and fMRI of the alpha rhythm. Neuroreport, 13(18), 2487–2492.
Goncalves, S. I., de Munck, J. C., Pouwels, P. J., Schoonhoven, R., Kuijer, J. P.,
Maurits, N. M., et al. (2005). Correlating the alpha rhythm to BOLD using simulta-
neous EEG/fMRI: Inter-subject variability. Neuroimage, 30, 203–213.
Gotman, J., Benar, C. G., & Dubeau, F. (2004). Combining EEG and FMRI in epilepsy:
methodological challenges and clinical results. J Clin Neurophysiol, 21(4), 229–240.
Guy, C. N., ffytche, D. H., Brovelli, A., & Chumillas, J. fMRI and EEG responses to
periodic visual stimulation. Neuroimage, 10(2), 125–148.
Halchenko, Y.O., Hanson, S.J. & Pearlmutter, B. A. (2005). MultimodalIntegration:
fMRI MRI, EEG, MEG, in Advanced Image Processing. In: Dekker (Ed.), Magnetic
Resonance Imaging. Boston, MA:MIT Press. (pp. 223–266).
Horwitz, B., & Glabus, M. F. (2005). Neural modeling and functional brain imaging:
the interplay between the data-fitting and simulation approaches. Int Rev Neurobiol,
66, 267–290.
Horwitz, B., & Poeppel, D.(2002). How can EEG/MEG and fMRI/PET data be combi-
ned? Hum Brain Mapp, 17(1), 1–3.
Horwitz, B., Tagamets, M. A., & McIntosh, A. R. (1999). Neural modeling, functional
brain imaging, and cognition. Trends Cogn Sci, 3(3), 91–98.
Horwitz, B., & Tagamets, M. A. (1999). Predicting human functional maps with
neural net modeling. Hum Brain Mapp, 8(2–3), 137–142.
Horwitz, B., Warner, B., Fitzer, J., Tagamets., M. A., Husain, F.T., & Long, T. W.
(2005). Investigating the neural basis for functional and effective connectivity.
Application to fMRI. Philos Trans R Soc Lond B Biol Sci, 360(1457), 1093–1108.
Husain, F. T., Tagamets, M. A., Fromm, S. J., Braun, A. R., & Horwitz, B. (2004). Rela-
ting neuronal dynamics for auditory object processing to neuroimaging activity:
a computational modeling and an fMRI study. Neuroimage, 21(4), 1701–1720.
Im, C. H., Jung, H. K., & Fujimaki, N. (2005). fMRI-constrained MEG source imaging
and consideration of fMRI invisible sources. Hum Brain Mapp, 26(2), 110–118.
Ioannides, A. A. (1999). Problems associated with the combination of MEG and fMRI
data: theoretical basis and results in practice. In: Yoshimoto, T., Kotani, M., Kuriki, S.,
296 MEG: An Introduction to Methods
Logothetis, N. K., Pauls, J., Augath, M., Trinath, T., & Oeltermann, A. (2001).
Neurophysiological investigation of the basis of the fMRI signal. Nature, 412(6843),
150–157.
Logothetis, N. K., & Pfeuffer, J. (2004). On the nature of the BOLD fMRI contrast
mechanism. Magn Reson Imaging, 22(10), 1517–1531.
Logothetis, N. K. (2003). The underpinnings of the BOLD functional magnetic
resonance imaging signal. J Neurosci, 23(10), 3963–3971.
Maquet, P., Peigneux, P., Laureys, S., Boly, M., Dang-Vu, T., Desseilles, M., et al.
(2003). Memory processing during human sleep as assessed by functional neuro-
imaging. Rev Neurol (Paris), 159(11 Suppl), 6S27–6S29.
Maquet, P., Ruby, P., Maudoux, A., Albouy, G., Sterpenich, V., Dang-Vu, T., et al. (2005).
Human cognition during REM sleep and the activity profile within frontal and
parietal cortices: a reappraisal of functional neuroimaging data. Prog Brain Res,
150, 219–227.
Martinez-Montes, E., Valdes-Sosa, P. A., Miwakeichi, F., Goldman, R. I., & Cohen,
M. S. (2004). Concurrent EEG/fMRI analysis by multiway Partial Least Squares.
Neuroimage, 22(3), 1023–34. Erratum in: (2005) Neuroimage, 26(3), 973.
Mizuhara, H., Wang, L. Q., Kobayashi, K., & Yamaguchi, Y. (2005). Long-range
EEG phase synchronization during an arithmetic task indexes a coherent cortical
network simultaneously measured by fMRI. Neuroimage, 27(3), 553–563.
Moradi, F., Liu, L. C., Cheng, K., Waggoner, R. A., Tanaka, K., Ioannides, A. A. (2003)
Consistent and precise localization of brain activity in human primary visual cortex
by MEG and fMRI. Neuroimage, 18(3), 595–609.
Neumann, J., Lohmann, G., Zysset, S., & von Cramon, D. Y. (2003). Within-subject
variability of BOLD response dynamics. Neuroimage, 19(3), 784–796.
Niazy, R. K., Beckmann, C. F., Iannetti, G. D., Brady, J. M., & Smith, S. M. (2005).
Removal of FMRI environment artifacts from EEG data using optimal basis sets.
Neuroimage, 28(3), 720–737.
Northoff, G., Witzel, T., Richter, A., Gessner, M., Schlagenhauf, F., Fell, J., et al. (2002).
GABA-ergic modulation of prefrontal spatio-temporal activation pattern during
emotional processing: a combined fMRI/MEG study with placebo and lorazepam.
J Cogn Neurosci, 14(3), 348–370.
Nunez, P. L., & Silberstein, R. B. (2000). On the relationship of synaptic activity to
macroscopic measurements: does co-registration of EEG with fMRI make sense?
Brain Topogr, 13(2), 79–96.
Parkes, L. M., Bastiaansen, M. C., & Norris, D. G. (2006). Combining EEG and fMRI
to investigate the post-movement beta rebound. Neuroimage, 29(3), 685–696.
Portas, C. M., Krakow, K., Allen, P., Josephs, O., Armony, J. L., & Frith, C. D. (2000).
Auditory processing across the sleep-wake cycle: simultaneous EEG and fMRI
monitoring in humans. Neuron, 28(3), 991–999.
Pouthas, V., Garnero, L., Ferrandez, A. M., & Renault, B. (2000). ERPs and PET ana-
lysis of time perception: spatial and temporal brain mapping during visual discri-
mination tasks. Hum Brain Mapp, 10(2), 49–60.
Riera, J., Aubert, E., Iwata, K., Kawashima, R., Wan, X., & Ozaki, T. (2005).
Fusing EEG and fMRI based on a bottom-up model: inferring activation and effec-
tive connectivity in neural masses. Philos Trans R Soc Lond B Biol Sci, 360(1457),
1025–1041.
298 MEG: An Introduction to Methods
Tuunanen, P. I., Kavec, M., Jousmaki, V., Usenius, J. P., Hari, R., Salmelin, R., et al.
(2003). Comparison of BOLD fMRI and MEG characteristics to vibrotactile stimu-
lation. Neuroimage, 19(4), 1778–1786.
Waites, A. B., Shaw, M. E., Briellmann, R. S., Labate, A., Abbott, D. F., & Jackson, G. D.
(2005). How reliable are fMRI-EEG studies of epilepsy? A nonparametric approach
to analysis validation and optimization. Neuroimage, 24(1), 192–199.
Wan, X., Iwata, K., Riera, J., Ozaki, T., Kitamura, M., & Kawashima, R. (2006).
Artifact reduction for EEG/fMRI recording: Nonlinear reduction of ballistocardio-
gram artifacts. Clin Neurophysiol, 117, 668–680.
12
Somatosensory and Motor Function
Ryusuke Kakigi and Nina Forss
Somatosensory Function
Introduction
In the twenty years since the averaged MEG values following somatosensory
stimulation, i.e., the somatosensory evoked magnetic field (SEF), were first
reported (Brenner et al., 1978; Kaufman et al., 1981; Hari et al., 1983, 1985;
Wood et al., 1985; Sutherling et al., 1988), many studies have been conducted,
and their number continues to increase. In this chapter, therefore, we will
introduce basic methods for the beginner; that is, how to record a clear SEF.
In addition, we will introduce basic information and findings related to SEF,
particularly unique and interesting aspects.
Since the landmark studies of Foerster (1936) and Penfield and Boldrey
(1937) on the motor and sensory representations in the human cerebral
300
Somatosensory and Motor Function 301
cortex based on direct electrical stimulation of the cortical surface, it has been
established that the primary sensorimotor cortex is organized in an orderly
somatotopic way, which has been termed the ‘homunculus’ representation of
the cutaneous body surface. SEFs following stimulation applied to various
parts of the body in normal subjects have been reported to examine the
homunculus noninvasively.
MEG detects only a specific orientation of brain current tangential to the
skull. Therefore, dipoles generated in area 3b of the primary somatosensory
cortex (SI) and/or 4 of the primary motor cortex (MI), each of which is
located on the posterior and anterior bank of the central sulcus, respectively,
are easily detected—but dipoles in area 1 or 3a in SI, which is located on the
crown and the bottom of the central sulcus, respectively, are not (Figure 12–1).
Postcentral gyrus
Central sulcus Intraparietal sulcus
Posterior
parietal lobule
SI
SII
Lateral
sulcus
Postcentral gyrus
Intraparietal
Central sulcus
sulcus
1
7
4 3b 2
5
3a
Cutaneous
input Deep input
There are many important reports on the receptive sites following stimulation
of the lower limb (Hari et al., 1984; Kaukoranta et al., 1986; Huttunen et al.,
1987; Rogers et al., 1994; Kakigi et al., 1995a; Hari et al., 1996; Shimojo et al.,
1996a), the urogenital organs (Nakagawa et al. 1998), the truncus (Itomi
et al., 2000a), the neck and shoulder (Itomi et al., 2000b), the upper limb
(Huttunen et al., 1987; Tiihonen et al., 1989; Rossini et al., 1989, 1994;
Baumgartner et al., 1991; Suk et al., 1991; Gallen et al., 1994; Buchner et al.,
1994; Akhtari et al., 1994; Schnitzler et al., 1995a,b; Kawamura et al., 1996;
Mauguiere et al., 1997a,b; Xiang et al., 1997a; Shimizu et al., 1997; Tecchio
et al., 1998; Jousmaki and Hari, 1999; Wasaka et al., 2003, 2005; Inui et al.,
2004), face (Karhu et al., 1991; Mogilner et al., 1994; Hoshiyama et al., 1995,
1996; Nihashi et al., 2001, 2003; Nguyen et al., 2004, 2005; Sakamoto et al.,
2008a,b) and multiple sites (Narich et al., 1991; Yang et al., 1993; Gallen et al.,
1994; Nakamura et al., 1998; Inoue et al., 2005).
For example, we made a complete homunculus in 5 normal subjects
(Nakamura et al., 1998). We recorded SEF following stimulation of 19 sites—
tongue, lower lip, upper lip, thumb, index finger, middle finger, ring finger,
304 MEG: An Introduction to Methods
little finger, radial palm, ulnar palm, forearm, elbow, upper arm, chest, thigh,
ankle, big toe, second toe and fifth toe—and put their ECD on the MRI of
each subject (Figure 12–3). These representative areas were generally arranged
in the above order from inferior to superior, lateral to medial, and anterior to
posterior. The changes in the coordinates were compatible with the anatomy
of the central sulcus and the homunculus. The location of the ECD for the
upper lip could be distinguished from that on the lower lip, with the former
positioned more superior than the latter in all subjects. Each representation
of the thumb, index finger, middle finger, ring finger and little finger was
distinguishable. They were represented sequentially from thumb to little
finger, ascending the postcentral sulcus. Next, we introduce some interesting
findings.
ligament—in 7 normal subjects (14 limbs; see Figure 12–4 and Shimojo
et al., 1996b). The ECDs of the 14 limbs were classified into two types accord-
ing to the distance of ECD between PT and FE; Type 1 (>1 cm, nine limbs)
and Type 2 (<1 cm, five limbs) (Figure 12–5). The ECD following FE stimula-
tion was located on the crown of the postcentral gyrus or at the edge of the
interhemispheric fissure in Type 1, and was close to the ECDs following PT
and SU stimulation along the interhemispheric fissure in Type 2. The ECD
following PE stimulation was located along the interhemispheric fissure in all
14 limbs, as for PT and SU. Its location was slightly but significantly higher
than that of PT and SU stimulation in Type 1, and was close to the ECDs fol-
lowing PT and SU stimulation in Type 2. The present findings indicated that
approximately 65 % (9 of 14) of the limbs showed particular receptive fields
compatible with the homunculus. The large inter- and intraindividual (left-
right) differences found in this study indicated significant anatomical varia-
tions in the area of the lower limb in the sensory cortex in humans.
Figure 12–4. Chart showing SEFs following stimulation of the posterior tib-
ial and sural nerve at the ankle, the peroneal nerve at the knee, and the
femoral nerve overlying the inguinal ligamentum of the right lower limb
in one subject. Waveforms recorded at 37 channels are superimposed.
Four components indicated by arrows are identified in each waveform.
Adapted from Shimojo et al. (1996a).
306 MEG: An Introduction to Methods
Figure 12–5. MRIs showing the location and direction of ECDs of the 1M
following stimulation of 4 nerves of the right lower limb in 2 subjects.
In Subject 1, the ECD following the femoral nerve stimulation is located
on the crown of the postcentral gyrus, directed to the inferior and post-
erior sides. In contrast, the ECDs following stimulation of the other nerves
are located along the interhemispheric fissure directed to the right hemi-
sphere. The ECDs of the other nerves are located very close together,
but that following peroneal nerve stimulation is slightly higher than that
following the stimulation of the other two nerves. This type of receptive
field is classified as Type 1. In Subject 2, the ECDs following the stimula-
tion of each nerve are located close together, along the interhemispheric
fissure. Those following stimulation of the posterior tibial and sural nerve
were directed to the right hemisphere horizontally, but those following
stimulation of the peroneal and femoral nerves were directed anteriorly
and posteriorly, respectively. This type of receptive field is classified as
Type 2. L = left, R = right.
Adapted from Shimojo et al. (1996a).
Somatosensory and Motor Function 307
Ear Stimulation
Figure 12–6. Procedures and results of the data analysis. (A) Sensor lay-
out. (B) Superimposed waveforms recorded from 37 channels (a), residual
magnetic fields obtained by a subtraction of those due to one (b), two
(c), three (d), four (e) and six (f) sources determined from the recorded
data. Isocontour maps at the peak latency of a selected deflection (verti-
cal bars) are shown on the right side of each trace. (C) Time course of each
strength. (D) Schematic drawings of the location and orientation of each
source. Bars indicate the direction of upward deflections of the corre-
sponding waveforms in (C). (E) Superimposition of sources on a subject’s
brain surface image. White circles in (A) and isocontour maps indicate the
position of the sensor (channel 3) that is just on the central sulcus. SII+,
secondary somatosensory cortex plus adjoining areas; PPC, posterior pari-
etal cortex; RV, residual variance.
Adapted from Inui et al. (2004).
Somatosensory and Motor Function 309
1
Source 1 2
Helix (posterior) 2
20 nAm
Lobulus (middle) Source 2 1
L R
Tragus (anterior) L R
50 100 msec 1
Analysis time=40–56 msec 2
GOF=95.4%
R
1
50 100 msec 2
Analysis time=44–60 msec
GOF=96.3% R
matter), a two-dipole model—(1) the neck area of SI and (2) face area of
SI—was found to be the most appropriate (Figures 12–7, 12–8, 12–9). These
results indicated that receptive fields of some parts of the ear, such as the
lobulus and tragus, might be present in both the neck and face areas of SI.
These findings suggested that the “ear area” of SI has variability between sub-
jects, unlike the other areas of SI, possibly because the ear is located on the
border between the neck and face. We confirmed this finding by fMRI
(Nihashi et al., 2002)
Lobulus Tragus
Source 1
Source 2
stimulation was located near that of the face stimulation (Figure 12–9). This
may be due to anatomical variations in the subjects.
Face Stimulation
Since Penfield and Boldrey (1937) revealed the somatotopic body surface
representation in the primary somatosensory cortex (SI), many studies using
various methods have confirmed this somatosensory homunculus. Regarding
the representation of the face in the SI, the face area drawn by Penfield and
Boldrey (1937) is organized along the central sulcus with the forehead in the
superiomedial region adjacent to the hand area, and the chin in the inferiolat-
eral region. Many authors have reported the locations in the SI of the lip
(Nakamura et al., 1998, Hoshiyama et al., 1995; Mogilner et al., 1994), tongue
(Karhu et al., 1991; Sakamoto et al., 2008a,b), oral cavity (Hari et al., 1993),
and ear (Nihashi et al., 2003, Nihashi et al., 2002, Nihashi et al., 2001).
However, there are only a few reports (Servos et al., 1999 and Yang et al.,
1994) on skin-covered areas of the face such as the forehead, cheek and chin
in humans. Interestingly, some results showed the representation of an
inverted face along the central sulcus of the human brain (Servos et al., 1999,
Yang et al., 1994, Pons et al., 1991), which is not consistent with the homun-
culus map drawn by Penfield and Boldrey (1937). Therefore, to investigate
the representation of facial skin areas in SI, we recorded magnetic fields
evoked by air-pressure-induced tactile stimulation applied to six points on
the face, lower lip and thumb (Figure 12–10). The thumb area in the SI was
located more medial and superior to the lip area, which was consistent with
Somatosensory and Motor Function 311
Topography of SII
One of the major advantages of SEF is that it easily records activities in the SII,
where it is difficult for SEP to detect activities due to the location and direction
Figure 12–10. Schematic drawing of the points stimulated on the face – six
sites were stimulated as shown in the figure.
Adapted from Nguyen et al. (2004).
312
Somatosensory and Motor Function 313
of dipole sources (Hari et al., 1983, 1990, 1993; Elbert et al., 1995a; Forss et al.,
1995, 1998; Mima et al., 1997, 1998a). A random or long interstimulus inter-
val stimulation rate (Wikstrom et al., 1996; Nagamine et al., 1998), and the
oddball paradigm, using rare and frequent, or target and non-target stimuli,
(Hari et al., 1990, 1993; Forss et al., 1995; Mima et al., 1998a) cause the SII
components to increase in amplitude. This finding suggested that SII activi-
ties are more affected by volitional or attention effects than SI activities.
We analyzed the topography of SII on somatosensory stimulation applied
to various parts of the body of normal subjects using SEF (Maeda et al., 1999;
Nguyen et al., 2005; Sakamoto et al., 2008a,b). SII components were found
about 80-100 ms after the stimulation as middle-latency components. SII in
the bilateral hemisphere was activated on stimulation of the unilateral side of
the body; that is, SII in humans has a “bilateral function.” Although there
were large interindividual differences, the receptive fields ranked (Figure 12–12)
as follows: (1) Anterior–posterior direction; lower lip – upper lip – thumb –
middle finger – foot; (2) Medial–lateral direction; foot – middle finger –
thumb – upper lip – lower lip; and (3) Lower–upper direction; lower
lip – upper lip – thumb – middle finger – foot. In general, these findings are
similar to those obtained in studies of animals (Whitzel et al., 1969) and
humans (Hari et al., 1993). However, the differentiation was not as clear as
that seen in the homunculus in the SI. The SII is located anterior to, medial
to, and above the auditory cortex.
Plasticity of the SI is one of the most interesting topics in the study of SEF.
A change of homunculus is reported to be due to limb deafferentation after
amputation (Yang et al., 1994; Elbert et al., 1994, 1997; Flor et al., 1995;
Knecht et al., 1995, 1996, 1998; Weiss et al., 1998). Yang et al. (1994) and
Elbert et al. (1994) first reported the marked intrusion of facial representa-
tions into the digit and hand area after upper limb amputation. Further,
Knecht et al. (1996) reported that phantom sensations could be evoked from
sites on the face and the trunk ipsilateral, but also contralateral to the ampu-
tation, and that the amount of reorganization strongly correlates with the
number of sites, be it ipsi- or contralateral, from where painful stimuli evoked
the referred sensation. These findings suggested the involvement of bilateral
pathways, and demonstrated that the perceptual changes go beyond what can
be explained by shifts in neighboring cortical representational zones.
Mogilner et al. (1993) reported somatosensory cortical plasticity in
patients who were studied before and after surgery for webbed fingers (syn-
dactyly). The presurgical maps displayed shrunken and nonsomatotopic rep-
resentations of the hand. Within weeks of the surgery, cortical reorganization
occurring over distances of 3–9 mm was evident, correlating with the new
functional status of the separated digits. Such a reorganization of the SI was
also reported in patients with stroke and neoplasm (Rossini et al., 1998a,b).
314 MEG: An Introduction to Methods
studied SEFs in blind multifinger Braille readers. They found that the cortical
somatosensory representation of the fingers was frequently topographically
disordered in these subjects; in addition, the subjects frequently misperceived
which of their fingers was being touched by a light tactile stimulus. Therefore,
use-dependent cortical reorganization can be associated with functionally rel-
evant changes in the perceptual and behavioral capacities of the individual.
subcortical structures in the area of the middle cerebral artery. Patients with
pure motor stroke showed no alterations in SEFs, but patients with pure sen-
sory stroke showed markedly attenuated or absent SEFs. Abnormal SEF find-
ings were more clearly correlated with an impairment of two-point discrimi-
nation than of joint-position or vibration senses.
A SEP of large amplitude (giant SEP) is recorded in patients with cortical
reflex myoclonus. We reported that giant SEPs are generated in area 3b of the
SI (Kakigi et al., 1987). Recent SEF studies confirmed this hypothesis (Uesaka
et al., 1993, 1996; Karhu et al., 1994; Mima et al., 1998b). In addition, Mima
et al. (1998b) found other components located in the anterior bank of the
central sulcus, and suggested the importance of the motor cortex for generation
of the cortical reflex myoclonus.
Karhu et al. (1992) reported SEFs in 10 patients with multiple sclerosis.
Seven patients showed SEFs of abnormally large amplitude at 60–80 ms; 5 of
them had multiple lesions around lateral ventricles. In 2 patients with plaques
at the level of the 3rd and 4th ventricles and medulla, the 30-ms responses
were enlarged. The results suggest that early- and middle-latency SEF compo-
nents reflect the parallel processing of somatosensory inputs. Significant
changes not only in the SI but also in the SII responses has been also reported
in patients with genetically verified progressive myoclonus epilepsy (Forss
et al., 2001).
Pain Processing
Recently a number of studies have appeared where MEG has been applied in
studies of the human nociceptive system.The results have shown that MEG is
very suitable to study pain-related cortical areas, especially the cortical areas
processing the sensory aspects of the nociceptive signal. Prior MEG studies
have observed activation of the SII, PPC and SI to painful electric or laser
stimuli (Kakigi et al., 1995, 2003, 2005; Bromm 1996, Ploner et al., 1999, 2000;
Forss et al., 2005). In contrast, activation of deep brain structures that are
more related to emotional aspects of the pain processing (like anterior cingu-
lated cortex, anterior insula and amygdala) may be difficult to detect in MEG
Somatosensory and Motor Function 317
Thulium laser stimulates selectively nociceptive fibers, and has fast rise time
and short duration (typically 0.5 ms). Commercially available laser stimula-
tion equipment is nowadays safe and easy to use. Short laser pulses to skin
provide highly selective and temporally precise noxious stimuli that evoke
prominent EEG and magnetoencephalographic (MEG) responses. In MEG
measurements we have used laser stimuli (1 msec in duration, 2,000 nm in
wavelength) that were produced by a thulium-YAG stimulator (BLM 1000
Tm:YAG; Baasel Lasertech, Starnberg, Germany), and the laser beam was
conducted to the magnetically shielded room via an optic fiber. To keep the
distance stable between the optic fiber and the stimulated skin area, the hand
piece can be connected by a wire to the top of the neuromagnetometer.
An assistant directs the laser beam of approximately 10 mm2 to the skin.
To avoid skin burns and adaptation, the stimulus site should be moved after
each pulse to a random direction in the skin area of approximately 5 cm in
diameter. Stimulus intensity can be adjusted individually to equal twofold the
subjective pain threshold.
The majority of functional brain imaging studies on pain have described cor-
tical activation to Aδ-fiber mediated pain, or to a combination of Aδ-and
C-fiber pain, because it has been difficult to selectively stimulate the C-fiber
system without significant activation of the Aδ-fibers. For example, Ploner
and his coworkers 1999 showed that brief painful laser stimuli evoke
318 MEG: An Introduction to Methods
the diaphragm. The benefit is that this method allows flexible usage of the
C-fiber stimulation to any part of the body. Using this method, the Aδ- and
C-fiber responses were compared in 10 healthy subjects (Forss et al., 2005).
Laser evoked fields were measured to 1-ms thulium-laser stimuli delivered to
the dorsum of the subject’s left hand. The earliest cortical responses peaked at
165 ± 7 ms, agreeing with the conduction velocity of Aδ-fibers. To stimulate
unmyelinated C-fibers, the total energy of the laser beam was decreased and
the size of the stimulated skin area was restricted to 0.2–0.3 mm2. The earliest
cortical responses to these stimuli peaked at 811 ± 14 ms. In addition to the
consistent activation of the SII cortices, activation was observed in the poste-
rior parietal cortex (PPC). Activation of PPC to painful stimuli could be
related to the sensorimotor coordination that is needed to precisely define the
site of the painful stimuli with respect to other parts of the body, and the
outer space to reduce or prevent the pain. In contrast to some earlier studies,
our data did not indicate participation of the primary somatosensory cortex
(SI) in processing of the painful laser stimuli. The results imply that the noci-
ceptive inputs mediated by the two fiber systems are processed in a common
cortical network in different time windows. Characterization of cortical
responses to first and second pain offers a practical tool for clinical neurosci-
ence to study the two distinctive pain fiber systems.
The nociceptive fiber terminals are located in the epidermis and superficial
layer of the dermis (Kruger et al., 1985; Novotny & Gommert-Novotny,
1988), while other fibers run more deeply in the dermis (Munger & Halata,
1983). Therefore, we have recently developed a method utilizing a pushpin-
like needle electrode to stimulate the epidermal area for activating Aδ-fibers
(epidermal electrical stimulation, ES) (Inui et al., 2002a b, 2003a b, 2006)
(Figure 12–13). The soft stop device protrudes 1.0 mm from the plate, and the
tip of the needle, in turn, protrudes 0.2 mm from the soft stop device. By
pressing the electrode plate against the skin gently, the needle tip is inserted
adjacent to the free nerve endings of the thin myelinated fibers in the epider-
mis and superficial part of the dermis. The insertion of the needle electrode
causes no bleeding or visible damage to the skin. The stimulus intensity is
very small, approximately 0.2–0.3 mA. Compared with other stimuli, such as
laser stimulation and CO2 gas stimulation of the nasal mucosa, time locking
of the ES method is much better, since it is an electrical stimulus. Recently, we
developed a system to produce this ES needle with Nihon Kohden, Inc., and
can now offer the needle on request.
The early processing of pain perception can be analyzed in detail using
the ES needle. We used a multidipole model, BESA (MEGIS Software GmbH,
Munich, Germany, 1995). We recorded MEG in detail following not only ES
stimulation (see Chapter 3) but also TS (transcutaneous electrical) stimulation
to compare the results (Inui et al., 2003a,b). The TS was a conventional SEP,
with signals ascending through cutaneous (Aβ) fibers. TS activated two
sources sequentially within the SI, areas 3b and 1. ES (Aδ fibers) activated one
source within the SI, whose location and orientation were similar to those of
the TS-activated area-1 source (Figure 12–14). Activities from this source
consisted of three components peaking at 88, 98 and 109 ms, before the SI
activity reported in previous MEG studies (Ploner et al., 2000, Kanda et al.,
2000). The reason these studies failed to identify the early SI components
largely lies in the fact that early SI components are very weak and easily
overlooked.
Then, a clear and large component, 1M, whose peak latency was approxi-
mately 160 msec, was identified in both hemispheres. The generators for 1M
were the SI, SII and insula—mainly the SII and insula. The 1M recorded from
the ipsilateral hemisphere was significantly longer in latency in all the sub-
jects, and the interhemispheric difference in latency was approximately
10-20 ms. This difference probably indicates transcallosal transmission. We
found that the dipole for insular activity was located more anterior following
ES than TS, though SII activity showed no significant difference between the
two (Figure 12–15). The results suggested that cortical processing was similar
between noxious and innocuous stimulation in the SII, but different in the
insular cortex. The anterior location of pain-related activation in the insula
was consistent with the results of most functional imaging studies (for review,
see Schnitzler & Ploner, 2000). Neuroimaging studies have reported that
vibrotactile stimulations activated more posterior parts of the insula (Coghill
et al., 1994; Davis et al., 1998) in human studies, supporting our results.
Following 1M, a rather complicated component, 2M, whose peak latency
was approximately 250 ms, was identified. The main generators for 2M were
the cingulate cortex and midtemporal region (MT) around the amygdala
and/or hippocampus. Neuroimaging techniques such as positron emission
tomography (PET) and functional magnetic resonance imaging (fMRI) have
found extensive activity in the cingulate cortex following laser stimulation
(Xu et al., 1997, Svensson et al., 1997; Sawamoto et al., 2000; Peyron et al.,
2002; Qiu et al., 2006). Intracranial EEG recordings also clearly identified
strong activity there (Lenz et al., 1998). However, it has been rather difficult
for the MEG to detect it (Bromm et al., 1996; Watanabe et al., 1998; Yamasaki
et al., 1999; Ploner et al., 1999; Kanda et al., 2000; Nakata et al., 2004; Kakigi
et al., 1995b, 1996) with some exceptions (Kitamura et al., 1995, 1997, Inui et al.,
2003a,b and Figure 12–16, 12–17). This is probably due to the fact that the
dipoles generated in the right and left cingulate cortex cancel each other out,
which is inconvenient for MEG. In addition, a dipole generated in a deep
region such as the cingulate cortex is not easily detected by MEG.
The role of the MT region around the amygdala and hippocampus is still
controversial. Watanabe et al. (1998) and Inui et al. (2003a & 2000b),detected
it using MEG (Figure 12–16, 12–17). Garcia-Larrea et al. (2003) agreed with
the possibility that this region is activated by painful stimulation, since the
amygdala is thought to contribute to the emotional processing (i.e., aversive
nature) of painful events, rather than the sensory-discriminative aspects of
pain (Büchel et al., 1998 and Bornhövd et al., 2002), while activation of the
hippocampal formation seems to be enhanced when the pain is not expected
Figure 12–14. Comparison of cortical responses to noxious (ES) and innocu-
ous (TS) stimulation in a single subject. A: magnetic fields following TS
(transcutaneous stimulation); Aa, superimposed waveform recorded
from 37 channels; Ab-d, residual magnetic fields obtained by subtrac-
tion of those due to the 20/30-SI source (b), 20/30-SI and 26/36-SI sources
(c) and all three sources (d) from the recorded data. Isocontour maps at
the peak latency of a selected deflection (vertical bars) are shown on
the right side of each trace. B: time course of the source strengths in TS.
C: magnetic fields following ES (epidermal stimulation); Ca, recorded
data; Cb, c, residual magnetic fields obtained by a subtraction of those
due to the SI source (b) and SI and SII sources from the recorded data.
D: time course of the source strengths in ES. E: source locations overlaid on
MR images. SI, primary somatosensory cortex; SII, secondary somatosen-
sory cortex; 20/30-SI, the first SI source in TS whose activity peaked at
21 and 30 ms; 26/36-SI, the second SI source in TS whose activity peaked
at 26 and 36 ms.
Adapted from Inui et al. (2003b).
322
Somatosensory and Motor Function 323
(Ploghaus et al., 2000) or when the painful stimulus is associated with anxiety
(Ploghaus et al., 2001).
A thorough evaluation of pain using phased painful stimuli has been per-
formed with MEG. The biggest advantage of MEG is that one is able to clarify
the temporal information on pain processing, in the order of ms. However,
MEG cannot be used to evaluate continuous tonic pain such as cancer pain.
Neuroimaging techniques such as PET and fMRI may be more useful in this
respect, but physiological functions cannot be evaluated by neuroimaging.
One promising method in electrophysiological studies for this particular
problem is to analyze a change of frequency band using the fast Fourier trans-
form (FFT) during some period, for example, the alpha wave power change
in each region between a nonpainful state and painful state (Stancak et al.,
2005). This method will probably be used for not only basic research but also
clinical studies in the near future.
Safety Issues
Although thulium laser is probably safer than CO2 laser as a pain stimulus, it
may cause skin burns if the stimulator is too close to the skin or, alternatively,
Figure 12–16. Temporal profile of cortical activities following painful epi-
dermal stimulation (ES). Cortical responses to ES in a subject. The upper
three traces are superimposed waveforms recorded from 37 channels in
both hemispheres, and evoked potentials recorded at Cz. The lower seven
traces are temporal profiles of each source strength. Filled circles indicate
a group of early SI activities. Arrowheads indicate the peak latency of
early and late SI activity. (Right) Locations of source generators overlaid
on MRI scans. Magnetic fields were recorded from two probes that were
centered on the C4 (hemisphere contralateral to the stimulation) and C3
(hemisphere ipsilateral to the stimulation) positions based on the Inter-
national 10–20 system.
Adapted from Inui et al. (2003a).
324
1M 2M 1M 2M
Con.
Con. SI
Ipsi.
Con. SII-insula
Ipsi. SII-insula
Con. SII-insula 10 nAm (SI, SII-insula)
Ipsi.SII-insula
Cingulate
Cingulate
Con. MT
Con. MT 30 nAm (cingulate, MT)
Ipsi. MT Ipsi. MT
Figure 12–17. Source generators analyzed by BESA following C fiber stimulation during Control and Distraction in one subject.
SI and SII-insula seem to be the main generators for 1M, while MT and cingulate cortex seem to be the main generators for 2M.
During Distraction, the activities of the SI, SII-insula, MT and cingulate cortex were much reduced. Note the difference in cur-
rent strength scale—10 nA for SI and SII-insula and 30 nA for MT and Cingulate. These source locations overlapped on MRI.
Adapted from Qiu et al. (2004).
326 MEG: An Introduction to Methods
stimulus intensity or frequency is too high. For example, we have noticed that
it is safe to use on average 52mJ/mm2 intensity to a skin area of 10 mm2 for
Aδ-stimulation (total energy 500 mJ), whereas for C-stimuli, in which the
area is restricted to 0.2-0.3 mm2, suitable energy is about 190 mJ/mm2 (total
energy 50 mJ). To avoid skin burns, the stimulus site can be slightly moved
after each stimulus to a random direction within a limited area, for example
10 cm2. In addition, the eyes of the subject and of the researcher handling the
stimulator need to be protected with goggles to avoid possible damage if the
laser beam is accidentally reflected toward the eye.
Motor Function
When motor function is examined by MEG, movement related cortical fields
(MRCFs) or background activities (brain rhythm) related to movement are
recorded.
MRCFs
MEF-I MEF-III
MF MEF-II RR
RF
Contralateral (left)
hemisphere
200fT
Ipsilateral (right)
hemisphere
10 mV
EMG
EOG (horizontal)
EOG (vertical)
for Readiness Potential cancel each other out, is inconvenient for MRCF.
However, Erdler et al. (2000) succeeded in detecting activity in the SMA using
an elegant analytical method.
We investigated MRCFs to identify the motor and sensory brain activities
at the instant of a unilateral finger movement (Hoshiyama et al., 1997). We
focused on the source of the events tightly linked to movement onset, and
used BESA to model the sources generating MRCFs during the interval from
200 ms before to 150 ms after the movement onset (Figures 12–19 and 12–20).
Four sources provided satisfactory solutions for MRCF activities in this inter-
val (Figures 12–19 and 12–20). Sources 1 and 2—which were located in the
pre-central regions in the hemisphere contralateral and ipsilateral to the
moved finger, respectively—generated readiness fields (RF), but Source 1 was
predominant just before movement onset. The motor field (MF), the peak of
which was just after movement onset, was mainly generated by Source 1.
Sources 3 and 4 were located in the post-central regions in the hemisphere
contralateral and ipsilateral to the moved finger, respectively. The first motor
328 MEG: An Introduction to Methods
evoked field (MEF-I), the peak of which was about 80 ms after the movement,
was mainly generated by Source 3, but with the participation of Sources 1, 2
and 4. The results indicated that the activities of both pre- and post-central
regions in both hemispheres were related to voluntary movements, although
the predominant areas varied over time.
We also investigated the vocalization-related cortical fields (VRCFs) fol-
lowing the vocalization of vowels (Gunji et al., 2000, 2001). A multiple-source
model, BESA, was used to elucidate the mechanism generating VRCF in
the period from 150 ms before to 150 ms after the onset of vocalization
(Figure 12–21). Six sources provided satisfactory solutions for VRCF activi-
ties during that period (Figure 12–21). Sources 1 and 2, which were activated
Source 1 10nAm
Source 2
Source 3
Source 4
GOF
GFP
99%
90%
10%
−200 0 +150 msec −200 0 +150 msec −200 0 +150 msec
Source 1 10nAm
Source 2
Source 3
Source 4
GOF
GFP
99%
90%
10%
−200 0 +150 msec −200 0 +150 msec −200 0 +150 msec
Subject 1 2 2 2
1 1 1
4 3 4 3 4
3
2 2 2
L R 34 1 L R 1 L R 1
34 34
1 2 1 2 1 2
3 4 3 4 3 4
2
Subject 4 3 1 2 31 2
4
3 1
4
4
2 32 32
L R 3 41 L R 4 1 L R 41
Figure 12–20. Head diagram indicating the locations of the dipole sources
of two subjects. The line and its length from each point indicate the direc-
tion and magnitude of the dipole current, respectively.
Adapted from Hoshiyama et al. (1997).
Subject 1 (GOF=91.4%)
Source 1
Source 2 6
5 24
3 1
Source 3
5 6
1 2
Source 4
3 4
Source 5 L R
5
6
1
Source 6 2
3
v.o. 4
20nAm
250ms R
Figure 12–21. The MEG waveforms and the results of the source analysis
of Subject 1. Upper right: the VRCF following the vocalization at the left
and right hemispheres (around C3 and C4) in Subject 1. The waveforms
recorded from 37 channels were superimposed. Left: temporal activity of
each source obtained by spatiotemporal source analysis (BESA). Lower
right: the localization and orientation of the dipole on the spherical head
model. The line from each point indicates the direction of the dipole cur-
rent. Sources 1 and 2 were located in the laryngeal motor areas. Sources
3 and 4 were located in the auditory areas, and sources 5 and 6 were in
the truncal motor areas. The sources in the motor areas (sources 1, 2, 5,
6) were activated approximately 100 ms prior to the vocalization onset,
while the activity of the auditory sources (sources 3 and 4) appeared after
the vocalization onset. All six sources temporally overlapped after the
onset of vocalization.
Adapted from Gunji et al. (2000).
see Hari & Salmelin, 1997). Reactivity of the cortical mu rhythm to external
stimuli can be studied with TSE-method (Temporal Spectral Evolution, see
page.x). For example, electric median nerve stimuli result in an initial decrease
of the mu rhythm level, followed by strong rebound within 1000 ms after
stimuli. The rebound is suppressed during finger movements (Salenius et al.
1997), motor imagery (Schnitzler et al. 1997), and even by viewing another
Somatosensory and Motor Function 331
person making movements (Hari et al. 1998). Previous MEG and TMS stud-
ies have suggested that increase of the 20-Hz rhythm after stimuli (“rebound”)
may reflect decreased excitability of the motor cortex (Salmelin and Hari,
1994; Chen et al. 1999). Therefore, reactivity of the 20 Hz rhythm has been
used as a tool to study the functional state of the motor cortex. For example,
analysis of the reactivity of the 20-Hz rhythm in patients with progressive
myoclonus epilepsy revealed abnormal excitation of the motor cortex (Silen
et al., 2000).
Many chronic pain patients show signs of motor dysfunction, such as decreased
muscle strength and restriction of the active range of movement. Further, a
motor cortex stimulator has been used to alleviate chronic pain, suggesting
that the pain and the motor systems are functionally coupled.
A recent study showed with 10 healthy subjects that acute pain modulates
the functions of the motor cortex (Raij et al. 2004); laser stimuli delivered to
the dorsum of the hand elicited long-lasting attenuation of the motor cortex
rhythm, indicating a prolonged activation of the motor cortex in association
with acute pain. In line with these findings, significantly altered reactivity of
the motor cortex has been shown in patients suffering from chronic pain: in
CRPS patients, rebound of the mu rhythm was diminished, suggesting that
inhibition of the motor cortex may be defective in chronic pain (Juottonen
et al., 2002). This view is in line with recent transcranial magnetic stimulation
studies showing signs of disinhibition or hyperexcitability of the motor cortex
in CRPS patients (Schwenkreis et al., 2003), and also agrees with the clinically
frequently observed deficits of motor functions in chronic pain patients. The
close interaction between the pain and motor systems may explain beneficial
effects of mirror therapy and motor imagery in rehabilitation of CRPS
(McCabe et al.., 2003; Moseley, 2004).
References
Akhtari, M., Mcnay, D., Mandelkern, M., Teeter, B., Cline, H. E., Mallick, J., et al.
(1994). Somatosensory evoked response source localization using actual cortical
surface as the spatial constraint. Brain Topogr, 7, 63–69.
Allison, T., McCarthy, G., Wood, C. C., Williamson, P. D., & Spencer, D. D.
(1989). Human cortical potentials evoked by stimulation of the median nerve.
1. Cytoarchitectonic areas generating short-latency activity. J Neurophysiol, 62,
694–710.
Arendt-Nielsen, L., Yamasaki, H., Nielsen, J., Naka, D., & Kakigi, R. (1999). Magne-
toencephalographic responses to painful impact stimulation. Brain Res, 839,
203–208.
332 MEG: An Introduction to Methods
Baumgartner, C., Sutherling, W. W., Di, S., & Barth, D. S. (1991). Spatiotemporal
modeling of cerebral evoked magnetic fields to median nerve stimulation. Electro-
encephalogr Clin Neurophysiol, 79, 27–35.
Bernard, J. F., & Besson, J. M. (1990). The spino (trigemino) pontoamygdaloid
pathway: electrophysiological evidence for an involvement in pain processes.
J Neurophysiol, 63, 473–490.
Biermann, K., Schmitz, F., Witte, O. W., Konczak, J., & Schnitzler, A. (1998). Interac-
tion of finger representation in the human first somatosensory cortex: a neuroma-
gnetic study. Neurosci Lett, 251, 13–16.
Bornhövd, K., Quante, M., Glauche, V., Bromm, B., Weiller, C, & Büchel, C. (2002).
Painful stimuli evoke different stimulus-response functions in the amygdala, pre-
frontal, insula and somatosensory cortex: a single-trial fMRI study. Brain, 125,
1326–1336.
Bragard, D., ACN Chen, & Plaghki, L. (1996). Direct isolation of ultra-late (C-fibre)
evoked brain potentials by CO2 laser stimulation of tiny cutaneous surface areas in
man. Neurosci Lett, 209, 81–84.
Brenner, D., Lipton, J., Kaufman, L., & Williamson, S. J. (1978). Somatically evoked
fields of the human brain. Science, 189, 81–83.
Bromm B., Lorenz J., & Scharein E. (1996). Dipole source analysis of brain activity
in the assessment of pain. In: J. Kimura and H. Shibasaki, eds. Recent Advances in
Clinical Neurophysiology. Amsterdam: Elsevier, pp. 328–335.
Büchel, C., Morris, J., Dolan, R. J., & Friston, K. J. (1998). Brain systems mediating
aversive conditioning: an event-related fMRI study. Neuron, 20, 947–957.
Buchner, H., Fuchs, M., Wischmann, H-A., Dossel, O., Ludwig, I., Knepper, A.,
et al. (1994). Source analysis of median nerve and finger stimulated somatosensory
evoked potentials: Multichannel simultaneous recording of electric and magnetic
fields combined with 3D-MR tomography. Brain Topogr, 6, 299–310.
Buchthal, F., & Rosenfalck, A. (1966). Evoked action potentials and conduction
velocity in human sensory nerves. Brain Res, 3, 1–122.
Chen R., Corwell B., & Hallett M. (1999) Modulation of motor cortex excitability by
median nerve and digit stimulation. Exp Brain Res. 129(1):77–86.
Cheyne, D., & Weinberg, H. (1989). Neuromagnetic fields accompanying unilateral
finger movements: pre-movement and movement-evoked fields. Exp Brain Res, 78,
604–612.
Coghill, R. C., Talbot, J. D., Evans, A. C., Meyer, E., Gjedde, A., Bushnell, M. C.,
et al. (1994). Distributed processing of pain and vibration by the human brain.
J Neurosci, 14, 4095–4108.
Curio, G., Mackert, B-M., Burghoff, M., Koetitz, R., Abraham-Fuchs, K., & Harer, W.
(1994). Localization of evoked neuromagnetic 600 Hz activity in the cerebral
somatosensory system. Electroencephalogr Clin Neurophysiol, 91, 483–487.
Davis, K. D., Kwan, C. L., Crawley, A. P., & Miklis D. J. (1998). Functional, study
of thalamic and cortical activations evoked by cutaneous heat, cold, and tactile
stimuli. J Neurophysiol, 80, 1533–1546.
Elbert, T., Flor, H., Birbaumer, N., Knecht, S., Hampson, S., Larbig, W., et al. (1994).
Extensive reorganization of the somatosensory cortex in adult humans after
nervous system injury. Neuroreport, 5, 2593–2597.
Elbert, T., Junghofer, M., Scholz, B., & Schneider, S. (1995a). The separation of over-
lapping neuromagnetic sources in first and second somatosensory cortices. Brain
Topogr, 7, 275–282.
Somatosensory and Motor Function 333
Elbert, T., Pantev, C., Wienbruch, C., Rockstroh, B., & Taub, E. (1995b). Increased cortical
representation of the fingers of the left hand in string players. Science, 270, 305–307.
Elbert, T., Sterr A., Flor H., Rockstroh B., Knecht S., Pantev C., et al. (1997). Input-
increase and input-decrease types of cortical reorganization after upper extremity
amputation in humans. Exp Brain Res, 117, 161–164.
Endo, H., Kato, Y., Kizuka, T., Masuda, T., & Takeda, T. (2004). Bilateral cerebral
activity for unilateral foot movement revealed by whole-head magnetoencephalo-
graphy. Somatosens Motor Res, 21, 33–43.
Erdler M., Beisteiner R., Mayer D., Kaindl T., Edward, V., Windischberger, C., et al.
(2000). Supplementary motor area activation preceding voluntary movement is
detectable with a whole-scalp magnetoencephalography system. Neuroimage, 11,
697–707.
Erne, S. N., Curio, G., Trahms, L., Trontelj, Z., & Aust, P. (1988). Magnetic activity
of a single peripheral nerve in man. In: K. Atsumi, M. Kotani, S. Ueno, T. Katila &
S.J. Williamson, eds. Biomagnetism. Tokyo: Denki University Press, pp. 166–169.
Flor, H., Elbert, T., Knecht, S., Wienbruch, C., Pantev, C., Birbaumer, N., et al. (1995).
Phantom-limb pain as a perceptual correlate of cortical reorganization following
arm amputation. Nature, 375, 482–484.
Foerster, O. (1936). The motor cortex in man in the light of Hughlings Jackson’s
doctrines. Brain, 56, 135–159.
Forss, N., Hari, R., Salmelin, R., Ahonen, A., Hamalainen, M., Kajola, M., et al.
(1994a). Activation of the human posterior parietal cortex by median nerve
stimulation. Exp Brain Res, 99, 309–315.
Forss, N., Salmelin, R., & Hari R. (1994b). Comparison of somatosensory evoked fields
to airpuff and electric stimuli. Electroencephalogr Clin Neurophysiol, 92, 510–517.
Forss, N., Jousmaki, V., & Hari, R. (1995). Interaction between afferent input from
fingers in human somatosensory cortex. Brain Res, 685, 68–76.
Forss, N., & Jousmaki, V. (1998). Sensorimotor integration in human primary and
secondary somatosensory cortices. Brain Res, 781, 259–267.
Gallen, C. C., Sobel, D. F., Waltz, T., Aung, M., Copeland, B., Schwartz, B. J., et al.
(1993). Noninvasive presurgical neuromagnetic mapping of somatosensory cortex.
Neurosurgery, 33, 260–268.
Gallen, C. C., Schwartz, B., Rieke, K., Pantev, C., Sobel, D., Hirschkoff, E., et al. (1994).
Intrasubject reliability and validity of somatosensory source localization using a large
array biomagnetometer. Electroencephalogr Clin Neurophysiol, 90, 145–156.
Gallien, P., Aghulon, C., Durufle, A., Petrilli, S., de Crouy, A. C., Carsin, M., et al.
(2003). Magnetoencephalography in stroke: a 1-year follow-up study. Eur J Neurol,
10, 373–382.
Garcia-Larrea, L., Frot, L. M., & Valeriani, M. (2003). Brain generators of laser-evoked
potentials: from dipoles to functional significance. Neurophysiol Clin, 33, 279–292.
Gunji A., Hoshiyama M., & Kakigi R. (2001). Auditory response following vocaliza-
tion; magnetoencephalographic study. Clin Neurophysiol, 111, 214–219.
Gunji, A., Kakigi, R., & Hoshiyama, M. (2000). Spatiotemporal source analysis of
vocalization-associated magnetic fields. Brain Res Cogn Brain Res, 9, 157–163.
Grimm, Ch., Schreiber, A., Kristeva-Feige, R., Mergner, Th., Henning, J., & Lucking,
C. H. (1998). A comparison between electric source localisation and fMRI during
somatosensory stimulation. Electroencephalogr Clin Neurophysiol, 106, 22–29.
Hamalainen, M. S. (1992). Magnetoencephalography: A tool for functional brain
mapping. Brain Topogr, 5, 95–102.
334 MEG: An Introduction to Methods
Hari, R., Kaukoranta, E., Reinikainen, K., Huopaniemie, T., & Mauno, J. (1983).
Neuromagnetic localization of cortical activity evoked by painful dental stimula-
tion in man. Neurosci Lett, 42, 77–82.
Hari, R., Reinikainen, K., Kaukoranta, E., Hamalainen, M., Ilmoniemi, R., Penttinen, A.,
et al. (1984). Somatosensory evoked cerebral magnetic fields from S1 and S2 in
man. Electroencephalogr Clin Neurophysiol, 57, 254–263.
Hari, R., & Kaukoranta, E. (1985). Neuromagneic studies of somatosensory system:
principles and example. Prog Neurobiol, 24, 233–256.
Hari, R., Hallstrom, J., Tiihonen, J., & Joutsiniemi, S-L. (1989). Multichannel detec-
tion of magnetic compound action fields of median and ulnar nerve. Electroence-
phalogr Clin Neurophysiol, 72, 277–280.
Hari, R., Hamalainen, H., Hamalainen, M., Kekoni, J., Sams, M., & Tiihonen, J.
(1990). Separate finger representations at the human second somatosensory cortex.
Neuroscience, 37, 245–249.
Hari, R. (1991). On brain’s magnetic responses to sensory stimuli. Journal of Clin
Neurophysiol, 8, 157–169.
Hari, R., Karhu., J, Hamalainen, M., Knuutila, J., Salonen, O., Sams, M., et al. (1993).
Functional organization of the human first and second somatosensory cortices: a
neuromagneic study. Eur J Neurosci, 5, 724–734.
Hari, R., Nagamine, T., Nishitani, N., Mikuni, N., Sato, T., Tarkiainen, A., et al.
(1996). Time-varying activation of different cytoarchitectonic areas of the human
SI cortex after tibial nerve stimulation. Neuroimage, 4, 111–118.
Hari, R., Portin, K., Kettenmann, B., Jousmaki, V., & Kobal, G. (1997). Right-
hemisphere preponderance of responses to painful CO2 stimulation of the human
nasal mucosa. Pain, 72, 145–151.
Hari R., & Salmelin R. (1997) i. Human cortical oscillations: a neuromagnetic view
through the skull. Trends Neurosc 20(1):44–49. Review.
Hari R., Forss N., Avikainen S., Kirveskari E., Salenius S., & Rizzolatti G. (1998).
Activation of human primary motor cortex during action observation: a neuro-
magnetic study. Proc Natl Acad Sci U S A Dec 8; 95(25), 15061–15065.
Hashimoto, I. (1988). Trigeminal evoked potentials following brief air puff: enhanced
signal-to-noise ratio. Ann Neurol, 23, 332–338.
Hashimoto, I., Okada, K., Gatayama, T., & Yokoyama, S. (1991). Multichannel
measurements of magnetic compound action fields of the median nerve in man.
Electroencephalogr Clin Neurophysiol, 81, 332–336.
Hashimoto, I., Mashiko, T., Mizuta, T., Imada, T., Iwase, K., & Okazaki, H. (1994).
Visualization of a moving quadrupole with magnetic mesurements of peripheral
nerve action fields. Electroencephalogr Clin Neurophysiol, 93, 459–467.
Hashimoto, I., Mashiko, T., & Imada, T. (1996). Somatic evoked high-frequency
magnetic oscillations reflect activity of inhibitory interneurons in the human soma-
tosensory cortex. Electroencephalogr Clin Neurophysiol, 100, 189–203.
Hoshiyama, M., Kakigi, R., Koyama, S., Kitamura, Y., Shimojo, M., & Watanabe, S.
(1995). Somatosensory evoked magnetic fields after mechanical stimulation of the
scalp in humans. Neurosci Lett, 195, 29–32.
Hoshiyama, M., Kakigi, R., Koyama, S., Kitamura, Y., Shimojo, M., & Watanabe, S.
(1996). Somatosensory evoked magnetic fields following stimulation of the lip in
humans. Electroencephalogr Clin Neurophysiol, 100, 96–104.
Hoshiyama, M., Koyama, S., Kitamura, Y., Watanabe, S., Shimojo, M., & Kakigi, R.
(1997). Activity in pariental cortex following somatosensory stimulation in man:
Somatosensory and Motor Function 335
Inui, K., Wang, X., Tamura, Y., Kaneoke, Y., & Kakigi, R. (2004). Serial processing in
the human somatosensory system. Cereb Cortex, 14, 851–857.
Inui, K, Tsuji, T, & Kakigi, R. (2006). Temporal analysis of cortical mechanisms for
pain relief by tactile stimuli in humans. Cereb Cortex, 33, 355-365.
Itomi, K., Kakigi, R., Maeda, K., & Hoshiyama, M. (2000a). Dermatome versus
homunculus: Detailed topography of the primary somatosensory cortex following
trunk stimulation. Clin Neurophysiol, 111, 405–412.
Itomi, K., Kakigi, R., Mukai, T., & Hoshiyama, M. (2000b). Magnetic responses to
shoulder and posterior head stimulation in humans. Brain Topogr, 14, 15–23.
Iwamura, Y., Iriki, A., & Tanaka, M. (1994). Bilateral hand representation in the post-
central cortex. Nature, 369, 554–556.
Jones, E. G., & Powell, T. P. S. (1970). Anatomical study of converging sensory
pathways within the cerebral cortex of the monkey. Brain, 93, 793–820.
Jones, S. J., Halonen, J-P., & Shawkat, F. (1989). Centrifugal and centripetal mecha-
nisms involved in the “gating” of cortical SEPs during movement. Electroencepha-
logr Clin Neurophysiol, 74, 36–45.
Jousmaki, V. & Hari, R. (1999). Somatosensory evoked fields to large-area vibrotactile
stimuli. Clin Neurophysiol, 110, 905–909.
Juottonen K., Gockel M., Silén T., Hurri H., Hari R., & Forss N. (2002) Altered central
sensorimotor processing in patients with complex regional pain syndrome. Pain
98(3):315–23.
Kakigi, R., & Shibasaki, H. (1984). Scalp topography of mechanically and electrically
evoked somatosensory potentials in man. Electroencephalogr Clin Neurophysiol, 59,
44–56.
Kakigi, R., & Jones, S. J. (1985). Effects on median nerve SEPs of tactile stimulation
applied to adjacent and remote areas of the body surface. Electroencephalogr Clin
Neurophysiol, 62, 252–265.
Kakigi, R. (1986). Ipsilateral and contralateral SEP components following median
nerve stimulation: Effects of interfering stimuli applied to the contralateral hand.
Electroencephalogr Clin Neurophysiol, 64, 246–259.
Kakigi, R., & Jones, S. J. (1986). Influence of concurrent tactile stimulation on soma-
tosensory evoked potentials following posterior tibial nerve stimulation in man.
Electroencephalogr Clin Neurophysiol, 65, 118–129.
Kakigi, R., & Shibasaki, H. (1987). Generator mechanisms of giant somatosensory
evoked potentials in cortical reflex myoclonus. Brain, 110, 1359–1373.
Kakigi, R., & Shibasaki, H. (1991). Effects of age, gender and stimulus side on the scalp
topography of somatosensory evoked potentials following median nerve stimula-
tion. J Clin Neurophysiol, 8, 320–330.
Kakigi, R., & Shibasaki, H. (1992). Effects of age, gender and stimulus side on the scalp
topography of somatosensory evoked potentials following posterior tibial nerve
stimulation in man. J Clin Neurophysiol, 9, 431–440.
Kakigi, R. (1994). Somatosensory evoked magnetic fields following median nerve
stimulation. Neurosci Res, 20, 165–174.
Kakigi, R., Koyama, S., Hoshiyama, M., Shimojo, M., Kitamura, Y., & Watanabe, S.
(1995a). Topography of somatosensory evoked magnetic fields following posterior
tibial nerve stimulation. Electroencephalogr Clin Neurophysiol, 95, 127–134.
Kakigi, R., Koyama, S., Hoshiyama, M., Kitamura, Y., Shimojo, M., & Watanabe, S.
(1995b). Pain-related magnetic fields following painful CO2 laser stimulation in
man. Neurosci Lett, 192, 45–48.
Somatosensory and Motor Function 337
Kakigi, R., Koyama, S., Hoshiyama, M., Kitamura, Y., Shimojo, M., Watanabe, S.,
et al. (1996a). Effects of tactile interference stimulation on somatosensory evoked
magnetic fields. Neuroreport, 7, 405–408.
Kakigi, R., Koyama, S., Hoshiyama, M., Kitamura, Y., Shimojo, M., & Watanabe S.
(1996) Pain-related brain responses following CO2 laser stimulation: Magnetoen-
cephalographic studies. In: I. Hashimoto, Y.C. Okada & S. Ogawa, eds. Visualization
of information processing in the human brain. Electroencephalogr Clin Neurophy-
siol (Suppl) 47, 110–120.
Kakigi, R., Shimojo, M., Hoshiyama, M., Koyama, S., Watanabe, S., Naka, D., et al.
(1997). Effects of movement and movement imagery on somatosensory evoked
magnetic fields following posterior tibial nerve stimulation. Brain Res Cogn Brain
Res, 5, 241–253.
Kakigi, R., Tran, T. D., Qiu, Y., Wang, X., Nguyen, T. B., Inui, K., et al. (2003).
Cerebral responses following stimulation of unmyelinated C-fibers in humans:
electro- and magneto-encephalographic study. Neurosci Res, 45, 255–275.
Kakigi, R., Inui, K., & Tamura, Y. (2005). Electrophysiological studies on human pain
perception. Clin Neurophysiol, 116, 743–763.
Kamada, K., Takeuchi, F., Kuriki, S., Oshiro, O., Houkin, K., & Abe, H. (1993). Func-
tional Neurosurgical simulation with brain surface magnetic resonance images and
magnetoencephalography. Neurosurgery, 33, 269–273.
Kanda, M., Nagamine, T., Ikeda, A., Ohara, S., Kunieda, T., Fujiwara, N., et al. (2000).
Primary somatosensory cortex is actively involved in pain processing in human. Brain
Res, 853, 282–289.
Kanno, A., Nakasato, N., Hatanaka, K., & Yoshimoto, T. (2003). Ipsilateral area 3b
responses to median nerve somatosensory stimulation. Neuroimage, 18, 169–177.
Karhu, J., Hari, R., Lu, S. T., Paetau, R., & Rif, J. (1991). Cerebral magnetic fields to
lingual stimulation. Electroencephalogr Clin Neurophysiol, 80, 459–468.
Karhu, J., Hari, R., Makela, J. P., Huttunen, J., & Knuutila, J. (1992). Cortical soma-
tosensory magnetic responses in multiple sclerosis. Electroencephalogr Clin Neuro-
physiol, 83, 192–200.
Karhu, J, Hari, R, Paetau, R, Kajala, M, & Mervaala, E. (1994). Cortical reactivity
in progressive myoclonus epilepsy. Electroencephalogr Clin Neurophysiol, 90,
93–102.
Kaufman, L., Okada, Y., Brenner, D., & Williamson, S. J. (1981). On the relation
between somatic evoked potentials and fields. Int J Neurosci, 15, 223–239.
Kaukolanta, E., Hari R., Hamalainen, M., & Huttunen, J. (1986). Cerebral magnetic
fields evoked by peroneal nerve stimulation. Somatosens Res, 3, 309–321.
Kawamura, T., Nakasato, N., Seki, K., Kanno, A., Fujita, S., Fijiwara, S., et al. (1996).
Neuromagnetic evidence of pre- and post-central cortical sources of somatosensory
evoked responses. Electroencephalogr Clin Neurophysiol, 100, 44–50.
Kenshalo Jr., D. R., & Willis Jr., W. D. (1991). The role of cerebral cortex in pain per-
ception. In: A. Peters and E. G. Jones, eds. The Cerebral Cortex. New York: Plenum,
pp. 153–212.
Kitamura, Y., Kakigi, R., Hoshiyama, M., Koyama, S., Shimojo, M., & Watanabe, S.
(1995). Pain-related somatosensory evoked magnetic fields. Electroencephalogr Clin
Neurophysiol, 95, 463–474.
Kitamura, Y., Kakigi, R., Hoshiyama, M., Koyama, S., & Nakamura, A. (1996). Effects
of sleep on somatosensory evoked responses in human: A magnetoencephalogra-
phic study. Brain Res Cogn Brain Res, 4, 275–279.
338 MEG: An Introduction to Methods
Kitamura, Y., Kakigi, R., Hoshiyama, M., Koyama, S., Shimojo, M., & Watanabe, S.
(1997). Pain-related somatosensory evoked magnetic fields following lower limb
stimulation. J Neurol Sci, 145, 187–194.
Knecht, S., Henningsen, H., Elbert, T., Flor, H., Hohling, C., Pantev, C., et al. (1995).
Cortical reorganization in human amputees and mislocalization of painful stimuli
to the phantom limb. Neurosci Lett, 201, 262–264.
Knecht, S., Henningsen, H., Elbert, T., Flor, H, Hohling, C., Pantev, C., et al.
(1996). Reorganizational and perceptional changes after amputation. Brain, 119,
1213–1219.
Knecht, S., Henningsen, H., Hohling, C., Elbert, T., Flor, H., Pantev, C., et al. (1998).
Plasticity of plasticity? Changes in the pattern of perceptual correlates of reorgani-
zation after amputation. Brain, 121, 717–724.
Kristeva, R., Cheyne, D., & Deecke, L. (1991). Neuromagnetic fields accompanying
unilateral and bilateral voluntary movements: topography and analysis of cortical
sources. Electroencephalogr Clin Neurophysiol, 81, 284–298.
Kristeva-Feige, R., Walter, H., Lutkenhoner, B., Hampson, S., Ross, B., Knorr, U., et al.
(1994). A neuromagnetic study of the functional organization of the sensorimotor
cortex. Eur J Neurosci, 6, 632–639.
Kruger, L., Sampogna, S. L., Rodin, B. E., Clague, J., Brecha, N., & Yeh, Y. (1985).
Thin-fiber cutaneous innervation and its intraepidermal contribution studies by
labeling methods and neurotoxin treatment in rats. Somatosens Res, 2, 335–356.
Lam, K., Kakigi, R., Kaneoke, Y,. Naka, D., Maeda, K., & Suzuki, H. (1999). Effects
of visual and auditory stimulation on somatosensory evoked magnetic fields. Clin
Neurophysiol, 110, 295–304.
Lenz, F. A, Rios, M., Zirh, A., Chau, D., Krauss, G., & Lesser, R, P. (1998). Painful
stimuli evoke potentials recorded over the human anterior cingulate gyrus. J Neu-
rophysiol, 79, 2231–2234.
Loose, R., Hamdy, S., & Enck, P. (2001). Magnetoencephalographic response charac-
teristics associated with tongue movement. Dysphagia, 16, 183–185.
Maclin, E., Rose, D. F., Knight, J. E., Orrison, W. W., & Davis, L. E. (1994). Soma-
tosensory evoked magnetic fields in patients with stroke. Electroencephalogr Clin
Neurophysiol, 91, 468–475.
Maeda, K., Kakigi, R., Hoshiyama, M., & Koyama, S. (1999). Topography of the
secondary somatosensory cortex in humans: a magnetoencephalographic study.
Neuroreport, 10, 301–306.
McCabe C.S., Haigh R.C., Ring E.F., Halligan P.W., Wall P.D., & Blake D.R. (2003).
A controlled pilot study of the utility of mirror visual feedback in the treatment of
complex regional pain syndrome (type 1). Rheumatology (Oxford). 42(1), 97–101.
Makela, J. P. & Hari, R. (1992). Neuromagnetic auditory evoked responses after a
stroke in the right temporal lobe. Neuroreport, 3, 94–96.
Mauguiere, F., Merlet, I., Forss, N., Vanni, S., Jousmaki, V., Adeleine, P., et al. (1997a).
Activation of a distributed somatosensory cortical network in the human brain:
a dipole modeling study of magnetic fields evoked by median nerve stimulation.
Part 1: Location and activation timing of SEF sources. Electroencephalogr Clin
Neurophysiol, 104, 281–289.
Mauguiere, F., Merlet, I., Forss, N., Vanni, S., Jousmaki, V., Adeleine, P., et al. (1997b).
Activation of a distributed somatosensory cortical network in the human brain:
a dipole modeling study of magnetic fields evoked by median nerve stimulation.
Part 2: Effects of stimulus rate, attention and stimulus detection. Electroencephalogr
Clin Neurophysiol, 104, 290–295.
Somatosensory and Motor Function 339
Mayville, J. M., Fuchs, A., & Kelso, J. A. (2005). Neuromagnetic motor fields accom-
panying self-paced rhythmic finger movement at different rates. Exp Brain Res, 166,
190–199.
Mima, T., Ikeda, A., Nagamine, T., Yazawa, S., Kunieda, T., Mikuni, N., et al. (1997).
Human second somatosensory area: Subdural and magnetoencephalographic
recording of somatosensory evoked responses. J Neurol Neurosurg Psychiatry, 63,
501–505.
Mima, T., Nagamine, T., Nakamura, K., & Shibasaki, H. (1998a). Attention modulates
both pimary and second somatosensory cortical activities in humans: A magneto-
encephalographic study. J Neurophysiol, 80, 2215–2221.
Mima, T., Nagamine, T., Nishitani, N., Mikuni, N., Ikeda, A., Fukuyama, H.,
et al. (1998b). Cortical myoclonus. sensorimotor hyperexcitability. Neurology, 50,
933–942.
Mogilner, A., Grossman, J. A. I., Ribary, U., Joliot, M., Volkmann, J., Rapaport, D.,
et al. (1993). Somatosensory cortical plasticity in adult humans revealed by magne-
toencephalography. Proc Natl Acad Sci USA, 90, 3593–3597.
Mogilner, A., Nomura, M., Ribary, U., Jagow, R., Lado, F., Rusinek, H., et al. (1994).
Neuromagnetic studies of the lip area of primary somatosensory cortex in humans:
evidence for an oscillotopic organization. Exp Brain Res, 99, 137–147.
Moseley G.L. (2004). Graded motor imagery is effective for long-standing complex
regional pain syndrome: a randomised controlled trial. Pain 108(1-2), 192–198.
Munger, B. L., & Halata, Z. (1983). The sensory innervation of primate facial skin. I.
Hairy skin. Brain Res Rev, 5, 45–80.
Nagamine, T., Toro, T., Balish, M., Deuschl, G., Wang, B., Sato, S., et al. (1994).
Cortical magnetic and electric fields associated with voluntary finger movements.
Brain Topogr, 6, 175–183.
Nagamine, T., Makela, J., Mima, T., Mikuni, N., Nishitani, N., Satoh, T., et al. (1998).
Serial processing of the somesthetic information revealed by different effects of sti-
mulus rate on the somatosensory-evoked potentials and magnetic fields. Brain Res,
791, 200–208.
Naka, D., Kakigi, R., Koyama, S., Xiang, J., & Suzuki, H. (1998). Effects of tactile inter-
ference stimulation on somatosensory evoked magnetic fields following tibial nerve
stimulation. Electroencephalogr Clin Neurophysiol, 109, 168–177.
Nakagawa, H., Namima, T., Aizawa, M., Uchi, K., Kaiho, Y., Yoshikawa, K., et al.
(1998). Somatosensory evoked magnetic fields elicited by dorsal penile, posterior
tibial and median nerve stimulation. Electroencephalogr Clin Neurophysiol, 108,
57–61.
Nakamura, A., Yamada, T., Goto, A., Kato, T., Ito, K., Abe, Y., et al. (1998). Somato-
sensory homunculus as drawn by MEG. Neuroimage, 7, 377–386.
Nakasato, N., Seki, K., Kawamura, T., Ohtomo, S., Kanno, A., Fujita, S., et al. (1996).
Cortical mapping using an MRI-linked whole head MEG system and presurgical
decision making. In: Hashimoto, Y. C. Okada & S. Ogawa, eds. Visualization of
Information Processing in the Human Brain. Electroencephalogr Clin Neurophysiol
Suppl 47, 333–341.
Nakasato, N., Kanno, A., Hatanaka, K., Ohtomo, S., Inoue, T., Shimizu, H., et al.
(1999). Preoperative MEG, fMRI and intraoperative cortical stimulation methods
for neurosurgical brain mapping. In: T. Yoshimoto, M. Kotani, S. Kuriki, H. Karibe
& N. Nakasato, eds. Recent Advances in Biomagnetism. Sendai: Tohoku University
Press. pp. 821–824.
340 MEG: An Introduction to Methods
Nakasato, N., Itoh, H., Hatanaka, K., Nakahara, H., Kanno, A., & Yoshimoto, T.
(2001). Movement-related magnetic fields to tongue protrusion. Neuroimage, 14,
924–935.
Nakata, H., Inui, K., Nishihira, Y., Hatta, A., Sakamoto, M., Kida, T., et al. (2004).
Effects of a go/nogo task on event-related potentials following somatosensory sti-
mulation. Clin Neurophysiol, 115, 361–368.
Narich, L., Madonna, I., Opsomer, R. J., Pizzella, V., Romani, G. L., Torrioli, G.,
et al. (1991). Neuromagnetic somatosensory homunculus: a non-invasive approach
in humans. Neurosci Lett, 121, 51–54.
Neshige, R., Luders, H., & Shibasaki, H. (1988). Recording of movement-related
potentials from scalp and cortex in man. Brain, 111, 719–736.
Nguyen, B. T., Tran T. D., Hoshiyama, M., Inui, K., & Kakigi R. (2004). Face represen-
tation in the human primary somatosensory cortex. Neurosci Res, 50, 227–232.
Nguyen, T. B., Inui, K., Hoshiyama, M., & Kakigi, R. (2005). Face representation in
the human secondary somatosensory cortex. Clin Neurophysiol, 116, 1247–1253.
Nihashi, T., Kakigi, R., Kawakami, O., Hoshiyama, M., Itomi, K., Nakanishi, H., et al.
(2001). Representation of the ear in human primary somatosensory cortex. Neuro-
image, 13, 295–304.
Nihashi, T., Kakigi, R., Okada, T., Sadato, N., Kashikura, K., Kajita, Y., et al. (2002).
Functional magnetic resonance imaging evidence for a representation of the ear in
human primary somatosensory cortex: comparison with MEG study. Neuroimage,
17, 1217–1226.
Nihashi, T., Kakigi, R., Hoshiyama, M., Miki K., Kajita, Y., Yoshida, J., et al. (2003). Effect
of tactile interference stimulation of the ear in human primary somatosensory
cortex: a magnetoencephalographic study. Clin Neurophysiol, 114, 1866–1878.
Nihashi, T., Naganawa, S., Sato, C., Kawai, H., Nakamura, T., Fukatsu, H., et al.
(2005). Contralateral and ipsilateral responses in primary somatosensory cortex
following electrical median nerve stimulation-an fMRI study. Clin Neurophysiol,
116, 842–848.
Novotny, G. E. K., & Gommert-Novotny, E. (1988). Intraepidermal nerves in human
digital skin. Cell Tissue Res, 254, 111–117.
Opsommer, E., Masquelier, E., & Plaghki, L. (1999). Study of nerve conduction velo-
city of C-fibers in humans from thermal thresholds to contact heat (thermode)
and from evoked brain potentials to radiant heat (CO2 laser). Neurophysiol Clin,
29, 411–422.
Opsommer, E., Weiss, T., Plaghki, L., & Miltner, W. H. R. (2001a). Dipole analysis
of ultralate (C-fibers) evoked potentials after laser stimulation of tiny cutaneous
surface areas in humans. Neurosci Lett, 298, 41–44.
Opsommer, E., Weiss, T., Miltner, W. H., & Plaghki, L. (2001b). Scalp topography of
ultralate (C-fibres) evoked potentials following thulium YAG laser stimuli to tiny
skin surface areas in humans. Clin Neurophysiol, 112, 1868–1874.
Pantev, C., Gallen, C., Hampson, S., Buchanan, S., & Sobel, D. (1991). Reproducibility
and validity of neuromagnetic source localization using a large array biomagneto-
meter. Am J EEG Tech, 31, 83–101.
Penfield, W., & Boldray, E. (1937). Somatic motor and sensory representation in the
cerebral cortex of man as studied by electrical stimulation. Brain, 60, 389–443.
Peyron, R., Frot, M., Schneider, F., Garcia-Larrea, L., Mertens, P., Barral, F.G., et al.
(2002). Role of operculoinsular cortices in human pain processing: Covering evi-
dence from PET, fMRI, dipole modeling, and intracerebral recordings of evoked
potentials. Neuroimage, 17, 1336–1346.
Somatosensory and Motor Function 341
Ploghaus, A., Tracey, I., Clare, S., Gati, J. S., Rawlins, J. N. P., & Matthews, P. M.
(2000). Learning about pain: The neural substrate of the prediction error for aver-
sive events. Proc Natl Acad Sci U S A, 97, 9281–9286.
Ploghaus, A., Narain, C., Beckmann, C. F., Clare, S., Bantick, S., Wise, R., et al. (2001).
Exacerbation of pain by anxiety is associated with activity in a hippocampal net-
work. J Neurosci, 21, 9896–9903.
Ploner, M., Schmitz, F., Freund, H. J., & Schnitzler, A. (1999). Parallel activation of
primary and secondary somatosensory cortices in human pain processing. J Neu-
rophysiol, 81, 3100–3104.
Ploner, M., Schmitz, F., Freund, H. J., & Schnitzler, A. (2000). Differential organiza-
tion of touch and pain in human primary somatosensory cortex. J Neurophysiol,
83, 1770–1776.
Praamstra, P, Schmitz, F, Freund, H. J., & Schnitzler, A. (1999). Magneto-encephalo-
graphic correlates of the lateralized readiness potential. Brain Res Cogn Brain Res,
8, 77–85.
Qiu, Y., Inui, K., Wang, X., Tran, T. D., & Kakigi, R. (2001). Conduction velocity of
the spinothalamic tract in humans as assessed by CO2 laser stimulation of C-fibers
in men. Neurosci Lett, 311, 181–184.
Qiu, Y., Inui, K., Wang, X., Tran, T. D., & Kakigi, R. (2002). Effects of attention,
distraction and sleep on CO2 laser evoked potentials related to C-fibers in human.
Clin Neurophysiol, 113, 1579–1585.
Qiu, Y., Fu, Q., Wang, X., Tran, T. D., Inui, K., Iwase, S., et al. (2003). Microneuro-
graphic study of C fiber discharges induced by CO2 laser stimulation in humans.
Neurosci Lett, 353, 25–28.
Qiu, Y., Inui, K., Wang, X., Nguyen, B. T., Tran, T. D., & Kakigi, R. (2004). Effects of
distration on MEG responses ascending through C-fibers in humans. Clin Neuro-
physiol, 115, 636–646.
Qiu, Y., Honda, M., Noguchi, Y., Nakata, H., Tamura, Y., Tanaka, S., et al. (2006).
Brain processing of the signals ascending through unmyelinated C fibers in humans,
an event-related fMRI study. Cereb Cortex, 16, 1289–1295.
Pons, T. P., Garraghty, P. E., Ommaya, A. K., Kaas, J. H., Taub, E., & Mishkin, M.
(1991). Massive cortical reorganization after sensory deafferentation in adult maca-
ques. Science, 252, 1857–1860.
Raij TT, Vartiainen NV, Jousmäki V, Hari R. (2003). Effects of interstimulus inter-
val on cortical responses to painful laser stimulation. J Clin Neurophysiol. 20(1),
73–79.
Raij T.T., Forss N., Stancák A., & Hari R. (2004). Modulation of motor-cortex oscilla-
tory activity by painful Adelta- and C-fiber stimuli. Neuroimage 23(2), 569–573.
Robinson, D. L. (1973). Electrophysiological analysis of interhemispheric relations on
the second somatosensory cortex of the cat. Exp Brain Res, 18, 131–144.
Rogers, R. L., Basile, L. F. H, Taylor, S., Sutherling, W. W., & Papanicolaou, A. C.
(1994). Somatosensory evoked fields and potentials following tibial nerve stimula-
tion. Neurology, 44, 1283–1286.
Rossini, P. M., Narici, L., Romani, G. L., Peresson, M., Torrioli, G., & Traversa, R.
(1989). Simultaneos motor output and sensory input: cortical interference site
resolved in humans via neuromagnetic measurements. Neurosci Lett, 96, 300–305.
Rossini, P. M., Narici, L., Martino, G., Pasquarelli, A., Peresson, M., Pizzella, V.,
et al. (1994). Analysis of interhemispheric asymmetries of somatosensory evoked
magnetic fields to right and left median nerve stimulation. Electroencephalogr Clin
Neurophysiol, 91, 476–482.
342 MEG: An Introduction to Methods
Rossini, P. M., Caltagirone, C., Castriota-Scanderbeg, A., Cicinelli, P., Gratta, C. D.,
Demartin, M., et al. (1998a). Hand motor cortical area reorganization in stoke: A
study with fMRI, MEG and TCS maps. Neuroreport, 9, 2141–2146.
Rossini, P. M., Tecchio, F., Pizzella, V., Lupoi, D., Cassetta, E., Pasqualetti, P., et al. (1998b).
On the reorganization of sensory hand areas after mono-hemispheric lesion: a
functional (MEG)/anatomical (MRI) integrative study. Brain Res, 782, 153–166.
Rossini, P. M., Tecchio, F., Pizzella, V., Lupoi, D., Cassetta, E., & Pasqualetti, P. (2001).
Interhemispheric differences of sensory hand areas after monohemispheric stroke:
MEG/MRI integrative study. Neuroimage, 14, 474–485.
Roth, B. J., Sepulveda, N. G., & Wikswo, J. P. (1989). Using a magnetometer to image
a two-dimensional current distribution. J Appl Physiol, 65, 361–372.
Sakamoto, K., Nakata, H., & Kakigi, R. (2008a) Somatotopic representation of the
tongue in human secondary somatosensory cortex. Clin Neurophysiol, 119,
1664–1673.
Sakamoto, K., Nakata, H., & Kakigi, R. (2008b). Somatotopic representation of
the tongue in human secondary somatosensory cortex. Clin Neurophysiol, 119,
2125–2134.
Sakuma, K., & Hashimoto, I. (1999a). High-frequency magnetic oscillations evoked
by posterior tibial nerve stimulation. Neuroreport, 10, 227–230.
Sakuma, K., Sekihara, K., & Hashimoto, I. (1999b). Neural source estimation from a
time-frequency component of somatic evoked high-frequency magnetic oscillati-
ons to posterior tibial nerve stimulation. Clin Neurophysiol, 110, 1585–1588.
Salenius S., Schnitzler A., Salmelin R., Jousmäki V., & Hari R. (1997) Modulation of
the human cortical rolandic rhythms during natural sensorimotor tasks. Neuro-
image 5, 221–228.
Salmelin, R. & Hari R. (1994) Spatiotemporal characteristics of sensorimotor MEG
rhythms related to thumb movement. Neuroscience 60, 537–550.
Sarvas, J. (1987). Basic mathematical and electromagnetic concepts of the biomagne-
tic inverse problem. Phys Med Biol, 32, 11–12.
Sawamoto, N., Honda, M., Okada, T., Hanakawa, T., Kanda, M., Fukuyama, H.,
et al. (2000). Expectation of pain enhances responses to nonpainful somatosen-
sory stimulation in the anterior cingulate cortex and parietal operculum/posterior
insula: an event-related functional magnetic resonance imaging study. J Neurosci,
20, 7438–7445.
Schmidt, R. F., Schaible, H. G., Messlinger, K., Heppelmann, B., Hanesch, U., &
Pawlak, M. (1994). Silent and active nociceptors: structure, functions, and clinical
implications. In: G. F. Gebhart, D. L. Hammond and T. S. Jensen, eds. Proceedings
of the 7th World Congress on Pain, Seattle, WA: IASP Press, pp. 213–250.
Schnitzler, A., Salmelin, R., Salenius, S., Jousmaki, V., & Hari, R. (1995a). Tactile
information from the human hand reaches the ipsilateral primary somatosensory
cortex. Neurosci Lett, 200, 25–28.
Schnitzler, A., Wittem, O. W., Cheyne, D., Haid, D., Vrba, J., & Freund, H. J. (1995b).
Modulation of somatosensory evoked magnetic fields by sensory and motor inter-
ference. Neuroreport, 6, 1653–1658.
Schnitzler A., Salenius S., Salmelin R., Jousmäki V., & Hari R. (1997) Involvement
of primary motor cortex in motor imagery: a neuromagnetic study. Neuroimage.
6(3):201–208.
Schnitzler, A., & Ploner, M. (2000). Neurophysiology and functional neuroanatomy
of pain perception. J Clin Neurophysiol, 17, 592–603.
Somatosensory and Motor Function 343
Seeck, M., Lazeyras, F., Michel, C. M., Blanke, O., Gericke, C. A., Ives, J., et al. (1998).
Non-invasive epileptic focus localization using EEG-triggered functional MRI and
electromagnetic tomography. Electroencephalogr Clin Neurophysiol, 106, 508–512.
Servos, P., Engel, S. A., Gati, J., & Menon, R. (1999). FMRI evidence for an inverted
face representation in human somatosensory cortex. Neuroreport, 10, 1393–1395.
Shibasaki, H., Barrett, G., Halliday, E., & Halliday, A. M. (1980). Components of the
movement-related cortical potential and their scalp topography. Electroencephalogr
Clin Neurophysiol, 49, 213–26.
Shimizu, H., Nakasato, N., Mizoi, K., & Yoshimoto, T. (1997). Localizing the central
sulcus by functional magnetic resonance imaging and magnetoencephalography.
Clin Neurol Neurosurg, 99, 235–238.
Shimojo, M., Kakigi, R., Hoshiyama, M., Koyama, S., Kitamura, Y., & Watanabe, S.
(1996a). Intracerebral interactions caused by bilateral median nerve stimulation in
man. A magnetoencephalographyic study. Neurosci Res, 24, 175–181.
Shimojo, M., Kakigi, R., Hoshiyama, M., Koyama, S., Kitamura, Y., & Watanabe, S.
(1996b), Differentiation of receptive fields in the sensory cortex following stimula-
tion of various nerves of the lower limb in man. Magnetoencephalographic study.
J Neurosurg, 5, 255–262.
Shimojo, M., Kakigi, R., Hoshiyama, M., Koyama, S., & Watanabe, S. (1997). Magne-
toencephalographic study of intracerebral interactions caused by bilateral posterior
tibial nerve stimulation in man. Neurosci Res, 28, 41–47.
Silén T., Forss N., Jensen O., & Hari R. (2000) Abnormal reactivity of the approxi-
mately 20-Hz motor cortex rhythm in Unverricht Lundborg type progressive myo-
clonus epilepsy. Neuroimage 12, 707–712.
Sobel, D. F., Gallen, C. C., Schwartz, B. J., Waltz, T. A., Copeland, B., Yamada, S., et al.
(1993). Locating the central sulcus: Comparison of MR anatomic and magnetoen-
cephalographic functional methods. Am J Neuroradiol, 14, 915–925.
Stancak, A., Raij, T. T., Pohja, M., Forss, N., & Hari, R. (2005). Oscillatory motor
cortex-muscle coupling during painful laser and nonpainful tactile stimulation.
Neuroimage, 26, 793–800.
Sterr, A., Muller, M. M., Elbert, T., Rockstroh, B., Pantev, C., & Taub, E. (1998a).
Perceptual correlates of changes in cortical representation of fingers in blind mul-
tifinger Braille readers. J Neurosci, 18, 4417–4423.
Sterr, A., Muller, M. M., Elbert, T., Rockstroh, B., Pantev, C., & Taub, E. (1998b).
Changed perception in Braille-readers. Nature, 391, 134–135.
Stippich, C., Freitag, P., Kassubek, J., Soros, P., Kamada, K., Kober, H., et al. (1998).
Motor, somatosensory and auditory cortex localization by fMRI and MEG. Neuro-
report, 9, 1953–1957.
Suk, J., Ribary, U., Cappell, J., Yamamoto, T., & Llinas, R. (1991). Anatomical locali-
zation revealed by MEG recordings of the human somatosensory system. Electroen-
cephalogr Clin Neurophysiol, 78, 85–196.
Sutherling, W., Crandall, P., Darcey, T., Becker, D., Levesque, M., & Barth, D. (1988).
The magnetic and electric fields agree with intracranial localizations of somatosen-
sory cortex. Neurology, 38, 1705–1714.
Svensson, P., Minoshima, S., Beydoun, A., Morrow, T. J., & Casey, K. L. (1997). Cerebral
processing of acute skin and muscle pain in humans. J Neurophysiol, 78, 450–460.
Tanosaki, M., Kimura, T., Takino, R., Iguchi, Y., Suzuki, A., Kurobe, Y., et al. (2002).
Movement interference attenuates somatosensory high-frequency oscillations:
contribution of local axon collaterals of 3b pyramidal neurons. Clin Neurophysiol,
113, 993–1000.
344 MEG: An Introduction to Methods
Tecchio, F., Rossini, P. M., Pizzella, V., Cassetta, E., Pasqualetti, P., & Romani, G. L.
(1998). A neuromagnetic normative data set for hemispheric sensory hand cortical
representations and their interhemispheric differences. Brain Res Protoc, 2, 306–314.
Tiihonen, J., Hari, R., & Hamalainen, M. (1989). Early deflections of cerebral magne-
tic responses to median nerve stimulation. Electroencephalogr Clin Neurophysiol,
74, 290–296.
Trahms, L., Erne, S. N., Trontel, Z., Curio, G., & Aust, P. (1989). Biomagnetic functio-
nal localization of a peripheral nerve in man. Biophys J, 55, 1145–1153.
Tran, T. D., Lam, K., Hoshiyama, M., & Kakigi, R. (2001). A new method for measu-
ring the conduction velocities of A-β, Aδ- and C-fibers following electric and CO2
laser stimulation in humans. Neurosci Lett, 301, 187–190.
Tran, T. D., Inui, K., Hoshiyama, M., Lam, K., & Kakigi, R. (2002a). Conduction
velocity of the spinothalamic tract following CO2 laser stimulation of C-fiber in
humans. Pain, 95, 125–131.
Tran, T. D., Inui, K., Hoshiyama, M., Lam, K., Qiu, Y., & Kakigi, R. (2002b). Cerebral
activation by the signals ascending through unmyelinated C-fibers in humans: a
magnetoencephalographic study. Neuroscience, 113, 375–386.
Treede, R. D., Meyer, R. A., & Lesser, R. P. (1994). Similarity of threshold tempera-
tures for first pain sensation, laser-evoked potentials, and nociceptor activation.
In: G.F. Gebhart, D.L. Hammond and T.S. Jensen, eds. Proceedings of the 7th world
congress on pain. Seattle: IASP Press, pp. 857–865.
Treede, R. D., Kenshalo, D. R., Gracely, R.H., & Jones, A. K. (1999). The cortical
representation of pain. Pain, 79, 105–111.
Uesaka, Y., Ugawa, Y., Yumoto, M., Sakuta, M., & Kanazawa, I. (1993). Giant soma-
tosensory evoked magnetic field in patients with myoclonus epilepsy. Electroence-
phalogr Clin Neurophysiol, 87, 300–305.
Uesaka, Y., Terao, Y., Ugawa, Y., Yumoto, M., Hanajima, R., & Kanazawa, I. (1996).
Magnetoencephalographic analysis of cortical myoclonic jerks. Electroencephalogr
Clin Neurophysiol, 99, 141–148.
Wang, X., Inui, K., Qiu, Y., & Kakigi, R. (2004). Cortical responses to noxious stimuli
during sleep. Neuroscience, 128, 177–186.
Wasaka, T., Hoshiyama, M., Nakata, H., Nishihira, Y., & Kakigi, R. (2003). Gating
of somatosensory evoked magnetic fields during the preparatory period of self-
initiated finger movement. Neuroimage, 20, 1830–1838.
Wasaka, T., Nakata, H., Akatsuka, K., Kida, T., Inui, K., & Kakigi, R (2005).
Differential modulation in human primary and secondary somatosensory cortices
during the preparatory period of self-initiated finger movement. Eur J Neurosci, 22,
1239–1247.
Watanabe, S., Kakigi, R., Koyama, S., Hoshiyama, M., & Kaneoke, Y. (1998). Pain
processing traced by magnetoencephalography in the human brain. Brain Topogr,
10, 255–264.
Weiss, T., Miltner, W. H. R., Dillmann, J., Meissner, W., Huonker, R., & Nowak, H.
(1998). Reorganization of the somatosensory cortex after amputation of the index
finger. Neuroreport, 9, 213–216.
Whitzel, B. L., Perrucelli, L., & Werner, G. (1969). Symmetry and connectivity in the
body surface in somatosensory cortex: identification by combined magnetic and
potential recordings. J Physiol (Lond), 32, 218–223.
Wikstrom, H., Huttunen, J., Korvenoja, A., Virtanen, J., Salonen, O., Aronen, H.,
et al. (1996). Effects of interstimulus interval on somatosensory evoked magnetic
Somatosensory and Motor Function 345
Introduction
It is usually assumed that when we see a familiar word, like ‘brain’, the visual
features must be processed first before the analysis can proceed to the con-
tent, apparently first at the level of single letters and then as a whole word,
which further activates the word’s meaning and its sound form. How much
these later processing stages interact, and whether they occur sequentially or
in parallel, as a single interactive process, is an issue currently under debate
(Coltheart et al., 1993; Plaut et al., 1996). These theoretical models of reading
are based largely on analysis of behavioral reaction times and error types in
acquired and developmental reading disorders.
346
MEG and Reading: From Perception to Linguistic Analysis 347
Figure 13–1 depicts a paradigm used to tease apart early pre-lexical processes
in reading (Tarkiainen et al., 1999). The subjects were shown (Finnish) words,
Figure 13–2. MEG responses to letter and symbol strings in one subject. The
orange curves show responses to words with no noise, the green curves to
words at the highest noise level, and black curves to symbol strings. A
clear response particularly to noisy words (Type I) was followed by a left-
hemisphere response that was strongest to the noiseless words (Type II).
Modified from Tarkiainen et al. (1999).
MEG and Reading: From Perception to Linguistic Analysis 349
the string, similarly for letters and symbols. Only 50 ms later, at about 150 ms
post-stimulus, the left inferior occipitotemporal cortex showed letter-string
sensitive activation (10 of 12 subjects). This signal increased with the visibility
of the letter strings. It was strongest for words, weaker for syllables, and still
weaker for single letters. Crucially, the activation was significantly stronger
for letter than symbol strings of equal length.
One may ask how specific these processes are to reading, or whether they
reflect a more general transformation from visual to cognitive analysis.
Category-specific occipitotemporal responses within the first 200 ms have been
reported not only for letter-strings but, for example, also for numbers and
faces—that is, for particularly important types of objects in our visual world
(e.g., Allison et al., 1994). Tarkiainen and colleagues chose faces as test stimuli.
The faces were masked the same way as the letter-strings, and the subjects’
task was to identify the expressions on the faces (Tarkiainen et al., 2002).
Here, the control stimuli were pictures of objects. Figure 13–4 compares early
processing of letter-strings and faces in the same individuals. The stage of
visual feature processing at about 100 ms was the same for both stimulus
types. The timing, activated areas, and increase of activation with noise were
indistinguishable. Thereafter, the processing routes diverged. The timing of
the category-specific processing stage at about 150 ms was exactly the same
for letter-strings and faces, and the activated areas showed large spatial over-
lap. However, the hemispheric balance was different. While both left and
right occipitotemporal cortex respond to these stimuli (Cornelissen et al.,
2003; Salmelin et al., 1996), letter-string-sensitive differentiation was detected
essentially in the left hemisphere, whereas face-sensitive processing was more
bilateral, with slight right-hemisphere predominance (Tarkiainen et al., 2002).
350 MEG: An Introduction to Methods
Figure 13–4. Letter-string and face analysis within 200 ms after stimulus
presentation. (a) Cortical sequence of activation collected from individ-
ual subjects. Red indicates processing of letter-string stimuli and orange
processing of face stimuli. (b) Mean (± SEM) location of category-specific
activation at the base of the occipitotemporal cortex.
Modified from Tarkiainen et al. (2002).
Lexical–Semantic Analysis
words, and essentially non-existent for the expected words. This response is
usually referred to as the N400 in EEG literature (N400m in MEG) and its
behavior is generally considered as a signature of lexical-semantic processing
(Kutas & Hillyard, 1980; Osterhout & Holcomb, 1995). In the right hemi-
sphere, only about half of the subjects showed a qualitatively similar N400-
type response (Helenius et al., 1998). In reading, lexical-semantic processing
thus seems to be fairly strongly lateralized to the left hemisphere.
In fact, the complete suppression of the N400m response to the expected
final words is far more remarkable than generation of the strong response to
the semantically wrong words. Isolated words and the first words of sentences
all elicit a strong N400/N400m response. When one proceeds along the sen-
tence the response is gradually reduced to each word (Van Petten, 1995) until
the expectation built by the context (semantic priming) is strong enough to
entirely suppress the response to the expected word.
The onset of the N400m response, characterized by the latency at half the
maximum response on the ascending slope, was positively correlated with the
reaction time for recognizing real words in a lexical decision task (Helenius
et al., 1998). In a series of MEG studies, Marantz, Pylkkänen and colleagues
have demonstrated that the onset latency of the N400m activation (referred
to as M350 in their studies) reflects lexical frequency. By varying both pho-
notactic probability and neighborhood density, these authors concluded that
the response is related to lexical access rather than postlexical processing
(Embick et al., 2001; Pylkkänen & Marantz, 2003; Pylkkänen et al., 2002).
When the active areas are modeled as Equivalent Current Dipoles (ECDs;
cf. Chapter 6), sources of the N400m response are consistently localized to the
superior temporal cortex (Halgren et al., 2002; Helenius et al., 1998; Pylkkänen
MEG and Reading: From Perception to Linguistic Analysis 353
& Marantz, 2003; Salmelin et al., 1996; Simos et al., 1997), in the immediate
vicinity of the auditory cortex (Helenius et al., 1998). Distributed models (cf.
Chapter 8) suggest further spreading of activation to the anterior temporal
and inferior frontal cortex (Halgren et al., 2002; Marinkovic et al., 2003).
Involvement of the left temporal pole in semantic processing would agree
with previous intracranial recordings (Halgren et al., 1994; Nobre & McCarthy,
1995) which did not, however, probe the superior temporal cortex.
Figure 13–7 summarizes the cortical dynamics of silent reading, as
revealed by these MEG studies. First, there is basic visual feature analysis
around the occipital midline, at about 100 ms; then, 50 ms later, lateralization
to the left occipitotemporal cortex for letter-string analysis. Reading compre-
hension is reflected in the subsequent activation of the left superior temporal
cortex at 200 to 600 ms.
The strong left-hemisphere lateralization of lexical-semantic processing
in reading, at least in Finnish-speaking subjects, could possibly serve as a
diagnostic tool. It has been used to evaluate cerebral implementation of read-
ing comprehension in an aphasic patient with deep dyslexia (Laine et al.,
2000). The central features of deep dyslexia are the abolishment of nonword
reading, and semantic errors in reading: the patient may, for example, read
the word “moon” as “crescent” (Coltheart, 1980). It has been suggested that
Behaviorally, it has been found that naming latencies are shorter for short than
long words, and that this length effect is markedly enhanced when naming non-
words (Weekes, 1997). The use of the phonological route, and thus the length
effect, is thought to be particularly pronounced in a regular orthography.
The Finnish language has an exceptionally regular one-to-one correspon-
dence between graphemes and phonemes and should, therefore, be well suited
for cortical evaluation of the potential lexical-semantic vs. phonological routes
of reading. Wydell and colleagues varied the letter-string length and lexicality
by presenting short and long real Finnish words (4 and 8 letters) and short and
long pronounceable nonwords in a randomized order (Wydell et al., 2003). The
subjects were occasionally prompted to read the string aloud, in an unpredict-
able fashion, thus emphasizing full phonological encoding of the letter-strings.
There were only two areas and time windows where the activation showed
systematic dependence on stimulus lexicality or length (Figure 13–9; cf.
Chapter 6 where this data set was used as an example in demonstrating source
analysis of complex cognitive data). An early length effect was evident in the
occipital midline at about 100 ms after stimulus onset. The long letter-strings
grouped together and evoked a stronger response than the short letter-strings,
regardless of letter-string type (lexicality). Based on previous knowledge
about cortical dynamics of silent reading, reviewed above, this response is
likely to reflect basic visual feature analysis. The subsequent activation in the
left inferior occipitotemporal cortex, interpreted as letter-string-sensitive acti-
vation (cf. Figure 13–7), did not vary with length or lexicality, in agreement
with existing MEG and intracranial data (Cornelissen et al., 2003; Nobre et al.,
1994; Salmelin et al., 1996).
As for the subsequent stages of semantic and phonological analysis, the MEG
data reviewed above suggest that the left superior temporal activation at 200
to 600 ms reflects both of those processes (Figure 13–14a). The response is
reduced by semantic priming, and there is stronger activation to pseudowords
than real words.
During the past few years, the representation of semantic and phonologi-
cal processes in reading has been addressed in a number of fMRI/PET studies.
Jobard and others, based on a meta-analysis of 35 reports (Jobard et al., 2003),
suggest the pattern sketched in Figure 13–14b. Semantic processing was most
consistently associated with activation of the triangular part of the inferior
frontal gyrus, posterior middle temporal gyrus, and basal temporal cortex,
whereas phonological processing was reflected in activation of the superior
temporal cortex, supramarginal gyrus, and opercular part of the inferior
frontal gyrus.
There is some agreement between neurophysiological and hemodynamic
methods for phonological processing, but apparently none for semantic anal-
ysis. The left superior temporal cortex is implicated in phonology both by
MEG and hemodynamic measures, but only MEG assigns it a role in seman-
tics. For this particular pattern, one could consider a rather simplistic account:
In fMRI/PET studies the active brain areas are determined by specific sub-
tractions. Based on the dual-route model of reading, areas involved in phono-
logical analysis are thought to be revealed by subtracting activations to real
words from those to pseudowords. Areas involved in semantic processing, on
the other hand, would be sought by the inverse subtraction. Based on the
MEG data, and assuming that both semantic and phonological manipulation
affects essentially the same neuronal population, such subtractions would
364 MEG: An Introduction to Methods
indeed show activation of the left superior temporal cortex for the compari-
son pseudowords > words (phonology) but none for the comparison words
> pseudowords (semantics).
On the other hand, one may also ask whether the differences might be
due to the choice of language. Most of the MEG studies were performed using
Finnish, whereas English has been the prevalent language in fMRI/PET studies.
A recent fMRI study used Japanese language, which has a highly regular
orthography, like Finnish (Ischebeck et al., 2004). One of the tasks—silent
articulation of visually familiar and unfamiliar words and pseudowords—was
very similar to that used in the MEG study of word and pseudoword reading
described above (Wydell et al., 2003). Nevertheless, in the fMRI pattern, there
was again a striking lack of left superior temporal activation, now also for
phonological processing. Clearly, it will be essential to establish the similari-
ties and differences between hemodynamic and MEG measures, what they tell
about the processes involved in reading, and how these processes are organized
in the brain.
MEG and Reading: From Perception to Linguistic Analysis 365
is generally much lower than in children. Since the ability to analyze speech
sounds is considered a prerequisite for reading acquisition (Bradley & Bryant,
1983), this intriguing correlation again points to an important role for the
left inferior occipitotemporal cortex in reading acquisition and fluent
reading.
Studies on children learning to read are likely to prove essential also for
understanding the relationship between perception of spoken and written
language. Cortical activation patterns in dyslexic adults seem to point to
impaired integration of auditory and visual information that may be specific
to language, or possibly reflects a more general difficulty in multisensory inte-
gration. This type of interactions between input modalities, particularly in the
interesting early time windows (< 300 ms), are quite difficult to assess in the
relatively rigid adult brain. The developing brain of a child could provide a
clearer view into such processes.
The different components of reading are typically probed with rather artificial
experimental setups that allow good control of the task and the stimulus
properties. While this is a well-argumented and necessary approach, the brain
correlates of language processing may appear quite different in more natural
contexts that the brain is tuned for. For example, how important is the occip-
itotemporal letter-string area when we are reading continuously, in more
realistic conditions?
MEG and Reading: From Perception to Linguistic Analysis 367
of brain areas that have been identified in activation studies on various aspects
of language function, using either MEG, fMRI or PET, or intracranial record-
ings: The occipitotemporal node corresponds approximately to the letter-
string area, and the superior, anterior and medial temporal areas have been
suggested to be involved in semantic and phonological analysis. Face motor
cortex, insula and the cerebellum are typically reported in language produc-
tion, thus suggesting a connection between sensory and motor processes even
in silent reading. The prefrontal and orbitofrontal cortex have been associated
especially with visual recognition and working memory.
The entire network was strongly interconnected (Figure 13–16a). The
connections were mostly bidirectional, but for part of the connections there
was a dominant direction of information flow, as evaluated with Granger
causality (Figure 13–16b). Intriguingly, the left inferior occipitotemporal cor-
tex, together with the cerebellum, turned out to be the main forward-driving
node of the network, again emphasizing the significance of this area in reading.
Indeed, in the light of this network structure it is not surprising that impaired
neural processing in this area in dyslexia may seriously affect the ability to
read fluently.
Rhythmic cortical activity also shows event-related modulation that is
not as tightly time-locked to the stimulus or task onset/offset as the strictly
phase-locked evoked responses, and may offer important complementary
information of brain function (see Chapters 6–9). Localization of rhythmic
activity, and especially of those areas in which rhythmic activity is reduced, is
not straightforward. Recently, methods have been developed that allow visu-
alization and localization of brain areas where the level of rhythmic activity
within a specific frequency range is suppressed below a predefined base level;
DICS, mentioned above, is one of those methods. Within the domain of MEG
as a measure of brain function, it will be essential to understand the relation-
ship between evoked responses, event-related modulation of cortical rhythms
and connectivity patterns, as regards their location, timing and functional
dependence on parametric variation of stimuli and tasks. Furthermore, the
relationship of these MEG signatures with hemodynamic measures, such as
the fMRI BOLD signal, will need to be established for efficient use of neu-
roimaging methods in unraveling the principles of neural processing.
the nearness of both visually and auditorily evoked N400m response to the
primary auditory cortex, deserve careful investigation.
Perhaps the time is now ripe for starting to let the brain inform us about
the ways in which it prefers to cope with written language, and the types of
processes implemented at the neuronal level. Combined spatial and temporal
information will be essential in that endeavor.
References
Allison, T., McCarthy, G., Nobre, A., Puce, A., & Belger, A. (1994). Human extrastri-
ate visual cortex and the perception of faces, words, numbers, and colors. Cerebral
Cortex, 5, 544–554.
Biermann-Ruben, K., Salmelin, R., & Schnitzler, A. (2005). Right rolandic activation
during speech perception in stutterers: an MEG study. Neuroimage, 25, 793–801.
Bradley, L., & Bryant, P. E. (1983). Categorizing sounds and learning to read – a causal
connection. Nature, 301, 419–421.
Brunswick, N., McCrory, E., Price, C. J., Frith, C. D., & Frith, U. (1999). Explicit and
implicit processing of words and pseudowords by adult developmental dyslexics –
A search for Wernicke’s Wortschatz? Brain, 122, 1901–1917.
Cohen, L., Dehaene, S., Naccache, L., Lehericy, S., DehaeneLambertz, G., Henaff M.,
et al. (2000). The visual word form area – Spatial and temporal characterization
of an initial stage of reading in normal subjects and posterior split-brain patients.
Brain, 123, 291–307.
Cohen, L., Lehericy, S., Chochon, F., Lemer, C., Rivaud, S., & Dehaene, S. (2002).
Language-specific tuning of visual cortex functional properties of the Visual Word
Form Area. Brain, 125, 1054–1069.
Coltheart, M. (1980). Deep dyslexia: A right-hemisphere hypothesis. In: K. Pat-
terson & J. C. Marshall (eds.). Deep dyslexia. London: Routledge & Kegan Paul.
pp. 326–380.
Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993). Models of reading aloud:
dual-route and parallel-distributed-processing approaches. Psychological Review,
100, 589–608.
Connolly, J. F., and Phillips, N. A. (1994). Event-related potential components reflect
phological and semantic processing of the terminal word of spoken sentences.
Journal of Cognitive Neuroscience, 6, 256–266.
Cornelissen, P., Tarkiainen, A., Helenius, P., & Salmelin, R. (2003). Cortical effects
of shifting letter-position in letter-strings of varying length. Journal of Cognitive
Neuroscience, 15, 731–746.
Embick, D., Hackl, M., Schaeffer, J., Kelepir, M, & Marantz, A. (2001). A magne-
toencephalographic component whose latency reflects lexical frequency. Cognitive
Brain Research 10, 345–348.
Gross, J., Kujala, J., Hämäläinen, M., Timmermann, L., Schnitzler, A, & Salmelin, R.
(2001). Dynamic imaging of coherent sources: Studying neural interactions in the
human brain. Proceedings of the National Academy of Sciences of USA, 98, 694–699.
Gross, J., Timmermann, J., Kujala, J., Dirks, M., Schmitz, F., Salmelin, R., et al. (2002).
The neural basis of intermittent motor control in humans. Proceedings of the Natio-
nal Academy of Sciences of USA, 99, 2299–2302.
370 MEG: An Introduction to Methods
Halgren, E., Baudena, P., Heit, G., Clarke, M., & Marinkovic, K. (1994). Spatio-temporal
stages in face and word processing. 1. Depth-recorded potentials in the human
occipital, temporal and parietal lobes. Journal of Physiology (Paris) 88, 1–50.
Halgren, E., Dhond, R. P., Christensen, N., Van Petten, C., Marinkovic, K., Lewine,
J. D., et al. (2002). N400-like magnetoencephalography responses modulated by
semantic context, word frequency, and lexical class in sentences. Neuroimage, 17,
1101–1116.
Hari, R. (1990). The neuromagnetic method in the study of the human auditory cor-
tex. In: F. Grandori, M. Hoke, and G. L. Romani (eds.). Auditory Evoked Magnetic
Fields and Electric Potentials. Vol 6. Basel: S. Karger, pp. 222–282.
Helenius, P., Salmelin, R., Richardson, U., Leinonen, S., & Lyytinen, H. (2002a).
Abnormal auditory cortical activation in dyslexia 100 msec after speech onset.
Journal of Cognitive Neuroscience, 14, 603–617.
Helenius, P., Salmelin, R., Service, E., & Connolly, J. F. (1998). Distinct time courses
of word and sentence comprehension in the left temporal cortex. Brain, 121, 1133–
1142.
Helenius, P., Salmelin, R., Service, E., & Connolly, J. F. (1999a). Semantic cortical
activation in dyslexic readers. Journal of Cognitive Neuroscience, 11, 535–550.
Helenius, P., Salmelin, R., Service, E., Connolly, J. F., Leinonen, S., & Lyytinen, H.
(2002b). Cortical activation during spoken-word segmentation in nonreading-
impaired and dyslexic adults. Journal of Neuroscience, 22, 2936–2944.
Helenius, P., Tarkiainen, A., Cornelissen, P., Hansen, P.C., & Salmelin, R. (1999b).
Dissociation of normal feature analysis and deficient processing of letter-strings in
dyslexic adults. Cerebral Cortex, 9, 476–483.
Ischebeck, A., Indefrey, P., Usui, N., Nose, I., Hellwig, F., & Taira, M. (2004). Reading
in a regular orthography: An fMRI study investigating the role of visual familiarity.
Journal of Cognitive Neuroscience, 16, 727–741.
Jobard, G., Crivello, F., & TzourioMazoyer, N. (2003). Evaluation of the dual route
theory of reading: a metanalysis of 35 neuroimaging studies. Neuroimage, 20,
693–712.
Kujala, J., Pammer, K., Cornelissen, P., Roebroeck, A., Formisano, E., & Salmelin R.
(2007). Phase coupling in a cerebro-cerebellar network at 8-13 Hz during reading.
Cerebral Cortex, 17, 1476–1485.
Kutas, M., and Hillyard, S. A. (1980). Reading senseless sentences: brain potentials
reflect semantic incongruity. Science, 207, 203–205.
Laine, B., Salmelin, R., Helenius, P., & Marttila R. (2000). Brain activation during
reading in deep dyslexia: An MEG study. Journal of Cognitive Neuroscience, 12,
622–634.
Marinkovic, K., Dhond, R. P., Dale, A. M, Glessner, M., Carr, V., & Halgren, E. (2003).
Spatiotemporal dynamics of modality-specific and supramodal word processing.
Neuron, 38, 487–497.
Nagarajan, S., Mahncke, H., Salz, T., Tallal, P., Roberts, T., & Merzenich, M. M. (1999).
Cortical auditory signal processing in poor readers. Proceedings of the National
Academy of Sciences of USA, 96, 6483–6488.
Nobre, A. C., Allison, T., & McCarthy, G. (1994). Word recognition in the human
inferior temporal lobe. Nature, 372, 260–263.
Nobre, A. C., and McCarthy, G. (1995). Language-related field potentials in the
anterior-medial temporal lobe: II. Effects of word type and semantic priming.
Journal of Neuroscience, 15, 1090–1098.
MEG and Reading: From Perception to Linguistic Analysis 371
Osterhout, L., & Holcomb, PJ. (1995). Event-related potentials and language compre-
hension. In: M. D. Rugg M. G. H. Coles (eds.). Electrophysiology of mind. Oxford:
Oxford University Press. pp, 171–215.
Parviainen, T., Helenius, P., Poskiparta, E., Niemi, P., & Salmelin, R. (2006). Corti-
cal sequence of word perception in beginning readers. Journal of Neuroscience, 26,
6052–6061.
Parviainen, T., Helenius, P., & Salmelin, R. (2005). Cortical differentiation of speech and
nonspeech sounds at 100 ms: implications for dyslexia. Cerebral Cortex, 15, 1054–1063.
Paulesu, E., Demonet, J. F., Fazio, F., McCrory, E., Chanoine, V., Brunswick, N., et al.
(2001). Dyslexia: Cultural diversity and biological unity. Science, 291, 2165–2167.
Paulesu, E., McCrory, E., Fazio, F., Menoncello, L., Brunswick, N., Cappa, S. F., et al.
(2000). A cultural effect on brain function. Nature Neuroscience, 3, 91–96.
Phillips, C., Pellathy, T., Marantz, A., Yellin, E., Wexler, K., Poeppel, D., et al. (2000).
Auditory cortex accesses phonological categories: An MEG mismatch study. Jour-
nal of Cognitive Neuroscience, 12, 1038–1055.
Plaut, D., McClelland, J., Seidenberg, M., & Patterson, K. (1996). Understanding
normal and impaired word reading: computational principles in quasi-regular
domains. Psychological Review, 103, 56–115.
Poeppel, D., Yellin, E., Phillips, C., Roberts, T. P. L., Rowley, H. A, Wexler, K.,
et al. (1996). Task-induced asymmetry of the auditory evoked M100 neuromagne-
tic field elicited by speech sounds. Cognitive Brain Research, 4, 231–242.
Price, C. J., & Devlin, J. T. (2003). The myth of the visual word form area. Neuroimage,
19, 473–481.
Puce, A., Allison, T., Asgari, M., Gore, J. C., & McCarthy, G. (1996). Differential
sensitivity of human visual cortex to faces, letterstrings, and textures: a functional
magnetic resonance imaging study. Journal of Neuroscience, 16, 5205–5215.
Pylkkänen, L., & Marantz, A. (2003). Tracking the time course of word recognition
with MEG. Trends in Cognitive Sciences, 7, 187–189.
Pylkkänen, L., Stringfellow, A., & Marantz, A.(2002). Neuromagnetic evidence for the
timing of lexical activation: an MEG component sensitive to phonotactic probabi-
lity but not to neighborhood density. Brain and Language, 81, 666–678.
Salmelin, R. (2007). Clinical neurophysiology of language: The MEG approach. Clinical
Neurophysiology, 118, 237–254.
Salmelin, R., & Helenius, P. (2004). Functional neuroanatomy of impaired reading in
dyslexia. Scientific Studies of Reading, 8, 257–272.
Salmelin, R., Helenius, P., & Service, E.(2000). Neurophysiology of fluent and impai-
red reading: a magnetoencephalographic approach. Journal of Clinical Neurophy-
siology, 17, 163–174.
Salmelin, R., & Kujala, J. (2006). Neural representation of language: activation versus
long-range connectivity. Trends in Cognitive Sciences, 10, 519–525.
Salmelin, R., Schnitzler, A., Parkkonen, L., Biermann, K., Helenius, P., Kiviniemi, K.,
et al.(1996). Native language, gender, and functional organization of the auditory
cortex. Proceedings of the National Academy of Sciences of USA, 96, 10460–10465.
Salmelin, R., Service, E., Kiesilä, P., Uutela, K., & Salonen O. (1996). Impaired visual
word processing in dyslexia revealed with magnetoencephalography. Annals of
Neurology, 40, 157–162.
Service, E., Helenius, P., Maury, S., & Salmelin, R. (2007). Localization of syntactic
and semantic brain responses using magnetoencephalography. Journal of Cognitive
Neuroscience 19, 1193–1205.
372 MEG: An Introduction to Methods
Shaywitz, S. E., Shaywitz, B. A., Pugh, K. R., Fulbright, R. K., Constable, R. T.,
Mencl, W. E., et al. (1998). Functional disruption in the organization of the brain
for reading in dyslexia. Proceedings of the National Academy of Sciences of USA, 95,
2636–2641.
Simos, P. G., Basile, L. F., & Papanicolaou, A. C. (1997). Source localization of the
N400 response in a sentence-reading paradigm using evoked magnetic fields and
magnetic resonance imaging. Brain Research, 762, 29–39.
Simos, P. G., Breier, J. I., Fletcher, J. M., Bergman, E., & Papanicolaou, A. C. (2000a)
Cerebral mechanisms involved in word reading in dyslexic children: A magnetic
source imaging approach. Cerebral Cortex, 10, 809–816.
Simos, P. G., Breier, J. I., Fletcher, J. M., Foorman, B. R., Bergman, E., Fishbeck, K.,
et al. (2000b). Brain activation profiles in dyslexic children during non-word rea-
ding: a magnetic source imaging study. Neuroscience Letters, 290, 61–65.
Simos, P. G., Breier, J. I., Fletcher, J. M., Foorman, B. R., Castillo, E. M., & Papani-
colaou, A. C. (2002a). Brain mechanisms for reading words and pseudowords: an
integrated approach. Cerebral Cortex, 12, 297–305.
Simos, P. G., Fletcher, J. M., Bergman, E., Breier, J. I., Foorman, B. R., Castillo, E. M.,
et al. (2002b). Dyslexia-specific brain activation profile becomes normal following
successful remedial training. Neurology, 58, 1203–1213.
Tarkiainen, A., Cornelissen, P. L., & Salmelin, R. (2002). Dynamics of visual feature
analysis and object-level processing in face versus letter-string perception. Brain,
125, 1125–1136.
Tarkiainen, A., Helenius, P., Hansen, P. C., Cornelissen, P. L., & Salmelin, R. (1999).
Dynamics of letter string perception in the human occipitotemporal cortex. Brain,
122, 2119–2131.
Tarkiainen, A., Helenius, P., & Salmelin, R. (2003). Category-specific occipitotemporal
activation during face perception in dyslexic individuals: an MEG study. Neuro-
image, 19, 1194–1204.
Van Petten, C. (1995). Words and sentences: event-related brain potential measures.
Psychophysiology, 32, 511–525.
Vihla, M., Lounasmaa, O. V., & Salmelin, R. (2000). Cortical processing of change
detection: Dissociation between natural vowels and two-frequency complex tones.
Proceedings of the National Academy of Sciences of USA, 97, 10590–10594.
Weekes, B., Coltheart, M., & Gordon, E. (1997). Deep dyslexia and right hemisphere
reading – a regional cerebral blood flow study. Aphasiology, 11, 1139–1158.
Weekes, B. S. (1997). Differential effects of number of letters on word and nonword
latency. Quarterly Journal of Experimental Psychology, 50A, 439–456.
Wilson, T. W., Leuthold, A. C., Lewis, S. M., Georgopoulos, A. P., & Pardo, P. J.
(2005). The time and space of lexicality: a neuromagnetic view. Experimental Brain
Research, 162, 1–13.
Wydell, T. N., Vuorinen, T., Helenius, P., & Salmelin, R. (2003). Neural correlates of
letter-string length and lexicality during reading in a regular orthography. Journal
of Cognitive Neuroscience, 15, 1052–1062.
14
The Use of MEG in Clinical Settings
Jyrki P. Mäkelä
Introduction
Search for clinical applications has paralleled the development of MEG from
its early phases in the seventies (Hughes et al., 1977). The possibilities of MEG
in clinical use were first demonstrated by studies of patients with several types
of epilepsy (Barth et al., 1982; Modena et al., 1982; Barth et al., 1984). At the
time, data were obtained sequentially by moving the one-sensor instrument,
373
374 MEG: An Introduction to Methods
General Aspects
The use of MEG in clinical settings requires an approach different from that
in research. Typically, individual features in measurements are of crucial
importance in clinical patient studies, whereas they are of minor interest or
even confounding in research settings that aim to reveal general features of
the brain function. Different approaches in the practical performance of mea-
surements are also apparent. Healthy control subjects are often familiar with
the recording environment, whereas patients seldom visit a MEG unit more
than one or two times. However, high motivation for obtaining personally
significant results from measurements often produces excellent cooperation
The Use of MEG in Clinical Settings 375
Data Quality
Magnetic Artifacts
MEG signals are extremely tiny, and the recordings are sensitive to artifacts
produced by moving magnetic materials. Even hair dyes or cosmetics may
376 MEG: An Introduction to Methods
Patient Safety
Unexpected events, such as epileptic seizures or drug-induced respiratory
suppression, are more probable in studies of patients than in control subjects.
378 MEG: An Introduction to Methods
If studied supine, safety belts or railings may prevent the subject from falling
off the bed in such instances. The examiners should be well versed on how to
get the subject quickly out of the gantry if needed. Resuscitation equipment,
oxygen system and, preferably, suction instrumentation need to be prepared
for an emergency.
MEG recordings are a useful adjunct in planning epilepsy surgery. MEG appears
to be particularly beneficial in the study of patients with non-lesional neocor-
tical epilepsy, and in patients with large lesions, where it may provide unique
information on the epileptogenic zone in relation to the lesion (for reviews,
see Pataraia et al., 2002; Barkley & Baumgartner, 2003; Knowlton & Shih,
2004). Naturally, the preoperative localization of eloquent cortices can be made
with the same methods in epileptic patients as in control subjects, and their
relation to epileptic zone can be visualized (Figure 14–1). Sources of epileptic
spikes can be integrated into neuronavigation systems (Iida et al., 2005).
It has been suggested that MEG source localization, using a single-dipole
model, can provide unique localization information not available with other
noninvasive methods in patients with epilepsy (Mamelak et al., 2002). The
clustering of the sources of individual interictal spikes (Figure 14–2) has dem-
onstrated a high correlation with electrocortigography (ECoG) (Lamusuo
et al., 1999; Mamelak et al., 2002). If source clusters are located in the nonre-
sectable eloquent cortex, residual seizures remain probable (Iida et al., 2005),
and a high correlation of the resection volume with the brain region contain-
ing MEG source clusters of epileptic spikes has been shown to predict favorable
outcome in epilepsy surgery (Fischer et al., 2005). MEG can also encourage
epilepsy surgery when displaying focal epileptiform activity, whereas tradi-
tional methods suggest multifocal activity or demonstrate bilateral, multifocal
or diffuse ictal onset, indicating an unfavorable candidate for epilepsy surgery
(Schwartz et al., 2008).
Figure 14–1. MEG signal from a triggered epileptic seizure. The patient
has partial epilepsy with seizures triggered by touching of the left gum
or corner of the mouth, including left facial jerking. The whole seizure
from the channel showing the maximum signal in the right hemisphere is
depicted in the box above. Below, the sections A, B and C show the devel-
opment of the seizure, as well as activity in the corresponding region in
the left hemisphere, in enlarged form. Before the seizure onset, spikes
emerge more frequently and become polyphasic in the right frontopari-
etal region. No notable activity over the left hemisphere is seen during
the first 6 s; afterwards, the spike discharge spreads to the left side as
well. After the seizure, interictal spikes are absent. The sources of epilep-
tic activity (spikes) cluster to the face motor cortex representation. Sources
of median nerve SEFs and AEFs are shown to indicate irretrievable areas.
Modified from Forss et al. (1995).
Figure 14–2. (A) Dipolar field patterns from two different epileptiform
spike types in the right parietal and temporal lobe in a patient with
intractable epilepsy, going through epilepsy surgery evaluation. (B) Clusters
of spike sources superimposed on the patient’s 3D MRI. (C) PET data dem-
onstrating hypometabolism in right temporal and parietal lobe (arrows).
SEEG demonstrated right temporal and parieto-occipital epileptiform
activity. MRI showed a small region of atrophy in the right parietal region.
Right parietal region was operated, resulting in worthwhile reduction of
seizures (Engel’s classification IIIa).
Modified from (Lamusuo et al. 1999).
One factor diminishing the yield of MEG in patients with epilepsy is the lack
of epileptiform activity in the limited time window of the recording. Usually,
a recording of at least 30 minutes of spontaneous activity, including periods
of drowsiness, is needed. Hyperventilation for 3 minutes is a practical way to
enhance the appearance of spikes in MEG measurements. MEG artifacts often
abound during hyperventilation, but the study of the post-hyperventilation
period may turn out to be useful. Photostimulation, routinely used in EEG
recordings, is more inconvenient to use effectively in the MEG setup. Sleep
deprivation during the night preceding the recording will increase the prob-
ability of the detection of the epileptic activity, although the quality of evoked
fields may decrease due to the lowered vigilance. Monitoring occipital MEG
signals to detect alpha rhythm changes may be useful to guarantee good quality
of the evoked fields.
Tapering of antiepileptic medication may provoke epileptiform activity;
follow-up on the hospital ward is needed for this procedure. Some epilepti-
form activity can be detected only during sleep; consequently, recording dur-
ing the night may prove useful. The yield of MEG recordings has been claimed
to approach 100% when patients with temporal lobe epilepsy were on sub-
therapeutic anticonvulsant levels, and sleep was encouraged (Assaf et al.,
2004). Video recording may provide useful clues of the epilepsy syndrome
and signal interpretation, if seizures occur during the MEG measurement.
Anesthesia may be needed when studying children between ages 6
months–5 years; an anesthetic regimen using propofol appears not to reduce
the occurrence of epileptiform activity, and has caused no problems in asso-
ciation with MEG recordings (Balakrisnan et al., 2007; Szmuk et al., 2003),
although it may produce seizures in rare cases (e.g., Mäkelä et al., 1993).
Appearance of epileptiform activity is enhanced by clonidine (Kettenmann
et al., 2005), and by some anesthetic agents and, consequently, recordings
under anesthesia may even enhance detection of the epileptic spikes. Naturally,
particularly careful monitoring of the patient by oximetry and heart rate
detection is required if anesthesia or sedation is used. Administration of anxi-
olytic drugs may also be necessary in some cases (e.g. Schwartz et al., 2008).
382 MEG: An Introduction to Methods
Invasive video EEG monitoring has been a gold standard for defining the epi-
leptogenic cortex prior to surgery. However, it is quite demanding for the
patient and may cause bleeding or infections (Hamer et al., 2002). As epilepsy
surgery is usually elective, such events should be prevented if possible.
Furthermore, a rough estimate concerning the epileptogenic cortical areas is
required before the insertion of the intracranial electrodes. MEG source local-
ization aids in selecting sites for grids to be used in subdural EEG recordings
(Mamelak et al., 2002; Vitikainen et al. 2009; Knowlton et al. 2009; Sutherling
et al. 2008). It has been suggested as particularly useful in the detection of
epileptic activity after lesionectomy, or unsuccessful removal of the epileptic
zone, because dural adhesions may hamper the insertion of subdural elec-
trode grids in these patients (Kirchberger et al., 1998).
Comparisons of preoperative MEG findings with ECoG have occasion-
ally found almost complete matches (Lamusuo et al., 1999; Otsubo et al.,
2001), whereas some others report lower values (Mamelak et al., 2002). The
patient populations have been quite variable, and the location of the seizure
focus probably affects the degree of correlation between MEG and invasive
EEG recordings.
Combination of MEG with ictal SPECT may replace invasive EEG moni-
toring (Knowlton, 2006). Combining MEG with navigated transcranial mag-
netic stimulation may also turn out to be useful in diminishing invasive EEG
measurements (Vitikainen et al., 2009). MEG appears to be as accurate as
interictal and ictal invasive video-EEG (Papanicolau et al., 2005)—and ictal
MEG recordings produced localization equivalent or superior to invasive
EEG in five out of six patients (Eliashiv et al., 2002). Nevertheless, although
MEG predicted 82% of findings in invasive recordings of 49 patients, epilep-
togenic cortex remained nonlocalized in 7 of them by MEG, whereas invasive
recordings were diagnostic (Knowlton et al., 2006).
The Use of MEG in Clinical Settings 383
MEG and EEG signals look similar, and the knowledge of different visual pat-
terns of epileptiform phenomena, collected since the 1920s, can be applied in
MEG analysis. Nevertheless, MEG source modeling provides information not
available in EEG. In simultaneous recordings of epileptiform activity by MEG
and subdural ECoG it has been estimated that at least 4 cm2 of synchronously
active cortex is needed to produce a detectable MEG spike (Mikuni et al.,
1997). However, even cortical spikes associated with 6–10 cm2 of synchro-
nous activity in ECoG rarely generate scalp-recordable EEG interictal spikes;
an area exceeding 10 cm2 is required for recognizable scalp potentials (Tao et al.,
2005). Indeed, in a cohort of 70 candidates for epilepsy surgery, whole-head
MEG detected epileptiform activity in 72%, and simultaneous 70-channel
EEG in 61% of the patients. MEG revealed epileptiform activity in one third
of the EEG-negative patients, particularly in patients having lateral neocortical
epilepsy or cortical dysplasia (Knake et al., 2006). Furthermore, a higher ratio
of spikes unique to MEG (8/12 patients) compared with EEG (2/12 patients)
is detected when the signals are overlapped by sleep changes; this has been
attributed to stronger contribution of radial sources of sleep spindles and
vertex waves on EEG (Ramantani et al., 2006). MEG may be useful in detect-
ing epileptic activity deep in the sulci, masked in EEG by more superficial
radial activity in the gyri (Merlet et al., 1997)—e.g., in Landau-Kleffner syn-
drome, in which the spike activity typically resides deep in the Sylvian fissure
(Paetau et al., 1999; Iwasaki et al., 2003). Consequently, MEG can be applied
also in patients with suspected epilepsy but with a normal EEG, and in
epilepsy patients whose epilepsy type remains unclassified on the basis of
the EEG.
The simultaneous recording of MEG and EEG, and the use of both meth-
ods in modeling the epileptiform activity, is crucial for a complete view of the
epileptogenic zone. Combined MEG and EEG can identify the source areas
and their activation sequences in more detail, thereby helping to select patients
with a single pacemaker area and prospects for good outcome after surgery
(Lewine et al., 1999; Paetau et al., 1999), and should at least be included into
analysis of particularly difficult cases of intractable epilepsy. Strategies for a
unified model of brain electric activity as recorded both by MEG and EEG
have been delineated (Huang et al., 2007); however, they have not yet been
utilized in the analysis of epileptiform activity.
Stereotactic radiation therapy with high single doses is suitable for lesions
with sharp boundaries, such as cortical meningioma, and for high-grade
glioma recidives. When applying high doses, it is important to avoid radiation
to the surrounding intact brain areas. MEG source locations may provide use-
ful information for the dose planning if the lesion is located close to eloquent
brain areas (Aoyama et al., 2003).
MEG landmarks (Rezai et al., 1996; Ganslandt et al., 1997; Schiffbauer et al.,
2002), and sources of epileptiform activity (Iwasaki et al. 2003) can be incor-
porated into image-guided stereotactic methods for a more precise naviga-
tion during operation. However it has been shown that the cortical surface
shifts 5–10 mm after dural opening during the surgery (Roberts et al., 1998).
The largest shift sometimes occurs near the center of the craniotomy, which
is usually the brain region of the greatest interest (Hill et al., 2000). These
shifts may introduce problems for pre- and intraoperative site comparison.
Depicting surface veins in combination with 3-D brain structures and
The Use of MEG in Clinical Settings 387
Functional MRI (fMRI) provides information about the location of the sen-
sorimotor cortex in about 90-95% of the studied patients (Krings et al., 2001;
Pujol et al., 2008). Although the method is not applicable in patients with a
complete hand paralysis, it is needed in about 7% of patients harboring an
intracranial mass (Pujol et al., 2008). Differences in central sulcus localization
by MEG and fMRI have been reported; a 15 ± 5 mm mean difference of the
SEF source location, and the somatosensory elicited activation in fMRI in the
same patients (Kober et al., 2001b) may exceed the gyral width. In patients
with tumors in the vicinity of the central sulcus, the fMRI and MEG localiza-
tion of the central sulcus were discordant in about 20% of the affected hemi-
spheres; MEG localizations matched with the intraoperative mapping (Inoue
et al., 1999; Korvenoja et al., 2006). As fMRI integrates brain activity over a
period of several seconds, it reveals the whole cortical network participating
in the processing of external stimuli or a task. Limited resolution in the time
domain may consequently result in difficulties in separating the primary
areas from secondary processing areas. Strong fMRI activations in nonpri-
mary areas may, therefore, confound the interpretation of activation maps
(Korvenoja et al., 2006). This drawback is avoided in MEG measurements
detecting cortical activity with millisecond temporal accuracy, which sepa-
rates the primary somatosensory cortex response from secondary activations
(Hari & Forss, 1999); MEG should be used when detailed central sulcus
localization is needed.
MEG and diffusion tensor imaging is also suggested to be more sensitive than
conventional imaging methods in detecting subtle neuronal lesions in mild
traumatic brain injury (Huang et al., 2009).
Strict criteria prolong the research project, whereas too loose a patient selection
makes interpretation of the results difficult.
Cortical Reorganization
Objective means to predict potential for rehabilitation and to follow its course
would be useful in the follow-up of neurological patients. It has been sug-
gested that extensive practice, or lack of use, of a certain body part may change
the somatotopic organization of the primary sensory cortex. Cortical repre-
sentations of fingers, as depicted by SEF sources elicited by tactile finger stim-
ulation, have been shown to differentiate after the treatment of syndactyly
(Mogilner et al., 1993). Moreover, the amputation of the arm modifies SEF
source structure in a manner suggesting plastic changes in the primary
sensory cortex (Flor et al., 1995). MEG studies have recently shown that mod-
ifications of SEF source locations occur in association with chronic pain
without nerve deafferentation (Juottonen et al., 2002; Maihöfner et al., 2003).
Furthermore, abnormal ipsilateral SEFs developed in association with the
mirror-like spread of chronic regional pain from one upper limb to another,
during a 3-year follow-up of one patient (Forss et al., 2005). Such changes
might be useful as objective correlates of perceived pain. The changes in SEF
source organization are, at least to some extent, reversible, suggesting that
rehabilitation should be targeted for regaining the orderly somatotopic
arrangement at the primary sensory cortex (Maihöfner et al., 2004). An objec-
tive follow-up of such changes is an intriguing new possibility for clinical
applications.
Experiments related with a possible cerebral plasticity underlying recov-
ery after stroke, have searched for changes in SEF source organization.
Narrowed cortical area harboring the sources of SEFs in patients with first-
ever monohemispheric stroke have been described (Rossini et al., 2003), but
so far, except for some source displacement possibly due to perilesional
edema, no unusual source structure has been demonstrated. Instead, source
strengths and response latencies appear to be correlated with the severity of
the clinical picture (Wikström et al. 1999; Gallien et al. 2003; Oliviero et al.
2005). Enhanced excitability of the affected hemisphere and spared posterior
parietal responses have been linked with high functionality of the affected
hand, whereas the enhanced excitability in the unaffected hemisphere has
been linked with large cortical lesions in the affected hemisphere (Oliviero
et al., 2004).
Use of a single dipole model in analyzing possible plastic changes may
not be as straightforward as it seems. For example, the orderly tonotopic
organization of AEF N100m sources has formed a basis for the investigation
of such phenomena as reorganization of auditory cortex in tinnitus subjects,
or possible anomalies of auditory cortex in schizophrenic patients. However,
a recent carefully executed study has shown that organization of AEF sources
demonstrated no significant frequency dependence at all in most healthy
392 MEG: An Introduction to Methods
Conclusions
MEG is a valuable tool to use in clinical settings. Nevertheless, it is clear that
it is only a part of a multifaceted clinical evaluation deriving information
from all available sources for the benefit of the patient. The relative weight of
MEG in this evaluation depends on individual clinical details of each patient.
The emerging applications are highly exciting, and will provide new opportu-
nities for MEG studies in clinical settings.
Acknowledgments
Supported by the Academy of Finland. E. Kirveskari made valuable comments
on the manuscript.
394 MEG: An Introduction to Methods
References
Ahonen, A., Hämäläinen, M., Kajola, M., Knuutila, J., Laine, P., Lounasmaa, O., et al.
(1993). 122-channel SQUID instrument for investigating the magnetic signals from
human brain. Physica Scripta, T49, 198–205.
American Academy of Neurology Professional Association (AANPA). (2009). Magne-
toencephalography (MEG) Policy. Recommended by the AANPA Medical Economics
and Management Committee. Approved by the AANPA Board of Directors on May 8,
2009. St. Paul, MN: AANPA.
Aoyama, H., Kamada, K., Shirato, H., Takeuchi, F., Kuriki, S., Iwasaki, Y., et al. (2003).
Visualization of the corticospinal tract pathway using magnetic resonance axono-
graphy and magnetoencephalography for stereotactic irradiation planning of arte-
riovenous malformations. Radiother Oncol, 68, 27–32.
Assaf B., Karkar K., Laxer K., Garcia P., Austin E., Barbaro N., et al. (2004). Magne-
toencephalography source localization and surgical outcome in temporal lobe
epilepsy. Clin Neurophysiol, 115, 2066–2076.
Balakrishnan, G., Grover, K. M., Mason, K., Smith, B., Barkley, G., Tepley, N., et al.
(2007). A retrospective analysis of the effect of general anesthetics on the successful
detection of interictal epileptiform activity in magnetoencephalography. (2007)
Anesth Analg, 104, 1493–1497.
Barkley, G., & Baumgartner, C. (2003). MEG and EEG in epilepsy. J Clin Neurophy-
siol, 20, 163–178.
Barth, D., Sutherling, W., Engel Jr., J., & Beatty, J. (1982). Neuromagnetic localization
of epileptiform spike activity in the human brain. Science, 218, 891–894.
Barth, D., Sutherling, W., Engel Jr., J., & Beatty, J. (1984). Neuromagnetic evidence of
spatially distributred sources underlying epileptiform spikes in the human brain.
Science, 223, 293–296.
Bittar, R., Olivier, A., Sadikot, A., Andermann, F., Comeau, R., Cyr, M., et al. (1999)
Localization of somatosensory function by using positron emission tomography
scanning; a comparison with intraoperative cortical stimulation. J Neurosurg, 90,
478–483.
Butz, M., Gross, J., Timmermann, L., Moll, M., Freund, H. & Witte, O. (2004). Peri-
lesional pathological oscillatory activity in the magnetoencepohalogram of patients
with cortical brain lesions. Neurosci Lett, 355, 93–96.
Charmichael, S., & Chesselet, M. (2002). Synchronous neuronal activity is a signal for
axonal sprouting after cortical lesions in the adult. J Neurosci, 22, 6062–6070.
Cheyne, D., Bostan, A., Gaetz, W., & Pang E. W. (2007). Event-related beamforming:
A robust method for presurgical functional mapping using MEG. Clin Neurophy-
siol, 118, 1691–1704.
Cramer, S. (2004). Functional Imaging in stroke recovery. Stroke, 35(Suppl. I), 2695–
2698.
Duffay, H. (2005). Lessons from brain mapping in surgery for low-grade glioma:
insights into associations between tumour and brain plasticity. Lancet Neurol, 4,
476–485.
Duffay, H., Capelle, L., Denvil, D., Sichez, N., Gaticnol, P., Taillandier, L., et al.
(2003). Usefulness of intraoperative electrical subcortical mapping during surgery
for low-grade gliomas located within eloquent brain regions: functional results in a
consecutive series of 103 patients. J Neurosurg, 98, 764–778.
The Use of MEG in Clinical Settings 395
Kamada, K., Todo, T., Masutani, Y., Aoki, S., Ino, K., Morita, A., et al. (2007). Visuali-
zation of the frontotemporal language fibers by tractography combined with func-
tional magnetic resonance imaging and magnetoencephalography. J Neurosurg,
106, 90–98.
Kamada, K., Houkin, K., Takeuchi, F., Ishii. N., Ikeda, J., Sawamura, Y., et al. (2003).
Visualization of the eloquent motor system by integration of MEG, functional, and
anisotropic diffusion-weighted MRI in functional neuronavigation. Surg Neurol,
59, 353–362.
Kamada, K., Saguer, M., Möler, M., Wichlow, K., Katenhauser, M., Kober, H. et al.
(1997). Functional and metabolic analysis of cerebral ischemia using magneto-
encephalography and proton magnetic resonance spectroscopy. Ann Neurol, 42,
554–563.
Kamada, K., Takeuchi, F., Kuriki, S., Oshiro, O., Houkin, K., & Abe, H. (1993). Func-
tional neurosurgical simulation with brain surface magnetic resonance images and
magnetoencephalography. Neurosurgery, 33, 269–273.
Kettenmann, B., Feichtinger, M., Tilz, C., Kaltenhäuser, M., Hummel, C., & Stefan, H.
(2005). Comparison of clonidine to sleep deprivation in the potential to induce
spike or sharp wave activity. Clin Neurophysiol, 116, 905–912.
Kim, J. S., & Chung, C. K. (2007). Robust source analysis of oscillatory motor cortex
activity with inherently variable phase delay. NeuroImage, 37, 518–529.
Kim, J. S., & Chung, C. K. (2008). Language lateralization using MEG beta frequency
desynchronization during auditory oddball stimulation with one-syllable words.
NeuroImage, 42, 1499–1507.
Kirchberger, K., Hummel, C., & Stefan, H. (1998). Postoperative multichannel
magnetoencephalography in patients with recurrent seizures after epilepsy surgery.
Acta Neurol Scand, 98, 1–7.
Kirveskari, E., Salmelin, R., & Hari, R. (2006). Neuromagnetic responses to vowels vs.
tones reveal hemispheric lateralization. Clin Neurophysiol,117, 643–648.
Klöppel, S., & Büchel, C. (2005). Alternatives to the Wada test: a critical view of func-
tional magnetic resonance imaging in preoperative use. Curr Opin Neurol, 18,
418–423.
Knowlton, R. (2006). The role of FDG-PET, ictal SPECT and MEG in the epilepsy
surgery evaluation. Epilepsy & Behav, 8, 91–101.
Knowlton, R., Laxer, K., Aminoff, M., Roberts, T., Wong, S., & Rowley, H. (1997).
Magnetoencephalography in partial epilepsy: clinical yield and localization
accuracy. Ann Neurol, 42, 622–631.
Knowlton, R., & Shih, J. (2004). Magnetoencephalography in epilepsy. Epilepsia,
45(Suppl. 4), 61–71.
Knowlton R., Elgavish R., Howell J., Blount J., Burneo J., Faught E., et al. (2006).
Magnetic source imaging versus intracranial electroencephalogram in epilepsy
surgery: a prospective study. Ann Neurol, 59, 835–842.
Knowlton, R., Radzan, S., Limdi, N., Elgavish, R., Kkillen, J., Blount, J., et al. (2009).
Effect of epilepsy magnetic source imaging on intracranial electrode placement.
Ann Neurol 65: 716–723.
Kober, H., Möller, M., Nimsky, C., Vieth, J., Fahlbusch, R., & Ganslandt, O.
(2001a). New approach to localize speech-relevant brain areas and hemispheric
dominance using spatially filtered magnetoencephalography. Hum Brain Mapp, 14,
236–250.
398 MEG: An Introduction to Methods
Kober, H., Nimsky, C., Moller, M., Hastreiter, P., Fahlbusch, R., & Ganslandt, O.
(2001b). Correlation of sensorimotor activation with functional magnetic resonance
imaging and magnetoencephalography in presurgical functional imaging: a spatial
analysis. NeuroImage, 14, 1214–1228.
Korvenoja, A., Kirveskari, E., Aronen, H., Avikainen, S., Brander, A., Huttunen, J.,
et al. (2006). Localization of primary sensorimotor cortex:comparison of magne-
toencephalography, functional MR imaging and intraoperative cortical mapping.
Radiology, 241, 213–222.
Kringelbach, M., Jenkinson, N., Owen, S. L. F., & Aziz, T. Z. (2007). Translational
principles of deep brain stimulation. Nature Rev Neurosci, 8, 623–635.
Kringelbach, M., Jenkinson, N., Greren A. L., Owen, S. L. F., Hansen, P. C.,
Cornelissen, P. L., et al. (2007). Deep brain stimulation for chronic pain investiga-
ted with magnetoencephalography. NeuroReport,18, 223–228.
Krings, T., Reiniges, M., Erberich, S., Kemeny, S., Rohde, V., Spetzger, U., et al. (2001).
Functional MRI for presurgical planning: problems, artefacts and solution strate-
gies. J Neurol Neurosurg Psych, 70, 749–760.
Lamusuo, S., Forss, N., Ruottinen, H.-M., Bergman, J., Mäkelä, J., Mervaala, E., et al.
(1999). 18-F FDG PET and whole-scalp MEG localization of epileptogenic cortex.
Epilepsia, 40, 921–930.
Lehericy, S., Duffay, H., Cornu, P., Capelle, L., Pidoux, B., Carpentier, A., et al. (2000).
Correspondence between functional magnetic resonance imaging somatotopy and
individual brain anatomy of the central region: comparison with intraoperative
stimulation in patients with brain tumors. J Neurosurg, 92, 589–598.
Lewine, J., Andrews, R., Chez, M., Patil, A., Devinsky, O., Smith, M., et al. (1999)
Magnetoencephalographic patterns of epileptiform activity in children with regres-
sive autism spectrum disorders. Pediatrics, 104, 405–418.
Lewine, J., Davis, J. T., Bigler, ED., Thoma, R., Hill, D., Funke, M., et al. (2007).
Objective documentation of traumatic brain injury subsequent to mild head
trauma:Multimodal brain imaging with MEG, SPECT and MRI. J Head Trauma
Rehabil, 22,141–155.
Lin, P. T., Berger, M. S., & Naragajan, S. S. (2006). Motor field sensitivity for preope-
rative localization of motor cortex. J Neurosurg, 105, 588–594.
Lin, Y.-Y., Chang, K.-P., Hsieh, J.-C., Yeh, T.-C., Hsiang, Y.-Y., Kwan, S.-Y., et al.
(2003). Magnetoencephalographic analysis of bilaterally synchronous discharges in
benign rolandic epilepsy of childhood. Seizure, 12, 448–455.
Linkenkaer-Hansen, K., Monto, S., Rytsälä, H., Suominen, K., Isometsä, E., & Kähkö-
nen, S. (2005). Breakdown of long-range temporal correlations in theta oscillations
in patients with major depressive disorder. J Neurosci, 25, 10131–10137.
Lütkenhöner, B., Krumbholz, K., & Seither-Preisler, A. (2003). Studies of tonotopy
based on wave N100 of the auditory evoked field are problematic. NeuroImage, 19,
935–949.
Maihöfner, C., Handwerker, H., Neundorfer, B., & Birklein, F. (2004). Cortical reor-
ganization during recovery from complex regional pain syndrome. Neurology, 63,
693–701.
Maihöfner, C., Handwerker, H., Neundörfer, B., & Birklein, F. (2003). Patterns
of cortical reorganization in complex regional pain syndrome. Neurology, 61,
1707–1715.
The Use of MEG in Clinical Settings 399
Mamelak, A., Lopez, N., Akhtari, M., & Sutherling, W. (2002) Magnetoencepha-
lography-directed surgery in patients with neocortical epilepsy. J Neurosurg, 97,
865–873.
Merlet, I., Paetau, R., Garcia-Larrea, L., Uutela.K, Granström, M.-L., & Maguiere, F.
(1997) Apparent asynchrony between interictal electric and magnetic spikes. Neu-
roReport, 8, 1071–1076.
Merrified, W. S., Simos, P. G., Papanicolau, A. C., Philpott, LM., & Sutherling, W. W.
(2007). Hemispheric language dominance in magnetoencephalography: Sensitivity,
specificity and data reduction techniques. Epilepsy Behav, 10, 120–128.
Mikuni, N., Nagamine, T., Ikeda, A., Terada, K., Taki, W., Kimura, J., et al. (1997).
Simultaneous recording of epileptiform discharges by MEG and subdural electro-
des in temporal lobe epilepsy. NeuroImage, 5, 298–306.
Modena, I., Ricci, G., Barbanera, S., Romani, G., & Carelli, P. (1982). Biomagnetic
measurements of spontaneous brain activity in epileptic patients. Electroenceph clin
Neurophysiol, 54, 622–628.
Mogilner, A., Grossman, A., Ribary, U., Joliot, M., Volkmann, J., Rapaport, D., et al.
(1993). Somatosensory cortical plasticity in adult humans revealed by magnetoen-
cephalography. Proc Natl Acad Sci USA, 90, 3593–3597.
Mäkelä J.P., Forss N., Jääskeläinen J., Kirveskari E., Korvenoja A., & Paetau R. (2006).
Magnetoencephalography in neurosurgery. Neurosurgery, 59, 493–510.
Mäkelä, J.P., Kirveskari, E., Seppä, M., Hämäläinen, M., Forss, N., Avikainen, S.,
et al. (2001). Three-dimensional integration of brain anatomy and function to faci-
litate intraoperative navigation around the sensorimotor strip. Hum Brain Mapp,
12, 180–192.
Mäkelä, J. P., Salmelin, R., Hokkanen, L., Launes, J., & Hari, R. (1998a). Neuroma-
gnetic sequelae of herpes simplex encephalitis. Electroenceph clin Neurophysiol, 106,
251–258.
Mäkelä, J. P., Salmelin, R., Kotila, M., Salonen, O., Laaksonen, R., Hokkanen, L., et al.
(1998b). Left thalamic infarctions modify neuromagnetic cortical signals. Electro-
enceph clin Neurophysiol, 106, 433–443.
Mäkelä, J.P., Taulu, S., Ahonen, A., Pohjola, J., & Pekkonen, E. (2007). Effects of
subthalamic nucleus stimulation on spontaneous sensorimotor MEG activity in a
Parkinsonian patient. International Congress Series, 1300, 345–348.
Mäkelä, J. P., Iivanainen, M., Pieninkeroinen, I. P., Waltimo, O., Lahdensuu, M.
(1993). Seizures associated with propofol anesthesia. Epilepsia, 34, 832–835.
Nakasato, N., Kumabe, T., Kanno, A., Ohtomo, S., Mizoi, K., & Yoshimoto, T. (1997).
Neuromagnetic evaluation of cortical auditory function in patients with temporal
lobe tumors. J Neurosurg, 86, 610–618.
Oliviero, A., Tecchio, F., Zappasodi, F., Pasqualetti, P., Salustri, C., Lupoi, D., et al.
(2004). Brain sensorimotor hand area functionality in acute stroke: insights from
magnetoencephalography. NeuroImage, 23, 542–550.
Osipova, D., Ahveninen, J., Jensen, O., Ylikoski, A., & Pekkonen, E. (2005). Altered
generation of spontaneous oscillations in Alzheimer’s disease. NeuroImage, 27,
835–8411.
Otsubo H., Ochi A., & Elliot, I. (2001). MEG predicts epileptic zone in
lesional extrahippocampal epilepsy: 12 pediatric surgery cases. Epilepsia, 42,
1523–1530.
400 MEG: An Introduction to Methods
Paetau, R., Granström, M.-L., Blomstedt, G., Jousmäki, V., Korkman, M., & Liukko-
nen, E. (1999). Magnetoencephalography in presurgical evaluation of children with
the Landau-Kleffner syndrome. Epilepsia, 40, 326–335.
Palva, J., Palva, S., & Kaila, K. (2005). Phase synchrony among neuronal oscillations
in the human cortex. J Neurosci, 25, 3962–3972.
Papanicolaou, A., Simos, P., Castillo, E., Breier, J., Sarkari, S., Pataraia, E., et al. (2004).
Magnetoencephalography: a noninvasive alternative to the Wada procedure. J Neu-
rosurg, 100, 867–876.
Papanicolau A., Pataraia E., Billingsley-Marshall R., Castillo E., Wheless J., Swank P.,
et al. (2005). Toward the substitution of invasive electroencephalography in
epilepsy surgery. J Clin Neurophysiol, 22, 231–237.
Parkkonen, L., Fujiki, N., & Mäkelä, JP. (2009). Sources of auditory brainstem responses
revisited: Contribution by magnetoencephalography. Human Brain Mapping 30,
1772–1782.
Parviainen, T., Helenius, P., & Salmelin, R. (2005). Cortical differentiation of speech
and nonspeech sounds at 100 ms: implications for dyslexia. Cerebral Cortex, 15,
1054–1063.
Pataraia, E., Baumgartner, C., Lidinger, G., & Deecke, L. (2002) Magnetoencephalo-
graphy in presurgical epilepsy evaluation. Neurosurg, Rev. 25, 141–159.
Pujol, J., Deus, J., Acebes, JJ., Villanueva, A., Aparicio, A., Soriano-Mas, C.,
et al. (2008). Identification of the sensorimotor cortex with functional MRI:
frequency and actual contribution in a neurosurgical context. J Neuroimaging, 18,
28–33.
Ramantani, G., Boor, R., Paetau, R., Ille, N., Feneberg, R., Rupp, A., et al. (2006).
MEG versus EEG: influence of background activity on interictal spike detection.
J Clin Neurophysiol, 23, 498–508.
Rezai, A., Hund, M., Kronberg, E., Zoneshayn, M., Cappell, J., Ribary, U., et al. (1996).
The interactive use of magnetoencephalography in stereotactic image-guided
neurosurgery. Neurosurgery, 39, 92–102.
Roberts, D., Hartov, A., Kennedy, F., Miga, M., & Paulsen, K. (1998). Intraoperative
brain shift and deformation: a quantitative analysis of cortical displacement in 28
cases. Neurosurgery, 42, 749–760.
Rossini, P., Calautti, C., Pauri, F., & Baron, J.-C. (2003) Post-stroke reorganization in
the adult brain. Lancet Neurol, 2, 493–502.
Salenius, S., Portin, K., Kajola, M., Salmelin, R., & Hari, R. (1997). Cortical control
of human motoneuron firing during isometric contractions. J Neurophysiol, 77,
3401–3405.
Salmelin, R., Hari, R., Lounasmaa, O. V., & Sams, M. (1994). Dynamics of brain
activation during picture naming. Nature, 368, 463–465.
Salmelin, R. (2007). Clinical neurophysiology of language: the MEG approach. Clin
Neurophysiol, 118, 237–254.
Schiffbauer, H., Berger, M., Ferrari, P., Freudenstein, D., Rowley, H., & Roberts, T.
(2002). Preoperative magnetic source imaging for brain tumor surgery: a quantita-
tive comparison with intraoperative sensory and motor mapping. J Neurosurg, 97,
1333–1342.
Schnitzler, A., & Gross, J. (2005). Normal and pathological oscillatory communication
in the brain. Nature Neurosci, 6, 285–296.
The Use of MEG in Clinical Settings 401
Shiraishi, H., Watanabe, Y., Watanabe, M., Inoue, Y., Fujiwara, T., & Yagi K. (2001).
Interictal and ictal magnetoencephalographic study in patients with medial frontal
lobe epilepsy. Epilepsia, 42, 875–882.
Stefan, H., Hummel, C., Scheler, G., Druschky, K., Tilz, C., Kaltenhäuser, M., et al.
(2003). Magnetic brain source imaging of focal epileptic activity: a synopsis of 455
cases. Brain, 126, 2396–2405.
Sutherling, W., Crandall, P., Darcey, T., Becker, D., Levesque, M., & Barth, D. (1988).
The magnetic and electric fields agree with intracranial localizations of somatosen-
sory cortex. Neurology 38, 1705–1714.
Sutherling, W., Mamelak, A., Thyerlei, D., Maleeva, T., Minazad, Y., Philpott, L.,
et al. (2008) Influence of magnetic source imaging for planning intracranial EEG in
epilepsy. Neurology 71, 990–996.
Schwartz, E. S., Dlugos, D. L., Storm, P. B., Dell, J., Magee, R., Flynn, T. P., et al.
(2008). Magnetoencephalography for pediatric epilepsy: how we do it. Am J
Neuroradiol 29(5), 832–837.
Szmuk, P., Kee, S., Pivalizza, E. G., Warters, R., Abramson, D. C., & Ezri, T. (2003).
Anaesthesia for magnetoencephalography in children with intractable seizures.
Paediatric Anaesth, 13, 811–817.
Szymanski, M., Perry, D., Cage, N. M., Rowley, H., Walker, J., Berger, M., et al. (2001).
Magnetic source imaging of late evoked field responses to vowels: toward an assess-
ment of hemispheric dominance for language. J Neurosurg, 94, 445–453.
Tang, L., Mantle, M., Ferrari, P., Schiffbauer, H., Rowley, A. A., Barbaro, N. M., et al.
(2003). Consistency of interictal and ictal onset localization using magnetoence-
phalography in patients with partial epilepsy. J Neurosurg, 98, 837–845.
Tanaka, N., Thiele, E., Madsen, J., Bourgeois, B., & Stufflebeam, S. (2009). Magneto-
encephalographic analysis in patients with vagus nerve stimulator. Pediatr Neurol
41, 383–387.
Tao, J. X., Ray, A., Hawes-Ebersole, S., & Ebersole, J. S. (2005) Intracranial EEG sub-
strates of scalp EEG interictal spikes. Epilepsia, 46, 669–676.
Taulu, S., Simola, J., & Kajola, M. (2005). Applications of the signal space separation
method. IEEE Trans Sign Proc., 53, 3359–3372.
Taulu, S., & Simola, J. (2006), Spatiotemporal signal space separation method for
rejecting nearby interference in MEG measurements. Phys Med Biol, 51,
1759–1768.
Tecchio, F., Zappasodi, F., Pasqualetti, P., Tombini, M., Salustri, C., Oliviero, A.,
et al. (2005). Rhythmic brain activity at rest from rolandic areas in acute mono-
hemispheric stroke: a magnetoencephalographic study. NeuroImage, 28, 72–83.
Tilz, C., Hummel, C., Kettenmann, B., & Stefan, H. (2002). Ictal onset localization of
epileptic seizures by magnetoencephalography. Acta Neurol Scand, 106, 190–195.
Uutela, K., Taulu, S., & Hämäläinen, M. (2001). Detecting and correcting for head
movements in neuromagnetic measurements. NeuroImage, 14, 1424–1431.
Vates, E., Lawton, M., Wilson, C., McDermott, M., Halbach, V., Roberts, T., et al.
(2002). Magnetic source imaging demonstrates altered cortical distribution of func-
tion in patients with arteriovenous malformations. Neurosurgery, 51, 614–622.
Vitikainen A-M., Lioumis P., Metsähonkala L., Salli E., Komssi S., Kicic D., et al. (2009).
Combined use of non-invasive techniques for improved functional localization for
a selected group of epilepsy surgery candidates. NeuroImage 45, 342–348.
402 MEG: An Introduction to Methods
Vrba, J., Betts, K., Burbank, M., Cheung, T., Fife, A., Haid, G., et al. (1993).
Whole cortex 64 channel SQUID biomagnetometer system. IEEE Trans Appl Sup.
3, 1878–1882.
Worrell, G. A., Parish, L., Cranstoun, S. D., Jonas, R., Baltuch G., & Litt, B. (2004).
High-frequency oscillations and seizure generation in neocortical epilepsy. Brain,
127, 1496–1506.
Yu, H., Nakasato, N., Iwasaki, M., Shamoto, H., Nagamatsu, K., & Yoshimoto, T.
(2004) Neuromagnetic separation of secondarily bilateral synchronized spike foci:
report of three cases. J Clin Neurosci, 11, 644–648.
15
Using Magnetoencephalography to Elucidate
the Principles of Deep Brain Stimulation
Morten L. Kringelbach, Peter C. Hansen,
Alex L. Green, and Tipu Z. Aziz
403
404 MEG: An Introduction to Methods
promise, the underlying neural mechanisms for DBS are not understood, and
in particular it is not at all clear how DBS of specific brain targets changes the
neural activity in wider cortical and subcortical regions.
DBS offers a novel and unique possibility for in vivo investigation of the
functional role of the underlying neural circuitry in humans, by using stimu-
lation parameters which yield different clinical results, and by switching
the stimulator on and off. The ensuing changes in whole-brain activity can
then be mapped using neuroimaging methods. Yet, some of the most-used
neuroimaging technologies—functional magnetic resonance imaging (fMRI)
and positron tomography (PET)—are, due to their intrinsic properties, less
ideal for this purpose.
In contrast, MEG is of a noninvasive nature, and with its high spatial and
temporal resolution holds great promise for elucidating the underlying whole-
brain neural mechanisms of DBS by, for example, measuring oscillatory
communication between brain regions (Schnitzler & Gross, 2005).
Here, we first provide an introductory overview of the current state-
of-art of DBS, and the previous use of neuroimaging techniques with DBS. We
then describe the methods and results of using MEG to measure both low- and
high-frequency stimulation. We discuss the importance of the findings, as well as
potential confounds and future possibilities of combining MEG and DBS.
The most efficacious targets for treating movement disorders with DBS have
been the structures in the basal ganglia. The internal globus pallidus (GPi)
and subthalamic nucleus (STN) have been demonstrated as safe and effica-
cious targets for Parkinson’s disease (Aziz et al., 1991; Bergman et al., 1990).
Long-term benefits of using high-frequency (130-185Hz) DBS for Parkinson’s
disease are well documented (Bittar et al., 2005a; Krack et al., 2003). Substantial
improvements in the symptoms of Parkinson’s disease (as measured by motor
and daily living scores) (Fahn & Elton, 1987), as well as reductions in the
patients’ level of medication for Parkinson’s disease, have been found in
extensive DBS trials for Parkinson’s disease (Benabid et al., 1996; Krack et al.,
2003; Siegfried & Lippitz, 1994). Recently, translational research has identi-
fied the brainstem pedunculopontine nucleus (PPN) as a potential new
Parkinson’s disease target in monkeys (Jenkinson et al., 2005; Jenkinson et al.,
2004; Jenkinson et al., 2006; Nandi et al., 2002) and humans (Mazzone et al.,
2005; Plaha & Gill, 2005; Stefani et al., 2007).
The preferred target for dystonia and spasmodic torticollis is the GPi
(Bittar et al., 2005c; Kumar et al., 1999). The DBS parameters for dystonia
differ from Parkinson’s disease, with a broader pulse width (200–400 μs) and
higher voltage (typically between 2.2–7V) (Krauss et al., 2004), leading to
rapid battery consumption. Blinded, controlled GPi trials have shown 30–50%
improvements in patients over 12 months (Vidailhet et al., 2005).
Essential tremor is usually treated with DBS in the ventral intermediate
nucleus of the thalamus (VIM) (Hassler, 1955; Lenz et al., 1994), while the DBS
target for Parkinson’s disease tremor is the STN (Krack et al., 1997). Long-term
effects of DBS in VIM have shown an average tremor reduction of over 80% in the
majority of patients (Koller et al., 1999; Rehncrona et al., 2003). Thalamic DBS was
found to significantly improve tremor compared to thalamotomy, and have fewer
adverse effects (Schuurman et al., 2000). A large multicenter study showed contin-
ued improvements in tremor ratings in patients with essential tremor after six
years of follow-up (Sydow et al., 2003).
DBS for chronic pain has been used for over fifty years since the initial studies
using DBS in the hypothalamus (Pool et al., 1956). More recent efficacious
targets are in the thalamus (Hosobuchi et al., 1973; Mazars et al., 1973; Mazars
et al., 1960) and periventricular-periaqueductal gray region (PVG/PAG)
(Hosobuchi et al., 1977; Richardson & Akil, 1977a; b; c). Following two failed
clinical trials (Coffey, 2001), FDA approval was not sought by device manu-
facturers. During the last decade only five centers outside the U.S. have pro-
duced case series of more than six patients: (Bittar et al., 2005b; Green et al.,
2006; Hamani et al., 2006; Krauss et al., 2002; Marchand et al., 2003; Nandi
et al., 2003; Owen et al., 2006a; Owen et al., 2006b; Tronnier, 2003). These
406 MEG: An Introduction to Methods
The specific methods used for DBS vary among neurosurgical teams. Here,
we present the methods adopted in Professor Aziz’s lab in Oxford and focus
specifically on the procedures used for DBS for pain relief (see Figure 15–1).
A T1-weighted MRI scan of each patient’s brain is performed several
weeks before surgery. For surgery, a Cosman-Roberts-Wells base ring is
applied to the patient’s head under local anesthesia. A stereotactic computed
tomography (CT) scan is then performed and, using the Radionics Image
Fusion® and Stereoplan® (Integra Radionics, Burlington, MA) program, the
coordinates for the PVG/PAG and ventro-posterior lateral thalamus (VPL)
are calculated. A double-oblique trajectory is used, with an entry point just
anterior to the coronal suture, and laterality of approach dictated by ventricu-
lar width. The PVG/PAG target is proximally located 2–3 mm lateral to the
wall of the third ventricle and 2 mm anterior to the level of the posterior com-
missure, and distally, the deepest electrode is placed in the superior colliculus.
The VPL is located 12 mm lateral and 5–8 mm posterior to the mid-commis-
sural point, at the depth of the anterior/posterior commissure plane. After
washing the patient’s scalp with alcoholic chlorhexidine, a parasaggital poste-
rior frontal scalp incision 3.0 cm from the midline is made, contralateral to
the side of pain.
The VPL is usually implanted with a Medtronic 3387 (Medtronic, Minneapolis,
MN) electrode, where stimulation induces parasthesia in the area of pain. The
PVG is also implanted with a Medtronic 3387 electrode where stimulation induces
relief of pain or a sensation of warmth in the area of pain. The deepest electrode
is noted to be in a satisfactory position if eye bobbing is induced at intensity of
stimulation at least twice that required for sensory effects. The electrodes are fixed
to the skull with a miniplate prior to externalization. In most patients, the
electrodes are externalized for a week of trial stimulation.
Pain is assessed before surgery and during stimulation by a self-rated
visual analog scale. If the patients are satisfied with the degree of pain relief, full
implantation of a Medtronic pulse generator is performed in the following
week under general anesthesia.
Figure 15–1. The neurosurgical procedures involved in DBS. (A) Schematic
of the principles of DBS. (B) Illustration of the process of the neurosurgi-
cal pre-planning. (C) Application of the CRW stereotactic head frame on
the patient. Note that the base ring is parallel to the orbitomeatal line.
(D) The precise positioning of the electrode through perforating the cal-
varium with a twist drill. (E) Securing the electrode to the skull with a
titanium miniplate and screws. (F) Placement of the implantable pulse
generator in a subcutaneous pectoral pouch.
407
408 MEG: An Introduction to Methods
The safety of the DBS procedure has been demonstrated in many worldwide
trials, and in the longterm follow-up in DBS for the treatment of chronic pain
(Hosobuchi, 1986). The long-term efficacy of DBS depends on the genera-
tors, where most will last around 3–5 years depending on the current demands
of the pulse protocol, although in the case of dystonia, this can be less than one
year. Radiofrequency rechargeable pulse generators are available for spinal
cord generators, and are being trialed for DBS.
Stereotactic procedures always carry a significant risk, and can lead to
intracranial bleeding, usually in around 2.0–2.5% of DBS implants (Benabid
et al., 1996; Beric et al., 2001). Other potential risks include hardware-related
complications such as dislocation, lead fracture, and infection (6%). The
infection rate is equal to that of other surgical procedures, but may necessitate
explantation of the stimulator (Hariz, 2002). Stimulation-induced side effects
(3%) are also quite common, such as aggression (Bejjani et al., 2002), mirth-
ful laughter (Krack et al., 2001), depression (Bejjani et al., 1999), penile
erection (Temel et al., 2004) and mania (Kulisevsky et al., 2002).
Neuroimaging methods such as PET and fMRI can measure indirect changes
of neural activity such as blood flow, blood oxygenation and glucose consump-
tion. It is presently not entirely clear how well these indirect measurements
correlate with various aspects of neural activity, but some progress in our
understanding has been made under normal physiological conditions
(Lauritzen, 2005; Logothetis & Wandell, 2004). In addition, it is important to
realize that these methods entail a number of assumptions that may or may not
prove to be important for interpreting the subsequent results.
It has become clear that fMRI studies pose a large degree of risk to DBS
patients, since the large magnitude of the magnetic fields will interfere with
active pulse generators and DBS electrodes. While several studies have been
published showing the feasibility of using fMRI of DBS (Rezai et al., 1999; Uitti
et al., 2002), there may well be significant problems using the BOLD signal as a
measurement, since near-infrared spectroscopy showed considerable variations
in blood oxygenation in the frontal cortex, following GPi and thalamic stimula-
tion (Sakatani et al., 1999). Another study has shown that extreme caution
Using MEG to Elucidate the Principles of DBS 409
must be exercised when studying fMRI with DBS, since strong heating, high
induced voltage, and even sparking at defects in the connecting cable have been
observed (Georgi et al., 2004). A case report has, however, shown that fMRI can
be used to study STN stimulation in a patient with Parkinson’s disease, who
showed increases in the BOLD signal in primary motor areas, and decreases in
supplementary motor areas, during stimulation (Stefurak et al., 2003).
PET is comparably safe, although not without health risks due to the ion-
ized radiation, and has been used for measuring the effects of DBS. Due to the
long acquisition periods for PET, usually up to a full minute, investigators
have to carefully address the potential movement artifacts when studying
movement disorders.
Using PET for Parkinson’s disease is therefore rather challenging, and
requires careful observation of any movement and removal of potentially
confounding scans. One PET study took such precautions while scanning 13
Parkinson’s disease patients and showed that STN stimulation led to increased
blood flow in the thalamus, GP and midbrain (including STN), and reduced
blood flow in frontal parietal and temporal cortices (Hershey et al., 2003).
Similarly, a PET study of VIM thalamic stimulation in patients with essential
tremor showed increases in blood flow in the thalamus and the cortical tar-
gets of thalamic output (Perlmutter et al., 2002). Other studies did not moni-
tor and consider the behavioral effects of patients during the PET scans, and
must therefore be cautiously interpreted (Fukuda et al., 2001; Hilker et al.,
2004). Taken together, however, these results suggest that STN stimulation
increases rather than inhibits the activity of STN output neurons, which in
turn leads to increases of inhibition of thalamocortical projections, with
subsequent decreases in blood flow in cortical regions.
Using PET to study DBS for affective disorders is less challenging in terms
of potential movement artifacts. One PET study investigated the effects of
hypothalamic stimulation for cluster headache in 10 patients (May et al.,
2006). Stimulation compared with no stimulation elicited significant increases
in activity in the ipsilateral hypothalamic gray (at the site of the stimulator
tip), as well as in structures in the known pain processing network, including
the ipsilateral thalamus, somatosensory cortex and praecuneus, the anterior
cingulate cortex, and in the ipsilateral trigeminal nucleus and ganglion.
Decreases in activity were found in the posterior cingulate cortex, middle
temporal gyrus, and contralateral anterior insula. The results suggest that
hypothalamic stimulation for cluster headache functions mainly through
modulating the pain processing network.
Another PET study, using stimulation of subgenual cingulate cortex for
treatment-resistant depression in four patients, showed marked reduction in
activity in cortical and subcortical areas (Mayberg et al., 2005). The results
are harder to interpret, given the small numbers of patients and the paucity
of knowledge about the brain structures involved in depression, but again
suggest that the mode of functioning of DBS would appear to be one of
modulating an existing network of interacting brain regions.
410 MEG: An Introduction to Methods
The 58-year old right-handed male patient, RM, was referred with a 4-year
history of severe phantom limb pain in the left leg, stemming from fracturing
his leg in May 2001 with subsequent complications including an MRSA infec-
tion culminating in an above-the-knee amputation in October 2001.
Sympathectomy, spinal cord nerve stimulation, hypnosis and a wide variety
of medications had provided little relief. Preoperative testing showed an
abnormal neuropsychological profile, with poor performance on all verbally
mediated tests and at a level of “caseness” for both anxiety and depression,
probably linked to his level of medication, which was repeated prescriptions
of Morphine sulphate 380 mgm over 24 hours.
The patient RM was then implanted with a DBS in the right PVG/PAG,
and he experienced excellent pain relief. The patient later fell, fracturing the
deep-brain electrode, and this caused immediate return of the pain, After
surgical revision, pain relief returned. Effective settings for stimulation in RM
were 1.5 volts, frequency, 7 Hz and pulse width, 300 μs. This has significantly
decreased the level of chronic pain in the patient to a manageable level, reducing
the patient’s McGill pain score by 74%.
The second patient was a 56-year old male, YY, with an 11-year history of
cluster-headache attacks. The headaches had a seasonal pattern starting in
September or October every year, and occurred 3–4 times a day, lasting
for 45 minutes on average. The pain originated over the right forehead and
radiated to the ipsilateral vertex, and was associated with lacrimation and
excess rhinorrhea.
Using MEG to Elucidate the Principles of DBS 411
Figure 15–2. DBS for chronic pain. (A) Axial MRI slice showing the implan-
tation of electrodes in PVG/PAG and thalamus in a patient. (B) Schematic
illustration of the vertical placement of electrodes in the PVG/PAG in a
series of chronic pain patients. (C) Three-dimensional rendering of human
brain showing the placement of the two electrodes in the PVG/PAG and
thalamus, as well as some of the important subcortical structures. (D)
Three-dimensional rendering showing the whole-brain DBS induced
activity from stimulation in the PVG/PAG. (E) The connectivity of the PVG/
PAG measured with diffusion tensor imaging.
Experimental Setup
Patient RM was initially scanned with MEG for 10 minutes, with DBS switched
on while resting. He was asked to continuously report his subjective experi-
ence of pain using a visual rating scale from 0–9 (where 0 is “not painful” and
9 is “very painful”) every 20 seconds. This was then repeated three times
with the stimulator switched off, during which his pain scores increased
with time.
Patient YY’s stimulator was turned off 30 minutes prior to scanning. We
attached electrodes to his forearm in order to measure EMG. He was then
scanned for 10 minutes. He rated his pain state in the same way as RM.
The patient’s stimulator was then turned on and set to the non-effective
setting of 7 Hz. After a further 5 minutes we began the second 10-minute scan
with the rating task. The protocol was then repeated a final time, with the
stimulator set at the effective frequency of 180 Hz.
Data Acquisition
MEG data were collected using a 275-channel CTF Omega system (CTF
Systems Inc., Port Coquitlam, Canada) at Aston University. Data were sampled
at 2400 Hz with an antialiasing cut-off filter of 200 Hz. The patient directly
viewed the visual pain rating scale on a computer monitor.
Before and after surgery the patients were scanned with MRI to get
high-resolution T1 volumes with 1x1x1 mm voxel dimensions. Immediately
after finishing data acquisition in the MEG laboratory, the head coils were
registered to the patients’ MRIs using a 3D digitizer (Polhemus Fastrack,
Polhemus Corporation, Colchester, VT, U.S.) to digitize the shape of the par-
ticipant’s head relative to the position of the head coils with respect to the
nasion, left and right ear on the headset.
Image Analysis
The MEG data were analyzed using Synthetic Aperture Magnetometry (SAM),
which is an adaptive beamforming technique for the analysis of EEG and
MEG data (Vrba & Robinson, 2001). SAM has been previously used in a vari-
ety of studies on the functions of the motor cortex (Taniguchi et al., 2000),
the human somatosensory cortex (Hirata et al., 2002), and visual word recog-
nition (Pammer et al., 2004). In addition, SAM has been shown to be able to
unveil changes in cortical synchronization that are spatially coincident with
the hemodynamic response found with fMRI (Singh et al., 2002).
The SAM analysis links each voxel in the brain to the detection array,
using an optimal spatial filter for that particular voxel (Robinson & Vrba,
1999). The data from the MEG sensors is then projected through this spatial
filter to give a weighted measure of current density, as a function of time, in the
target voxel—which means that the time series for each voxel has the same
Using MEG to Elucidate the Principles of DBS 413
Neuroimaging Data
In the pain condition when DBS was switched off, and when RM’s subjective
pain ratings were significantly higher than during the fourth period after the
DBS had been switched on, significant activity was found in regions involved
in the pain network that have previously been identified using fMRI and PET
(see Figure 15–3 and Table 15–1 for list of activations). These regions included
the insula, and the primary and secondary somatosensory, lateral orbitofron-
tal, and anterior cingulate cortices. Activity was also found in the motor net-
works related to the rating process. In the 10–20 Hz frequency band significant
differences in activity were found in somatosensory cortices (SI, SII), intrapa-
rietal cortex, motor cortex, premotor area (PMA), middle and posterior
414 MEG: An Introduction to Methods
insula cortices, occipital lobe, and middle frontal gyrus. In the 20–30Hz band,
significant differences were found in motor, parietal, insula, fusiform and
motor cingulate cortices, as well as lateral orbitofrontal cortex/anterior insula.
In the 30–60Hz band we found significant differences in the parahippocam-
pal, motor and fusiform cortices.
In the pain relief condition, during the fourth period after the DBS had
been switched on, significant activity was found in brain regions previously
identified with fMRI as the pain relief network (see Table 15–1 for list of
activations). In the 10–20Hz band, significant differences in activity were
found in SI, brainstem, the mid-anterior orbitofrontal cortex and subgenual
cingulate cortex. In the 20–30Hz band there were differences in activity in a
motor network comprising supplementary motor area (SMA), parietal and
Figure 15–3. Brain activity when the patient reported subjective pain relief
(DBS on) and pain (DBS off). Top part of figure shows that in the pain
relief condition there was significant activity in the left mid-anterior orb-
itofrontal cortex and right subgenual cingulate cortex. Activity in these
regions was not found in the pain condition (bottom part of figure).
Using MEG to Elucidate the Principles of DBS 415
Activations are significant at t>2.3, uncorrected; unless indicated with * for a priori predicted
regions. All brain coordinates are in the standard space of MNI (Montreal Neurological Institute).
416 MEG: An Introduction to Methods
The brain regions active in patient YY measured with MEG in all conditions
regardless of stimulator activity, which were very similar to those found in
patient RM. In all conditions somatosensory and motor cortices were acti-
vated in the active (button-press) condition, compared with passive (100ms
before the button press). We also found significant activity in the 10–20Hz
frequency band in the PAG only when the patient’s stimulator was turned off.
When the stimulator was turned on, there was activity in frontal brain regions
including the orbitofrontal cortex, associated using fMRI with the pain relief
network.
Discussion
These studies show that it is feasible to use MEG to map whole-brain changes
in neural activity induced by both low- and high-frequency DBS. We found
significant changes in brain activity in patient RM with implanted low-fre-
quency DBS in the right PVG/PAG, for severe phantom limb pain in the left
leg. This patient reported significantly more pain relatively soon after the DBS
was turned off compared to when DBS was switched on.
Similarly, we were able to measure the whole-brain changes during low-
and high-frequency stimulation of the posterior hypothalamus for cluster-
headache, on MEG recordings. The posterior hypothalamus contains several
neurochemically distinct cell groups. One of these is the Hypocr/Orx neurons
that are activated by nociceptive stimuli and reach structures involved in
nociceptive relay and modulation, including the PAG (Baldo et al., 2003). We
found activity in the PAG only when patient YY’s stimulator was turned off,
which may be related to these patterns of connectivity. In an fMRI study it
was found that PAG was activated if participants were anticipating a painful
stimulus, even before they were subjected to pain (Fairhurst et al., 2007).
Patient YY in the present study was aware that his stimulator had been turned
off, and thus the activity in his PAG may have been related to the anticipation
of his pain returning.
The obtained changes in brain activity related to both pain and pain relief
are consistent with previous findings reported with fMRI and PET (Petrovic
& Ingvar, 2002). Testifying to the utility of our method, irrespective of DBS
we also found activity in brain regions that are part of the motor network,
related to the patient using button-presses to rate his pain.
Pain-specific activity was found when DBS was switched off, and the sub-
jective pain scores were significantly higher than in the pain relief condition
Using MEG to Elucidate the Principles of DBS 417
Potential Problems
Future Possibilities
While the MEG results presented here have been obtained from DBS elec-
trodes with both low- and high-frequency stimulation, the experiments were
conducted in patients who did not move in the scanner. It would be of great
interest to extend these studies to movement disorders such as Parkinson’s
disease, dystonia and essential tremor—but as mentioned above, great care
will have to be taken to minimize head movements in these patients. Another
exciting possibility will be to use the recordings from DBS electrodes (while
still externalized) as an EEG channel, which can then be used in the subsequent
analysis together with the MEG signals.
418 MEG: An Introduction to Methods
Conclusion
The results of these studies have demonstrated the feasibility of using MEG to
map changes in whole-brain activity induced by both low- and high-frequency
DBS. The findings have highlighted the mid-anterior orbitofrontal cortex as
potential future candidate for DBS in patients with chronic pain. From a systems
neuroscience point of view, the combination of the causal, interventional
nature of DBS with the high temporal resolution of MEG, has the makings of
a powerful and sophisticated tool for unraveling the fundamental mechanisms
of normal human brain function.
References
Ackermans, L., Temel, Y., Cath, D., van der Linden, C., Bruggeman, R., Kleijer, M.,
et al. (2006). Deep brain stimulation in Tourette’s syndrome: two targets? Mov
Disor, 21, 709–713.
Anderson, W. S., & Lenz, F. A. (2006). Surgery insight: Deep brain stimulation for
movement disorders. Nat Clin Pract Neurol, 2, 310–320.
Andy, O. J., & Jurko, F. (1987). Thalamic stimulation effects on reactive depression.
Appl Neurophysiol, 50, 324–329.
Aziz, T. Z., Peggs, D., Sambrook, M. A., & Crossman, A. R. (1991). Lesion of the subtha-
lamic nucleus for the alleviation of 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine
(MPTP)-induced parkinsonism in the primate. Mov Disord, 6, 288–292.
Baldo, B. A., Daniel, R. A., Berridge, C. W., & Kelley, A. E. (2003). Overlapping
distributions of orexin/hypocretin- and dopamine-beta-hydroxylase immunoreac-
tive fibers in rat brain regions mediating arousal, motivation, and stress. J Comp
Neurol, 464, 220–237.
Bejjani, B. P., Damier, P., Arnulf, I., Thivard, L., Bonnet, A. M., Dormont, D., et al.
(1999). Transient acute depression induced by high-frequency deep-brain stimula-
tion. N Engl J Med, 340, 1476–1480.
Bejjani, B. P., Houeto, J. L., Hariz, M., Yelnik, J., Mesnage, V., Bonnet, A. M., et al.
(2002). Aggressive behavior induced by intraoperative stimulation in the triangle
of Sano. Neurology, 59, 1425–1427.
Benabid, A. L., Pollak, P., Gao, D., Hoffmann, D., Limousin, P., Gay, E., et al. (1996).
Chronic electrical stimulation of the ventralis intermedius nucleus of the thalamus as
a treatment of movement disorders. J Neurosurg, 84, 203–214.
Bergman, H., Wichmann, T., & DeLong, M. R. (1990). Reversal of experimental
parkinsonism by lesions of the subthalamic nucleus. Science, 249, 1436–1438.
Beric, A., Kelly, P. J., Rezai, A., Sterio, D., Mogilner, A., Zonenshayn, M., et al. (2001).
Complications of deep brain stimulation surgery. Stereotact Funct Neurosurg, 77,
73–78.
Bittar, R. G., Burn, S. C., Bain, P. G., Owen, S. L., Joint, C., Shlugman, D., et al.
(2005a). Deep brain stimulation for movement disorders and pain. J Clin Neurosci,
12, 457–463.
Bittar, R. G., Otero, S., Carter, H., & Aziz, T. Z. (2005b). Deep brain stimulation for
phantom limb pain. J Clin Neurosci, 12, 399–404.
Using MEG to Elucidate the Principles of DBS 419
Bittar, R. G., Yianni, J., Wang, S., Liu, X., Nandi, D., Joint, C., et al. (2005c).
Deep brain stimulation for generalised dystonia and spasmodic torticollis. J Clin
Neurosci, 12, 12–16.
Coffey, R. J. (2001). Deep brain stimulation for chronic pain: results of two multicen-
ter trials and a structured review. Pain Med, 2, 183–192.
Fahn, S., & Elton, R. L. (1987). Members of the UPDRS Development Committee.
Unified Parkinson’s disease rating scale. In: Fahn, S., Marsden, C. D., Goldstein, M.,
& Calne, D. B. (eds.). Recent developments in Parkinson’s disease. Vol. 2. (p. 153).
Florham Park, New York: MacMillan.
Fairhurst, M., Wiech, K., Dunckley, P., & Tracey, I. (2007). Anticipatory brainstem
activity predicts neural processing of pain in humans. Pain, 128, 101–110.
Franzini, A., Ferroli, P., Leone, M., & Broggi, G. (2003). Stimulation of the posterior
hypothalamus for treatment of chronic intractable cluster headaches: first reported
series. Neurosurgery, 52, 1095–1099; discussion 1099–1101.
Fritsch, G., & Hitzig, E. (1870). Über die elektrische Erregbarkeit des Grosshirns. Arch.
Anat. Physiol, 37, 300–332.
Fukuda, M., Mentis, M. J., Ma, Y., Dhawan, V., Antonini, A., Lang, A. E., et al. (2001).
Networks mediating the clinical effects of pallidal brain stimulation for Parkinson’s
disease: a PET study of resting-state glucose metabolism. Brain, 124, 1601–1609.
Georgi, J. C., Stippich, C., Tronnier, V. M., & Heiland, S. (2004). Active deep brain
stimulation during MRI: a feasibility study. Magn Reson Med, 51, 380–388.
Gildenberg, P. L. (2005). Evolution of neuromodulation. Stereotact Funct Neurosurg,
83, 71–79.
Green, A. L., Owen, S. L., Davies, P., Moir, L., & Aziz, T. Z. (2006). Deep brain stimu-
lation for neuropathic cephalalgia. Cephalalgia, 26, 561–567.
Hamani, C., Schwalb, J. M., Rezai, A. R., Dostrovsky, J. O., Davis, K. D., & Lozano,
A. M. (2006). Deep brain stimulation for chronic neuropathic pain: Long-term
outcome and the incidence of insertional effect. Pain, 125, 188–196.
Hariz, M. I. (2002). Complications of deep brain stimulation surgery. Mov Disord 17,
S162–6.
Hassler, R. (1955). The influence of stimulations and coagulations in the human
thalamus on the tremor at rest and its physiopathologic mechanism. Proceedings of
the Second International Congress of Neuropathology, 2, 637–642.
Hershey, T., Revilla, F. J., Wernle, A. R., McGee-Minnich, L., Antenor, J. V., Videen,
T. O., et al. (2003). Cortical and subcortical blood flow effects of subthalamic
nucleus stimulation in PD. Neurology, 61, 816–821.
Hilker, R., Voges, J., Weisenbach, S., Kalbe, E., Burghaus, L., Ghaemi, M., et al. (2004).
Subthalamic nucleus stimulation restores glucose metabolism in associative and
limbic cortices and in cerebellum: evidence from a FDG-PET study in advanced
Parkinson’s disease. J Cereb Blood Flow Metab, 24, 7–16.
Hillebrand, A., & Barnes, G. R. (2002). A quantitative assessment of the sensitivity of
whole-head MEG to activity in the adult human cortex. Neuroimage, 16, 638–650.
Hirata, M., Kato, A., Taniguchi, M., Ninomiya, H., Cheyne, D., Robinson, S. E., et al.
(2002). Frequency-dependent spatial distribution of human somatosensory evoked
neuromagnetic fields. Neurosci Lett, 318, 73–76.
Hosobuchi, Y. (1986). Subcortical electrical stimulation for control of intractable pain
in humans. Report of 122 cases (1970–1984). J Neurosurg, 64, 543–553.
Hosobuchi, Y., Adams, J. E., & Linchitz, R. (1977). Pain relief by electrical stimulation of
the central gray matter in humans and its reversal by naloxone. Science, 197, 183–186.
420 MEG: An Introduction to Methods
Hosobuchi, Y., Adams, J. E., & Rutkin, B. (1973). Chronic thalamic stimulation for
the control of facial anesthesia dolorosa. Arch Neurol, 29, 158–161.
Jenkinson, N., Nandi, D., Aziz, T. Z., & Stein, J. F. (2005). Pedunculopontine nucleus:
a new target for deep brain stimulation for akinesia. Neuroreport, 16, 1875–1876.
Jenkinson, N., Nandi, D., Miall, R. C., Stein, J. F., & Aziz, T. Z. (2004). Pedunculo-
pontine nucleus stimulation improves akinesia in a Parkinsonian monkey. Neuro-
report, 15, 2621–2624.
Jenkinson, N., Nandi, D., Oram, R., Stein, J. F., & Aziz, T. Z. (2006). Pedunculopon-
tine nucleus electric stimulation alleviates akinesia independently of dopaminergic
mechanisms. Neuroreport, 17, 639–641.
Jimenez, F., Velasco, F., Salin-Pascual, R., Hernandez, J. A., Velasco, M., Criales,
J. L., et al. (2005). A patient with a resistant major depression disorder treated with
deep brain stimulation in the inferior thalamic peduncle. Neurosurgery, 57, 585–93;
discussion 585–593.
Koller, W. C., Lyons, K. E., Wilkinson, S. B., & Pahwa, R. (1999). Efficacy of unilate-
ral deep brain stimulation of the VIM nucleus of the thalamus for essential head
tremor. Mov Disord, 14, 847–850.
Krack, P., Batir, A., Van Blercom, N., Chabardes, S., Fraix, V., Ardouin, C., et al.
(2003). Five-year follow-up of bilateral stimulation of the subthalamic nucleus in
advanced Parkinson’s disease. N Engl J Med, 349, 1925–1934.
Krack, P., Kumar, R., Ardouin, C., Dowsey, P. L., McVicker, J. M., Benabid, A. L.,
et al. (2001). Mirthful laughter induced by subthalamic nucleus stimulation. Mov
Disord, 16, 867–875.
Krack, P., Pollak, P., Limousin, P., Benazzouz, A., & Benabid, A. L. (1997). Stimulation
of subthalamic nucleus alleviates tremor in Parkinson’s disease. Lancet, 350, 1675.
Krauss, J. K., Pohle, T., Weigel, R., & Burgunder, J. M. (2002). Deep brain stimulation
of the centre median-parafascicular complex in patients with movement disorders.
J Neurol Neurosurg Psychiatry, 72, 546–548.
Krauss, J. K., Yianni, J., Loher, T. J., & Aziz, T. Z. (2004). Deep brain stimulation for
dystonia. J Clin Neurophysiol, 21, 18–30.
Kringelbach, M. L. (2005). The orbitofrontal cortex: linking reward to hedonic expe-
rience. Nature Reviews Neuroscience, 6, 691–702.
Kringelbach, M. L., Jenkinson, N., Green, A., Hansen, P. C., Cornelissen, P. L., Holliday,
I. E., et al. (2006). Deep brain stimulation and chronic pain mapped with MEG.
Society for Neuroscience, 782.1.
Kringelbach, M. L., Jenkinson, N., Green, A. L., Owen, S. L. F., Hansen, P. C.,
Cornelissen, P. L., et al. (2007a). Deep brain stimulation for chronic pain
investigated with magnetoencephalography. Neuroreport, 18, 223–228.
Kringelbach, M. L., Jenkinson, N., Owen, S. L. F., & Aziz, T. Z. (2007b). Translational
principles of deep brain stimulation. Nature Reviews Neuroscience, 8, 623–635.
Kringelbach, M. L., O’Doherty, J., Rolls, E. T., & Andrews, C. (2003). Activation of the
human orbitofrontal cortex to a liquid food stimulus is correlated with its subjective
pleasantness. Cerebral Cortex 13, 1064–1071.
Kringelbach, M. L., Owen, S. L. F., & Aziz, T. Z. (2007c). Deep brain stimulation.
Future Neurology, 2, 633–646.
Kulisevsky, J., Berthier, M. L., Gironell, A., Pascual-Sedano, B., Molet, J., & Pares, P.
(2002). Mania following deep brain stimulation for Parkinson’s disease. Neurology,
59, 1421–1424.
Using MEG to Elucidate the Principles of DBS 421
Kumar, R., Dagher, A., Hutchison, W. D., Lang, A. E., & Lozano, A. M. (1999). Globus
pallidus deep brain stimulation for generalized dystonia: clinical and PET investi-
gation. Neurology, 53, 871–874.
Lauritzen, M. (2005). Reading vascular changes in brain imaging: is dendritic calcium
the key? Nat Rev Neurosci, 6, 77–85.
Lenz, F. A., Kwan, H. C., Martin, R. L., Tasker, R. R., Dostrovsky, J. O., & Lenz, Y. E. (1994).
Single unit analysis of the human ventral thalamic nuclear group. Tremor-related
activity in functionally identified cells. Brain, 117, 531–543.
Leone, M., Franzini, A., Broggi, G., May, A., & Bussone, G. (2004). Long-term follow-up
of bilateral hypothalamic stimulation for intractable cluster headache. Brain, 127,
2259–2264.
Litvak, V., Eusebio, A., Jha, A., Oostenveld, R., Barnes, G. R., Penny, W. D., et al.
(2010). Optimized beamforming for simultaneous MEG and intracranial local field
potential recordings in deep brain stimulation patients. Neuroimage, in press.
Logothetis, N. K., & Wandell, B. A. (2004). Interpreting the BOLD signal. Annu Rev
Physiol, 66, 735–769.
Marchand, S., Kupers, R. C., Bushnell, M. C., & Duncan, G. H. (2003). Analgesic and
placebo effects of thalamic stimulation. Pain, 105, 481–488.
May, A., Leone, M., Boecker, H., Sprenger, T., Juergens, T., Bussone, G., et al.
(2006). Hypothalamic deep brain stimulation in positron emission tomography.
J Neurosci, 26, 3589–3593.
Mayberg, H. S., Lozano, A. M., Voon, V., McNeely, H. E., Seminowicz, D., Hamani, C.,
et al. (2005). Deep brain stimulation for treatment-resistant depression. Neuron,
45, 651–660.
Mazars, G., Merienne, L., & Ciolocca, C. (1973). Intermittent analgesic thalamic sti-
mulation. [Preliminary note]. Rev Neurol (Paris) 128, 273–279.
Mazars, G., Roge, R., & Mazars, Y. (1960). Results of the stimulation of the spinotha-
lamic fasciculus and their bearing on the physiopathology of pain. Rev Prat, 103,
136–138.
Mazzone, P., Lozano, A. M., Stanzione, P., Galati, S., Scanati, E., Peppe, A., et al.
(2005). Implantation of human pedunculopontine nucleus: a safe and clinically
relevant target in Parkinson’s disease. Neuroreport, 16, 1877–1881.
Nandi, D., Aziz, T., Carter, H., & Stein, J. (2003). Thalamic field potentials in chronic
central pain treated by periventricular gray stimulation — a series of eight cases.
Pain,101, 97–107.
Nandi, D., Liu, X., Winter, J. L., Aziz, T. Z., & Stein, J. F. (2002). Deep brain stimu-
lation of the pedunculopontine region in the normal non-human primate. J Clin
Neurosci, 9, 170–174.
Nuttin, B. J., Gabriels, L. A., Cosyns, P. R., Meyerson, B. A., Andreewitch, S., Sunaert, S. G.,
et al. (2003). Long-term electrical capsular stimulation in patients with obsessive-
compulsive disorder. Neurosurgery, 52, 1263–1272.
Owen, S. L., Green, A. L., Stein, J. F., & Aziz, T. Z. (2006a). Deep brain stimulation for
the alleviation of post-stroke neuropathic pain. Pain, 120, 202–206.
Owen, S. L. F., Green, A. L., Nandi, D., Bittar, R. G., Wang, S., & Aziz, T. Z. (2006b).
Deep brain stimulation for neuropathic pain. Neuromodulation, 9, 100–106.
Pammer, K., Hansen, P. C., Kringelbach, M. L., Holliday, I., Barnes, G., Hillebrand, A.,
et al. (2004). Visual word recognition: the first half second. Neuroimage, 22,
1819–1825.
422 MEG: An Introduction to Methods
Perlmutter, J. S., & Mink, J. W. (2006). Deep brain stimulation. Annu Rev Neurosci,
29, 229–257.
Perlmutter, J. S., Mink, J. W., Bastian, A. J., Zackowski, K., Hershey, T., Miyawaki, E.,
et al. (2002). Blood flow responses to deep brain stimulation of thalamus.
Neurology, 58, 1388–1394.
Petrovic, P., & Ingvar, M. (2002). Imaging cognitive modulation of pain processing.
Pain, 95, 1–5.
Petrovic, P., Kalso, E., Petersson, K. M., & Ingvar, M. (2002). Placebo and opioid
analgesia—imaging a shared neuronal network. Science, 295, 1737–1740.
Plaha, P., & Gill, S. G. (2005). Bilateral deep brain stimulation of the pedunculopon-
tine nucleus for idiopathic Parkinson’s disease. Neuroreport, 16, 1883–1887.
Pool, J. L., Clark, W. D., Hudson, P., & Lombardo, M. (1956). Steroid hormonal
response to stimulation of electrodes implanted in the subfrontal parts of the brain.
In: Fields, W. S., Guillemin, R., & Carton, C. A. (eds.). Hypothalamic-Hypophysial
Interrelationships (pp. 114–124). Springfield, IL: Charles C. Thomas.
Ray, N. J., Kringelbach, M. L., Jenkinson, N., Owen, S. L. F., Davies, P., Wang, S., et al.
(2007). Using magnetoencephalography to investigate deep brain stimulation for
cluster headache. Biomedical Imaging and Intervention Journal, 3, e25.
Rehncrona, S., Johnels, B., Widner, H., Tornqvist, A. L., Hariz, M., & Sydow, O.
(2003). Long-term efficacy of thalamic deep brain stimulation for tremor: double-
blind assessments. Mov Disord, 18, 163–170.
Rezai, A. R., Lozano, A. M., Crawley, A. P., Joy, M. L., Davis, K. D., Kwan, C. L., et al.
(1999). Thalamic stimulation and functional magnetic resonance imaging: locali-
zation of cortical and subcortical activation with implanted electrodes. Technical
note. J Neurosurg, 90, 583–590.
Richardson, D. E., & Akil, H. (1977a). Long term results of periventricular gray
self-stimulation. Neurosurgery, 1, 199–202.
Richardson, D. E., & Akil, H. (1977b). Pain reduction by electrical brain stimulation
in man. Part 1: Acute administration in periaqueductal and periventricular sites.
J Neurosurg, 47, 178–183.
Richardson, D. E., & Akil, H. (1977c). Pain reduction by electrical brain stimulation
in man. Part 2: Chronic self-administration in the periventricular gray matter.
J Neurosurg, 47, 184–194.
Robinson, S. E., & Vrba, J. (1999). Functional neuroimaging by synthetic aperture
magnetometry (SAM). In: Yoshimoto, T., Kotani, M. Kuriki, S., Karibe, H., & N.
Nakasato. (eds.). Recent Advances in Biomagnetism. (pp. 302–305). Sendai: Tohoku
University Press.
Sakatani, K., Katayama, Y., Yamamoto, T., & Suzuki, S. (1999). Changes in cerebral
blood oxygenation of the frontal lobe induced by direct electrical stimulation of tha-
lamus and globus pallidus: a near infrared spectroscopy study. J Neurol Neurosurg,
Psychiatry 67, 769–773.
Schnitzler, A., & Gross, J. (2005). Normal and pathological oscillatory communica-
tion in the brain. Nat Rev Neurosci, 6, 285–296.
Schuurman, P. R., Bosch, D. A., Bossuyt, P. M., Bonsel, G. J., van Someren, E. J.,
de Bie, R. M., et al. (2000). A comparison of continuous thalamic stimulation and
thalamotomy for suppression of severe tremor. N Engl J Med, 342, 461–468.
Siegfried, J., & Lippitz, B. (1994). Bilateral chronic electrostimulation of ventropo-
sterolateral pallidum: a new therapeutic approach for alleviating all parkinsonian
symptoms. Neurosurgery, 35, 1126–1129.
Using MEG to Elucidate the Principles of DBS 423
Singh, K. D., Barnes, G. R., Hillebrand, A., Forde, E. M., & Williams, A. L. (2002).
Task-related changes in cortical synchronization are spatially coincident with the
hemodynamic response. Neuroimage, 16, 103–114.
Stefani, A., Lozano, A. M., Peppe, A., Stanzione, P., Galati, S., Tropepi, D., et al.
(2007). Bilateral deep brain stimulation of the pedunculopontine and subthalamic
nuclei in severe Parkinson’s disease. Brain, 130, 1596–1607.
Stefurak, T., Mikulis, D., Mayberg, H., Lang, A. E., Hevenor, S., Pahapill, P., et al.
(2003). Deep brain stimulation for Parkinson’s disease dissociates mood and motor
circuits: a functional MRI case study. Mov Disord, 18, 1508–1516.
Sydow, O., Thobois, S., Alesch, F., & Speelman, J. D. (2003). Multicentre European
study of thalamic stimulation in essential tremor: a six year follow up. J Neurol
Neurosurg Psychiatry, 74, 1387–1391.
Taniguchi, M., Kato, A., Fujita, N., Hirata, M., Tanaka, H., Kihara, T., et al. (2000).
Movement-related desynchronization of the cerebral cortex studied with spatially
filtered magnetoencephalography. Neuroimage, 12, 298–306.
Temel, Y., van Lankveld, J. J., Boon, P., Spincemaille, G. H., van der Linden, C., &
Visser-Vandewalle, V. (2004). Deep brain stimulation of the thalamus can influ-
ence penile erection. Int J Impot Res, 16, 91–94.
Tronnier, V. M. (2003). Deep brain stimulation. Amsterdam, London: Elsevier.
Uitti, R. J., Tsuboi, Y., Pooley, R. A., Putzke, J. D., Turk, M. F., Wszolek, Z. K., et al.
(2002). Magnetic resonance imaging and deep brain stimulation. Neurosurgery, 51,
1423–1428.
Vidailhet, M., Vercueil, L., Houeto, J. L., Krystkowiak, P., Benabid, A. L., Cornu, P.,
et al. (2005). Bilateral deep-brain stimulation of the globus pallidus in primary
generalized dystonia. N Engl J Med, 352, 459–467.
Visser-Vandewalle, V., Temel, Y., Boon, P., Vreeling, F., Colle, H., Hoogland, G.,
et al. (2003). Chronic bilateral thalamic stimulation: a new therapeutic approach in
intractable Tourette syndrome. Report of three cases. J Neurosurg, 99, 1094–1100.
Vrba, J., & Robinson, S. E. (2001). Signal processing in magnetoencephalography.
Methods, 25, 249–271.
Author Index
424
Author Index 425
Clarke, J., 38 Enatsu, R., 384 Gross, J., 161, 162, 163, 216,
Coffey, R. J., 406 Endo, H., 329 221, 226, 228, 230, 232, 233,
Coghill, R. C., 321 Engel, A. K., 216 234, 238, 239, 267, 268, 367,
Cohen, D., 1, 34, 84 Erdler, M., 326, 327 393, 404
Cohen, L., 362, 363 Ermer, J. J., 112 Guilford, J. P., 114
Coltheart, M., 133, 134, 346, Gunji, A., 328, 330
353, 354 Fahn, S., 405 Guy, C. N., 275
Connolly, J. F., 351 Fairhurst, M., 417
Cooper, L. N., 27 Favorov, O. V., 16 Hadamard, J., 90
Corchs, S., 285 Ferlazzo, F., 267 Hadjipapas, A., 226, 230
Cornelissen, P. L., 142, 349, Fernandez, A., 381, 393 Halata, Z., 320
351, 355, 362, 363 Feynman, R. P., 87 Halchenko, Y. O., 282
Cosmelli, D., 222, 225, 240 Fieseler, T., 225 Halgren, E., 142, 175, 230,
Cramer, S., 392 Fischer, M., 378 352, 353
Cuffin, B. N., 84 Fischl, B. R., 50, 116, 200, 204, Halliday, D. M., 218
Curio, G., 17, 39, 60, 109, 316 248, 256 Hämäläinen, M. S., 13, 16, 17,
Czisch, M., 286 Flor, H., 313, 315, 391 26, 29, 86, 87, 110, 127, 159,
Foerster, O., 300 187, 191, 249, 375
Dahlhaus, R., 218, 219 Forss, N., 130, 313, 316, 318, Hamani, C., 406
Dale, A. M., 102, 132, 159, 161, 319, 379, 387, 391 Hamer, H. M., 382
176, 187, 188, 192, 193, 194, Foucher, J. R., 219, 278, 287, Hara, K., 378, 379
197, 200, 204, 230, 249, 254, 289 Hari, R., 1, 8, 15, 71, 147, 149,
274, 283 Franzini, A., 406 150, 164, 169, 300, 303, 310,
Darvas, F., 86, 106, 112, 114, Freiwald, W. A., 16 313, 317, 330, 331, 357, 375,
115, 267 Fries, P., 216 378, 387
Daunizeau, J., 118, 187, Friston, K. J., 116, 217, 231, Hariz, M., 408
197, 284 240, 249, 250, 252, 253, 254, Harrison, L. M., 240
Davatzikos, C., 256 255, 257, 285 Hashimoto, I., 17, 39, 60,
David, O., 118, 197, 285, 222, Fuchs, M., 106, 193, 201, 316
225, 240 249 Hassler, R., 405
Davis, K. D., 321 Fujimaki, N., 283 Hayasaka, S., 258, 260, 261,
Davison, A. C., 114 Fukuda, M., 409 262, 267
Debener, S., 226, 289 Haykin, S., 162
Deco, G., 285 Gaetz, W., 178 Helenius, P., 138, 142, 175, 352,
Deecke, L., 19 Galan, L., 231, 265 353, 354, 356, 357, 359, 361,
Dehaene, S., 275, 280 Gallen, C. C., 303, 315, 374, 362, 363, 389
Del Gratta, C., 280 381, 384 Heller, L., 193, 256
Delorme, A., 226 Gallien, P., 315, 391 Hershey, T., 409
Demoment, G., 104 Ganslandt, O., 384, 386 Hilker, R., 409
de Munck, J. C., 13 Garcia-Larrea, L., 321 Hill, D., 386
De Tiège, X., 46 Garnero, L., 222, 225 Hillebrand, A., 86, 100, 110,
Devlin, J. T., 363 Garreffa, G., 287 161, 178, 193, 226, 230, 231,
Devor, A., 198, 290 Geddes, L. A., 113 250, 253, 262, 410
Diamond, M. E., 16 Genovese, C., 265, 266 Hillyard, S. A., 175, 351, 352
Ding, M., 222 George, N., 219 Hinkley, D. V., 114
DiNocera, F., 267 Georgi, J. C., 409 Hirata, M., 388, 389, 413
Dogdas, V., 112 Geselowitz, D. B., 84, 404 Hochberg, Y., 260, 265
Douiri, A., 225 Gill, S. G., 405 Hoechstetter, K., 226
Downing, P., 280 Glabus, M. F., 285 Holcomb, P. J., 352
Duffay, H., 387, 392 Goldman, R. I., 286, 289, 291 Holmes, A. P., 115, 255, 263,
Duong, T. Q., 276 Golub, C. F., 101 231, 232
Gomez, J. F., 166 Horwitz, B., 275, 276, 285
Edgington, M., 263 Gommert-Novotny, E., 320 Hosaka, H., 84
Edwards, E., 266 Goncalves, S. I., 112, 279 Hoshiyama, M., 302, 303, 310,
Eichler, M., 218 Gootjes, L., 388 327, 328, 329
Elbert, T., 313, 314 Gotman, J., 290 Hosobuchi, Y., 406, 406,
Elger, C. E., 18, 19 Grasman, R. P., 226 408
Eliashiv, D. S., 382 Gray, C. M., 16 Hoyer, D., 222
Elton, R. L., 405 Green, A. L., 406 Huang, M. X., 4, 112, 196, 315,
Embick, D., 352 Greenblatt, R. E., 232 326, 383, 388, 393
426 Author Index
Hughes, J., 373 Karhu, J. A., 18, 303, 310, 316 Lenz, F. A., 321, 404, 405
Hui, H., 268 Karnisky, W., 265 Leone, M., 406
Huizenga, H. M., 95, 226 Karp, P. J., 108 Le Van Quyen, M., 219
Hund, M., 384 Kaufman, L., 300 Lewine, J. D., 383, 392
Hurtado, J. M., 222 Kaukoranta, E., 303 Liang, H., 222
Husain, F. T., 285 Kawamura, T., 303 Liljeström, M., 149, 152, 176
Huttunen, J., 303, 317 Kay, S. M., 95, 113 Limpiti, T., 201
Hyvärinen, A., 62 Kelhä, V. O., 46 Lin, F. H., 102, 106, 132, 161,
Kelso, J. A., 216 171, 172, 193, 194, 198, 201,
Iadecola, C., 277 Kettenmann, B., 381 225, 378, 386
Iida, K., 378 Kiebel, S. J., 115, 231, 246, 250, Linkenkaer-Hansen, K., 393
Ikeda, A., 326 251, 252, 254, 257 Lippitz, B., 405
Ikeda, H., 17 Kilner, J. M., 262 Liston, A. D., 289
Ilmoniemi, R. J., 49, 143, 159, Kim, D. S., 274, 277 Liu, A. K., 160, 176, 187, 188,
187, 197, 249 Kim, J. S., 386, 388 197, 283
Im, C. H., 284 Kirchberger, K., 382 Liu, L. C., 226
Ingvar, M., 417 Kircher, T. T., 278 Lobaugh, N. J., 114, 232
Inoue, K., 303, 316 Kirveskari, E., 388 Logothetis, N. K., 217, 276, 409
Inoue, T., 387 Kitamura, Y., 317, 321 Loose, R., 329
Inui, K., 303, 308, 317, 318, 320, Kitney, R. I., 166 Lopes da Silva, F. H., 2, 3, 5, 8,
321, 322, 323, 324 Klöppel, S., 388, 389 18, 20, 67, 147, 164, 218
Ioannides, A. A., 18, 160, 193, Knake, S., 210, 383 Lorente de Nó, R., 4
222, 225, 226 Knecht, S., 313 Lütkenhöner, B., 166, 392
Ischebeck, A., 364 Knowlton, R., 378, 380, 382, Lutzenberger, W., 265
Itier, R. J., 265 393
Itomi, K., 303, 309 Kober, H., 277, 384, 387, 389 MacKay, N., 202
Iwasaki, M., 383, 386 Koller, W. C., 405 Maclin, E., 315
Kooi, K. A., 84 Maeda, K., 313, 314
Jacobs, J., 266 Kornhuber, H. H., 19 Magerl, 318
Jacques, C., 266 Korvenoja, A., 387 Maihöfner, C., 315, 391
Jasper, H. H., 67 Krack, P., 405, 408 Mainen, Z. F., 12
Jenkinson, N., 405 Krakow, K., 274 Mair, 318
Jensen, O., 166, 167, 168, 171, Krauss, J. K., 405, 406 Makeig, S., 226
176, 177, 178 Kringelbach, M. L., 377, 393, Mäkelä, J. P., 126, 315, 376, 381,
Jerbi, K., 108, 117, 209, 268 404, 405, 410, 417 384, 385, 386, 387, 392
Jessel, J. P., 301 Krings, T., 387 Mamelak, A., 378, 382
Jimenez, F., 406 Kristeva, R., 326 Mangin, J. F., 116
Jobard, G., 363 Kristeva-Feige, R., 326 Maquet, P., 286
Johnson, H. J., 256 Kruger, L., 320 Marantz, A., 352, 353
Jokisch, D., 178 Kujala, J., 163, 226, 230, 364, Marchand, S., 406
Josephson, B. D., 28 367 Marinkovic, K., 353, 357
Joshi, A. A., 248, 256 Kulisevsky, J., 408 Maris, E., 263, 265, 268
Jousmäki, V., 57, 71, 303, Kumar, R., 405 Martinez-Montes, E., 291
375 Kurths, J., 219 Massimini, M., 19
Jun, S. C., 95 Kutas, M., 175, 351, 352 Matsuura, K., 159, 160, 187, 195
Juottonen, K., 315, 331, 391 Kwong, K. K., 186 Mattout, J., 118
Jurko, F., 406 Kyuhou, S., 13 Mauguiere, F., 303
May, A., 409
Kager, H., 19 Lachaux, J. P., 198, 216, 219, 221 Mayberg, H. S., 406, 410
Kaiser, J., 265 Lahaye, P. -J., 290 Mayville, J. M., 326
Kajola, M., 49, 166 Laine, B., 353, 354 Mazars, G., 406
Kakigi, R., 303, 316, 318, 321 Lamusuo, S., 378, 380, 382, 383 Mazziotta, J. C., 50
Kakuda, N., 234 Laufs, H., 274, 289 Mazzone, P., 405
Kalcher, J., 169 Lauritzen, M., 409 McCabe, C.S., 331
Kamada, K., 292, 315, 374, 384, Leahy, R. M., 113, 190, 200, McCarthy, G., 353
387, 392 204, 250, 268, 277 McIntosh, A. R., 114, 232
Kaminski, M., 222 Lehericy, S., 384 Mechelli, A., 217
Kanda, M., 321 Lehmann, D., 99 Medendorp, W. P., 178
Kandel, E., 301 Lehtelä, L., 149 Medvedovsky, M., 377
Kang, J. K., 290 Leistner, S., 19 Merlet, I., 383
Kanno, A., 307 Lemieux, L., 280, 286, 287 Merrifield, W. S., 388
Author Index 427
Meunier, S., 92, 99, 115 Ossadtchi, A., 102, 352 Qiu, Y., 318, 321, 325
Mikuni, N., 383 Otsubo, H., 382
Miller, W. T., 84 Ou, W., 196 Raij, T. T., 319, 331
Mima, T., 313, 316 Owen, S. L. F., 406 Ramantani, G., 383
Mink, J. W., 404 Rappelsberger, P., 222
Mitra, P. P., 166, 219, 233 Paetau, R., 378, 383 Rauschecker, J. P., 207
Mizuhara, H., 293 Palus, M., 222 Ray, N. J., 410
Modena, I., 373, 374 Palva, J., 393 Rehncrona, S., 405
Mogilner, A., 303, 310, 313, 391 Pammer, K., 413 Rennie, C. J., 240
Molenberghs, G., 255 Pantazis, D., 115, 117, 231, 232, Rezai, A. R., 385, 386, 409
Mölle, M., 19 250, 253, 254, 257, 259, 262, Richardson, D. E., 406
Moller, E, 222 263, 264, 265, 266 Riera, J. J., 197, 198, 285
Moon, T. K., 195 Pantev, C., 126 Roberts, D., 386
Moradi, F., 277, 280 Papanicolaou, A. C., 382, 388 Robinson, S. E., 48, 161, 162,
Morgan, D. E., 56 Park, H., 253, 262, 265 226, 231, 240, 413
Moseley, G. L., 331 Parkes, L. M., 289 Rodriguez, E., 219, 221
Mosher, J. C., 101, 102, 113, Parkkonen, L., 49, 384 Rogers, R. L., 17, 303
190, 250 Parviainen, T., 357, 362, 365, Rosenberg, J. R., 218
Mountcastle, V. B., 13, 16 366, 388 Rosenblum, M. G., 219, 221
Mumford, J. A., 255, 255 Pascual-Marqui, R. D., 106, Rossini, P. M., 280, 303, 313,
Munger, B. L., 320 160, 187, 194, 250, 315, 391
Murakami, S., 10, 11, 12, 13, 255, 262 Rossion, B., 266
15, 17 Passingham, R. E., 240 Rubchinsky, L. L., 222
Pataraia, E., 378 Rudrauf, D., 225
Nagamine, T., 313, 326 Paulesu, E., 360, 362, 363
Nagarajan, S. S., 226, 230, 362 Penfield, W., 85, 300, 310 Saarinen, M., 152
Nakagawa, H., 303, 304 Percival, D. B., 166 Sakamoto, K., 302, 303, 310,
Nakamura, A., 310 Perlmutter, J. S., 404, 409 313
Nakasato, N., 315, 329, 389 Perrin, F., 93 Sakatani, K., 409
Nakata, H., 321 Pesaran, B., 166, 219, 233 Sakman, B., 12
Nandi, D., 405, 406 Petit-Dutaillis, D., 84 Sakuma, K., 316
Narich, L., 303 Petrovic, P., 417 Salek-Haddadi, A., 286
Neshige, R., 326 Peyron, R., 321 Salenius, S., 149, 330, 386
Neumann, J., 290 Pflieger, M. E., 232 Salmelin, R., 2, 126, 142, 143,
Nguyen, B. T., 302, 303, 312 Pfurtscheller, G., 147, 169, 218 146, 147, 149, 150, 151, 152,
Nguyen, T. B., 313 Phillips, C., 357 164, 169, 330, 331, 347, 349,
Niazy, R. K., 286 Phillips, N. A., 351 351, 353, 355, 356, 357, 358,
Nichols, T. E., 115, 231, 232, Pikovsky, A. S., 221 359, 360, 362, 363, 364,
255, 258, 260, 262, Plaha, P., 405 367, 389
263, 267 Plaut, D., 346 Sameshima, K., 222
Nicholson, C., 14 Pleger, 315 Sammer, G., 276
Niedermeyer, E., 3, 5, 8, 18, 67, Ploghaus, A., 323 Sams, M., 152
164, 178, 179 Ploner, M., 316, 317, 321 Sarvas, J., 110
Nihashi, T., 303, 307, 309, Plonsey, R., 8, 9, 20 Saunders, G. H., 56
310, 311 Poeppel, D., 275, 357 Sawamoto, N., 321
Nobre, A. C., 350, 351, 353, Poghosyan, V., 222, 226 Schack, B., 222
355, 362 Poline, J. B., 262 Scherg, M., 99, 307
Nolte, G., 109 Pollok, B., 233, 239 Schiffbauer, H., 385, 386
Novotny, G. E. K., 320 Ponomarenko, V., 221 Schmidt, R. F., 318
Nunez, P. L., 276 Pons, T. P., 310 Schmidt, R. O., 101
Nuttin, B. J., 406 Pool, J. L., 406 Schmitz, A., 222
Popper, K. R., 89 Schmitz, F., 232
Ochoa, 318 Portas, C. M., 286 Schnitzler, A., 130, 216, 303,
Oja, E., 62 Posthuma, D., 165 321, 330, 393, 404
Okabe, Y., 159, 160, 187, 195 Pouthas, V., 282 Schreiber, T., 222
Okada, Y. C., 13, 16, 17, 39, 60, Praamstra, P., 67, 326 Schrieffer, J. R., 27
85, 99, 187 Press, W. H., 97 Schultz, M., 278, 281
Oliviero, A., 391 Price, C. J., 217, 363 Schuurman, P. R., 405
Oostenveld, R., 67, 263, 265 Puce, A., 350 Schwartz, D., 109
Opsommer, E., 318 Pujol, J., 387 Schwartz, E. S., 378, 381
Osipova, D., 178, 393 Pylkkänen, L., 352 Schwenkreis, P., 331
428 Author Index
Seber, G. A. F., 253 Tagamets, M. A., 285 Van Petten, C., 352
Sejnowski, T. J., 12 Takanashi, Y., 280 van Rotterdam, A., 2, 9, 14, 20
Sekihara, K., 182, 226, 230, 250, Tallon-Baudry, C., 164, 165, Van Veen, B. D., 100, 101,
253, 265 169, 231 161, 161, 164, 226, 230, 231,
Seppä, H., 38 Tamhane, A. C., 260 249, 255
Sereno, M. I., 102, 159, 157, Tanaka, N., 376 Varela, F., 216, 219, 268
192, 283 Tang, L., 382 Vates, E., 392
Sergent, C., 106 Taniguchi, M., 413 Vaughan, H. G., 285
Service, E., 368 Tanosaki, M., 316 Verbeke, G., 255
Servos, P., 310 Tao, J. X., 107, 383 Vidailhet, M., 405
Shattuck, D. W., 200, 204 Tarantola, A., 89, 90, 105, 118 Vieth, J. B., 149
Shaywitz, S. E., 359, 360 Tarkiainen, A., 140, 141, 153, Vihla, M., 357
Shen, D., 256 347, 348, 362, 349, 350, 351, Visser-Vandewalle, V., 406
Shibasaki, H., 326 360, 362 Vitikainen, A-M., 377, 382
Shih, J., 378 Tass, P., 221, 225 Volegov, P., 62
Shimizu, H., 303 Taulu, S., 49, 62, 74, 202, 376 von Cramon, D., 99
Shimojo, M., 303, 305, Tecchio, F., 303, 392 Vrba, J., 48, 161, 162, 226, 231,
306 Temel, Y., 408 374, 413
Shiraishi, H., 210, 382 Teolis, A., 246
Shmuel, A., 276 Terry, J. R., 218, 222 Wagner, M., 201
Siegfried, J., 405 Tesche, C., 18, 166, 197 Waites, A. B., 278
Silberstein, R. B., 276 Thatcher, R. W., 166 Waldorp, L. J., 99, 248, 253
Silén, T., 331 Thompson, P. M., 256 Walter, W. G., 19
Simola, J., 62, 376 Thomson, D. J., 219 Wan, X., 286
Simon, N. R., 19 Tian, B., 207 Wandell, B. A., 276, 409
Simos, P. G., 353, 356, 360 Tiihonen, J., 303 Wang, J. Z., 106
Singer, W., 216 Tikhonov, A., 90, 249 Wang, X., 317
Singh, K. D., 232, 250, Tilz, C., 382 Wasaka, T., 303
253, 257, 262, 263, 265, Timmermann, L., 228, 234, Watanabe, S., 321
280, 413 238, 239 Wax, M., 101
Singh, M., 275 Toga, A. W., 50 Weekes, B. S., 139, 141,
Sobel, D. F., 315 Toma, K., 282 354, 355
Somjen, G. G., 19 Tononi, G., 240 Weisend, M. P., 190, 206
Soto, J. L. P., 253 Torquati, K., 282 Weiss, T., 313
Speckmann, E. J., 18, 19 Tran, T. D., 318, 319 Welch, P. D., 166, 227
Spencer, M. E., 100 Traub, R. D., 9 Wessberg, J., 234, 236
Sporns, O., 240 Treede, R. D., 318 Whitzel, B. L., 313
Stam, C. J., 218, 222 Tronnier, V. M., 406 Wikstrom, H., 313, 315, 391
Stancak, A., 323 Tuch, D. S., 112, 292 Wikswo, J. P., 9
Stefan, H., 380 Turner, R., 115 Wilson, T. W., 356
Stefani, A., 405 Tuunanen, P. I., 290 Witzel, T., 193, 194, 225
Stefurak, T., 409 Wood, C. C., 93, 300
Stenbacka,, L., 171 Uesaka, Y., 316 Worrell, G. A., 381
Stephan, K. E., 240 Uitti, R. J., 409 Worsley, K. J., 231, 248,
Steriade, M., 19 Uusitalo, M. A., 49, 143, 197 261, 262
Sterr, A., 314 Uutela, K., 74, 109, 132, 159, Wydell, T. N., 133, 141, 142,
Stippich, C., 274, 277 160, 161, 173, 174, 187, 188, 351, 355, 356, 357,
Stirling, W. C., 195 192, 195, 201, 377 362, 364
Stuart, G. J., 12
Sudmeyer, M., 233, 239 Vallbo, A. B., 234, 236 Xiang, J., 303
Suk, J., 303 Van ‘t Ent, D., 2 Xu, X., 321
Sutherling, W., 300, 374, 380, van Dijk, H., 178, 180
382, 385, 393 van Drongelen, W., 226, Yamasaki, H., 321
Svensson, P., 321 230, 231 Yang, T. T., 303, 310, 313
Swinney, K. R., 9 Van Essen, D. C., 50 Yekutieli, D., 266
Sydow, O., 405 Vanhatalo, S., 19 Yu, H., 378
Symms, M. R., 278 van Hulsteyn, D. B., 193
Szmuk, P., 381 Vanni, S., 166, 167, 168, 171, Zimmerman, J. E., 29
Szymanski, M., 388 173, 174 Zimmerman, J. T., 99
Subject Index
Note: Page number followed by “f” and “t” refers to figures and tables, respectively.
Activation maps, for network Auditory evoked magnetic fields BEM. See Boundary Element
analysis, 223, 225–26 (AEFs), 33–34, 34f Method
dipole models, 226 language lateralization of, Berger bands, 39
magnetic field tomography, 388 BESA. See Brain electric source
225–26 Auditory evoked responses, analysis
minimum-norm techniques, replicability of, 73–74, Bessel filter, 61. See also F ilters
225 74f, 126–29, 126–29f Biophysical source modeling,
spatial filter, 226 Auditory stimulation system, 157–58
Active noise compensation 54–56 Biophysics, 5–9
systems, 48 Autoregressive models, 222 BOLD (blood-oxygen-level
AEFs. See Auditory evoked Averaging, 59 dependent), 76f, 78,
magnetic fields Axial gradiometer, 32, 32f. See 276–77, 280, 281,
Affective disorders also Gradiometer 282–83, 287, 288–89,
deep brain stimulation for, lead fields of, 33f 290–91, 368, 409
406, 409–10 Bonferroni correction method,
Aliasing, 41 Bandpass filter, 60. See also 231, 260
Alpha rhythm, 147 Filters Boundary Element Method
Ampère’s law, 5 Band-stop filter, 60. See also (BEM), 111, 111f
Amplitude quantization, Filters Brain activation statistical maps,
41–44, 43f BDS. See Deep brain creation of, 248–57, 249f
Analog-to-digital conversion, 42f stimulation General Linear Modeling,
Anatomical MRI data, Beamformers, 100–101, 157, 280 250–54
acquisition and Capon’s, 162 contrast statistic and
processing of, Beamforming, 161–64 normalization, 254–55
200–201 for identifying prestimulus general univariate
Anti-alias filtering, 41 induced oscillatory formulation, 253–54
Artifacts activity, 178–80, 179f, mass-univariate approach,
checking for, 69 180f 252–53
identifying and avoiding, linearly constrained minimum multivariate approach, 253
71–73, 72f variance, 162, 226 MEG observations, creation
-related problems, of MEG Behavioral responses, of, 249–50
signals, 375–77 recording, 58 multisubject studies, 255–57
429
430 Subject Index
complications of, 408 Dynamic range, 39–40 Face stimulation, 310–11, 312f
mechanics of, 406–8 Dynamic statistical parametric False discovery error rate
MEG study of maps (dSPMs), 175, (FDR), 265–66, 267f
discussion, 416–17 187, 194, 225 Family-wise error rate (FWER)
experimental setup, Dystonia. See also Movement controlling, 116–17. See also
412–13 disorders Bonferroni correction
future possibilities, 417–18 deep brain stimulation for, 405 method
in patients with cluster- strong, 258
headache, 411–12 Ear stimulation, 307–9, 309–11f weak, 258
in patients with phantom ECD. See Equivalent current Fast Fourier transform (FFT),
limb chronic pain, dipole model 61, 165, 218, 219f
410–11, 411f EEGlab, 230 FEM. See Finite Element
potential problems, 417 Electric current, 25 Method
results, 413–16 Electrocorticography Ferromagnetism, 25
for movement disorders, comparison with FFT. See Fast Fourier transform
405, 409 magnetoence- Field pattern of MEG, 8–9, 8f
for pain relief, 407f phalography, 382 Fieldtrip, 230, 232
problems with fMRI, 408–10 Electroencephalography (EEG) Filtering, 59–61
problems with PET, 408–10 in clinical settings Filters
safety of, 408 complementary properties bandpass, 60
Depression. See also Affective of, 383 band-stop, 60
disorders Electrophysiological basis, of Bessel, 61
deep brain stimulation for, MEG signals, 1–20 Chebyshev, 61
406, 410 Elekta Neuromag®system, 35f comb, 60
Developmental dyslexia Epileptiform activity, MEG for finite impulse response, 61
cortical correlates of, 358–62, comparison with electrocor- frequency-domain, 61
359–61f ticography, 382 high-pass, 60
Dewar, 36–37, 37f, 52 complementary properties low-pass, 60
DI. See Directionality index of, 383 notch, 60
Diamagnets, 25 enhancing gain of, 381 Finite Element Method (FEM),
DICS. See Dynamic imaging of interictal vs. ictal MEG, 382 111, 111f
coherent sources mesial epileptiform activity, Finite impulse response filter,
Direct readout, 38 detection of, 383–84 61. See also Filters
Directed transfer function sensitivity of, 380–81 Flux modulation, 38
(DTF), 225 temporal dynamics of, Flux transformer, 31f
Directionality index (DI), 221–22 378–79, 379f, 380f geometrics of, 31–32, 32f
Distributed current estimates, EPSPs. See Excitatory principle of, 31f
159–61, 159f. See also postsynaptic potentials fMRI. See Functional magnetic
Current estimates Equivalent current dipole resonance imaging
principle (ECD) model, 124, Forward modeling problem, 87,
bias in estimating superficial 125–26, 187, 226, 229 107–13
sources, 160–61 for cortical rhythmic activity head tissues, modeling,
dSPMs. See Dynamic statistical analysis, 147–52, 109–13, 111f, 112f
parametric maps 148–51f sensor array, modeling, 108–9
DTF. See Directed transfer ERD. See Event-related source, modeling, 107–8
function desynchronization FreeSurfer, 200, 204
Dual-route model, of reading, ERS. See Event-related Frequency-domain filter, 61.
133, 134f, 354f. See synchronization See also Filters
also Reading; Reading Essential tremor. See also Functional magnetic resonance
familiar words, Movement disorders imaging (fMRI), 404
cortical dynamics of; deep brain stimulation for, BOLD signals and MEG/EEG
Reading unfamiliar 405, 409 signals, comparison
words, cortical Event-related between, 75–77, 76f
dynamics of desynchronization for deep brain stimulation,
Dynamic causal modeling (ERD), 169 problems with,
(DCM), 240 Event-related synchronization 408–10
Dynamic imaging of coherent (ERS), 169 in localization of central
sources (DICS), 162, Excitatory postsynaptic suclus, 387
176, 178, 226–29, potentials (EPSPs), FWER. See Family-wise
227f, 230, 367 3, 3f, 4 error rate
432 Subject Index
Multi-dipole modeling in MEG, NIOI. See Near-infrared optical dipole reconstruction with
124–54 imaging fMRI information,
cortical rhythmic activity, Noise cancellation within MEG 282–83
ECD analysis for, devices, 48–50 distributed models of, 280
147–52, 148–51f Noise covariance matrix, distributed reconstruction
reading words aloud, 142–47, estimation of, 202–4, with fMRI
143–47f 203f information, 283–84
reading words and Noise-normalization, 193–94 similarities, finding, 281
nonwords, 134–42 Noninvasive functional Notch filter, 60. See also Filters
differences between tomographic Nutmeg, 230
experimental connectivity analysis Nyqvist frequency, 41
conditions, testing with MEG, 216–40
for, 139–42, 140–42f activation maps for network Obsessive compulsive disorder.
source model, analysis, 223, 225–26 See also Affective
constructing, 135f, dipole models, 226 disorders
136–39, 137f, magnetic field deep brain stimulation for,
139–42f tomography, 225–26 406
responses to auditory minimum-norm Ohm’s law, 5, 6
stimulation, 126–29, techniques, 225 Open field, 4, 5f
126–29f spatial filter, 226 Orthonormal right-handed
responses to right-median discussion, 239–40 head coordinate
nerve stimulation, dynamic imaging of coherent system, 51f
130–33, 130–32f sources, 226–29,
Multiple signal classification 227f, 230 Painful laser stimulation
(MUSIC) algorithm, example, 234–39 Aδ -and C-fiber pain, 317–19
101–2, 250 localization method, choice interstimulus interval, impact
Multivariate Analysis of of, 229–30 of, 319
Covariance nonparametric dependency Paramagnetism, 25
(MANCOVA), 253 measures Parametric dependency
Multivariate autoregressive coherence, 218–19, 219f measures, of
model (MVAR), 222 information flow, functional
Mu rhythm, 147, 329–31 direction of, 221–22 connectivity, 222
MUSIC. See Multiple signal phase synchronization, Parkinson’s disease. See also
classification 219–21, 220f Movement disorders
algorithm parametric dependency deep brain stimulation for,
MVAR. See Multivariate measures, 222 405, 409
autoregressive model regions of interest, Partial least squares (PLS), 232
identification of, PCA. See Principal component
NAI. See Neural activity index 230–33 analysis
Near-infrared optical imaging statistical considerations, Periventricular-periaqueductal
(NIOI), 290 222–23, 224f gray region (PVG/
Negative feedback, 29–30 strategies and practical PAG), 406, 408
Neural activity index (NAI), considerations, implantation of electrodes in
159, 162, 164 233–34 for patients with phantom
Neuroimaging techniques, Nonparametric dependency limb chronic pain,
combining, 273–93 measures, of 410–11, 411f
brain imaging techniques, functional Permeability, 25
288–89 connectivity Permutation methods, 231–32,
converging evidence, 275 coherence, 218–19, 219f 262–65, 264f, 265f
difficulties of, 274–75 information flow, direction PET. See Positron emission
enticement of, 273–74 of, 221–22 tomography
generative models, 276 phase synchronization, Phantom limb chronic pain. See
quantified data with shared 219–21, 220f also Chronic pain
information, 275 Non-simultaneous MEG/EEG deep brain stimulation for,
Neurological diseases and fMRI, 277–85 410–11, 411f
modifications of spontaneous assumptions of, 278–79 neuroimaging data,
MEG by, 392–93 beamformer (scanning) 414–16, 414f, 415t
Neuronal modeling, 9–13, analysis, 280 subjective pain
10f, 11f dipolar analyses of, 279–80 experience, 413–14
Subject Index 435