Multichannel SQUID Biomagnetic Systems Author Jiri VRB
Multichannel SQUID Biomagnetic Systems Author Jiri VRB
Multichannel SQUID Biomagnetic Systems Author Jiri VRB
JIRI VRBA
CTF Systems Inc.
15-1750 McLean Ave, Port Coquitlam, B.C., Canada, V3C 1M9
Abstract. The field of biomagnetism advanced considerably since the first recordings
of magnetic fields of the human heart in 1963 and of the human brain in 1968. Since
the introduction of whole-cortex magnetoencephalography (MEG) systems in 1992, the
number of installed channels has dramatically increased, and the magnetic evaluation of
the human brain has been gradually finding its place in clinical work. MEG is presently
the most important biomagnetic application, and sophisticated MEG systems with large
numbers of channels have been developed commercially. The MEG systems must meet
certain specifications on noise, dynamic range, slew rate and linearity because they are
exposed to environmental noise even when they are operated within shielded rooms. The
systems are designed to meet these specifications through optimized design of SQUID
flux transformers, SQUID control electronics and data acquisition, and development of
various synthetic noise cancellation techniques. The interpretation of the resulting mag-
netic data is enhanced by combining the MEG results with information from electroen-
cephalography (EEG) and other imaging modalities. In addition, an engineering effort is
devoted to the development of various items of MEG peripheral equipment (stimulators,
patient support, head positioning, etc.).
1. Introduction
The detection of the electromagnetic activity of the human heart and brain has a long
history. The first known recording of the human electrocardiogram (ECG) was reported
in 1887 [1], and more extensive recordings were made early in the 20th century after the
perfection of the string galvanometer [2]. However, it took nearly 80 years after the first
reported ECG before the first magnetocardiogram (MCG) was recorded using conven-
tional coils with large numbers of turns [3]. Similarly, the electroencephalogram (EEG)
was discovered in 1929 [4] and its magnetic counterpart, the magnetoencephalogram
(MEG) was first recorded about 40 years later in 1968 using room-temperature coils [5].
In order to perform more detailed measurements of biomagnetic fields, more sensi-
tive instruments were needed. The significant increase in sensitivity was made possible
by the Nobel-prize-winning discovery of the Josephson effect in 1962 [6] and the subse-
quent development of highly sensitive detectors of magnetic field, called SQUIDs
(S uperconducting QUantum Interference Devices) [7 - 11] in 1964. SQUIDs were first
used for detection of MCG in 1970 [12], and for detection of MEG in 1972 [13].
The time-line of the ECG/MCG and EEG/MEG history is shown in Figure 1.1.
The first magnetic recordings with superconducting SQUID detectors were made using
simple systems consisting of one or a few channels, and it was about two decades later
before the present multi-channel systems were developed.
Since the first recordings of the magnetic fields of the human heart and brain were
made by SQUID detectors, magnetic fields were detected from numerous other body or-
gans, for example from the eye (magnetooculogram and magnetoretinogram), the stom-
ach (magnetogastrogram, MGG), the small intestine (magnetoenterogram, MENG), the
skeletal muscle (magnetomyogram, MMG), the peripheral nerves (magnetoneurogram,
MNG), the fetal heart (fMCG), the fetal brain (fMEG), the lungs (magnetopneumo-
gram), etc. The biomagnetic fields range from 10 fT for the spinal chord to more than a
nT for magnetized lungs, and the frequency range is from less than 0.1 Hz to several
hundreds of Hz [14, 15] - see Figure 1.2.
Figure 1.2. Representative biomagnetic fields and their frequency ranges. From references [14 and 15].
Even though magnetic fields were detected from many organs, so far the most im-
portant application of biomagnetism has been to the brain, and MEG recordings have
been gradually moving from research laboratories to clinical practice. Examples of clini-
cal or near-clinical applications are use of MEG in pre-surgical mapping [16], localiza-
tion of epileptic seizure foci [17], and the detection of abnormal slow waves associated
with stroke, head trauma, and transient ischemic attack in the brain [18-21].
Since MEG is presently the most important biomagnetic application, and since it
encompasses the smallest signal range and a large span of frequencies, it will be used as
a model for discussion of biomagnetic systems - problems and technological solutions
for other biomagnetic applications are similar to that encountered in MEG. MEG also
has spurred intensive technological development in the commercial sector, which has led
to the development of sophisticated MEG systems with large numbers of channels
which cover the whole head surface. These helmet systems became commercially avail-
able in 1992 [22, 23, 24]. Presently, there are three commercial suppliers of MEG sys-
tems in Europe and North America (Biomagnetic Technologies, Inc. [25], CTF Sys-
tems, Inc. [26] and Neuromag Ltd. [27]), and approximately 30 whole-head systems
have been sold worldwide to date.
The introduction of the helmet-type MEG systems has led to a dramatic increase in
the use of SQUID sensors. It is estimated that the present number of SQUID sensors in
the commercially-sold whole-cortex MEG systems is approximately 4,300. This num-
ber should be put into perspective by considering that before 1992 (a period of 22 years
since 1970 when the first SQUID recording of the magnetic field of human heart was
made) the total number of functional SQUIDs in all applications was only slightly
more than 1,000 [15]. Even though the SQUID sensors and superconducting compo-
nents are the "enabling" devices for biomagnetism, and without them the detection of
small magnetic fields would not be possible, they constitute only a small part of the
4
overall cost of the present MEG systems (about 3 to 4 %), while the cost of nonsuper-
conducting components (cryogenics, electronics, software, peripheral equipment and
shielding) is dominant.
Currently, all commercial MEG systems are based on low-temperature SQUIDs,
however, systems based on high-temperature superconductors also are being developed.
Even though the present high-temperature SQUIDs are not as reliable and sensitive as
their low-temperature counterparts, their performance is steadily improving, and they are
already suitable for some applications.
In addition to commercial activity, SQUID biomagnetometers with large numbers of
channels have been constructed in a number of laboratories world-wide using both low-
and high-temperature superconductors. A few examples, not intended to be exhaustive,
of such systems based on low-temperature superconductors and cooled by liquid helium
are: 19-channel system developed at the University of Twente [30], a double 14-channel
system developed by Dornier [31, 32], a 19-channel system build by the Russian com-
pany, Cryoton [33], an 83-channel system developed at PTB Berlin [34], a double 31-
channel system built by Philips [35], a double 37-channel system commercially pro-
duced by BTi [25], a 129-channel system by Shimadzu [36], a 128-channel system de-
veloped by the Superconducting Sensor Laboratory in Japan [37]; systems based on
low-temperature sensors, but cooled by cryocoolers: 32- and 61-channel systems by
Daikin Industries Ltd. [38, 39]; and systems based on high-temperature superconductors
and cooled by liquid nitrogen: e.g., a 32-channel system by a Japanese group [40].
The MEG system measures magnetic fields on the head surface. However, to the
user, the more relevant information is the current distribution in the brain responsible
for the measured fields. Unfortunately, the field inversion problem is ill-posed - the in-
version is not unique - and the MEG data must be supplemented by additional informa-
tion or additional physiological constraints, or the mathematical source models must be
simplified.
One way to supply more useful information is to use also the EEG [28], an older
electrical counterpart of MEG. Both MEG and EEG measure the same sources of neu-
ronal activity, and their information is complementary and additive [29]. For these rea-
sons, EEG is usually also included as a part of a biomagnetic system. Additional infor-
mation to assist the field inversion also can be supplied by other imaging techniques.
For structural information one uses the brain atlas [41], magnetic resonance imaging
(MRI, [43-47]), and computed axial tomography (CAT or CT, [42]); for functional in-
formation one uses positron emission tomography (PET, [48, 49]), single photon
emission computed tomography (SPECT, [48 - 50]), and functional MRI (fMRI, [43,
49, 51-53]).
MEG and EEG provide a direct measure of neuronal electrical activity with excellent
temporal resolution, but spatial characterization is dependent on a non-unique inverse
problem. MRI and CT provide high-resolution volumetric data defining anatomy, and
PET, SPECT, and fMRI provide a 3-dimensional representation of functional activity in
the brain in terms of metabolic and hemodynamic variables. In comparison with MEG
and EEG, the PET, SPECT and fMRI are limited by the slow time-course of the meta-
bolic and hemodynamic processes, and by the poorly defined relationship between them
and neural activation.
A comparison of temporal and spatial resolutions of various imaging methods is
shown in Figure 1.3. CT and MRI provide the best spatial resolutions (≤ 1 mm), but
they can be used only for the determination of anatomical structures. MEG, fMRI, and
5
Figure 1.3. Comparison of temporal and spatial resolutions for various imaging techniques (* the inversion
problem for MEG and EEG is poorly defined). PET temporal resolution ranges from about 30 sec to 10 min,
where longer times are required for good S/N ratio. The temporal resolution for SPECT is given by the long
lifetime of the radioisotopes used. CT does not provide functional imaging and its temporal resolution only
reflects the scanning time. Light shaded area for fMRI extends to the minimum scan acquisition time of
about 40 msec.
PET have comparable spatial resolutions of several mm - however, the MEG spatial
resolution depends on a poorly-defined inverse problem. Standard SPECT has the worst
spatial resolution, about 10 mm. The spatial resolutions of all methods are more or less
comparable, yet their temporal resolutions are vastly different and suggest that the inte-
gration of spatially-robust techniques (MRI, fMRI and PET) with temporally superior
techniques (MEG and EEG) has a potential to yield imaging capabilities with optimal
spatial and temporal resolutions [54, 55].
Biomagnetometers are complex devices and their construction involves multidisci-
plinary engineering efforts in areas of sensor development, cryogenics, noise cancella-
tion, electronics, system design, software, and data handling, interpretation and process-
ing. The aspects of cryogenics, SQUID-sensor fundamentals, and biomagnetic signal
analysis are covered by other chapters in this volume. The present chapter will concen-
trate on system questions associated with biomagnetic instrumentation. First, in Sec-
tion 2 the origins of biomagnetic signals and the relationship between MEG and EEG
will be discussed. The general configuration of a biomagnetic system will be shown in
Section 3. Biomagnetometers are exposed to environmental noise, and the system de-
sign must allow for noise reduction, the electronics must be capable of handling the en-
vironmental dynamic ranges and slew rates, and flux transformers must be optimized for
the best signal detection in the presence of noise. These questions will be discussed in
Sections 4, 5 and 6. Miscellaneous features of biomagnetic systems (the number of
channels, the measurement of head position, the registration with MRI images, stimula-
tion equipment, patient support and handling, etc.) will be discussed in Section 7, and
selected examples of MEG results will be shown in Section 8. Where applicable, the
discussion of various design issues will be illustrated by examples of CTF's instrumen-
tation, because the relevant technical details are accessible to the author.
6
In this section, the origins of magnetic fields generated by the brain are briefly recapitu-
lated, and field patterns produced by a simple model of equivalent current dipole in a per-
fectly spherical medium are explained. A more detailed treatment of the cellular electrical
properties can be found elsewhere [56-58]. The section is concluded by the discussion of
the relationship between the magnetic (MEG) and electric (EEG) measurements of the
brain.
The human brain is a poorly understood complex system. The cortex contains rela-
tively well-aligned pyramidal cells organized in cortical columns. The cells consist of
dendrites, a cell body and an axon, and there are approximately 105 to 106 cells in an area
of about 10 mm 2 of the cortex [59] - see Figure 2.1.a and b. There are many connec-
tions between various parts of the brain, mediated by the nerve fibers which impinge on
other dendrites and cell bodies via synapses. There are approximately 104 synapses per
cell, or about 1010 synaptic connections in several cubic mm volume of cortex. In the
whole brain there are more than 1010 cells and about 1014 synaptic connections.
In a pyramidal cell, the dendrite (or a cell body) can be thought of as a tube whose
wall is a membrane. Because of the existence of the Na-K pump [56], there is an excess
concentration of K inside and of Na outside the cell, and these concentration gradients
cause diffusion of K+ and Na + ions across the membrane. Since the membrane perme-
ability to K is significantly larger than to Na, the cell characteristics are mostly deter-
mined by the dynamics of the K+ ions. According to the Nernst equation [56], the diffu-
sion of K+ ions across the membrane leads to the establishment of negative potentials
inside the cell with an equilibrium value of about -70 mV.
Stimulation of the membrane potential (chemical, electrical or even mechanical) al-
ters the transmembrane potential and can cause cell repolarisation. Such repolarisation
happens, e.g., at the synapse, when neurotransmitters are released. Because the cell is
conductive, there will be a current between the positive depolarized region and the nor-
mal negative region of the cell. The current flow inside the cell is called the impressed
or intracellular current, and the return current flow outside the cell is called the volume
or extracellular current - see Figure 2.1.c. In the most common detection situation,
when the radial magnetic fields are measured, the detected fields (MEG) are produced by
the impressed currents, while the volume currents are mostly responsible for the electric
potentials (or EEG). The action potentials or axonal currents are usually not observable
magnetically, because the polarization is symmetrical, and the associated magnetic fields
have a fast spatial decay.
Figure 2.1. Origins of the brain's magnetic fields. (a) coronal section through the brain indicating the cortex
consisting of gray matter; (b) small section of cortex expanded to indicate pyramidal cells, dendrites and
synaptic connections; (c) intracellular and extracellular dendritic currents.
7
Figure 2.2. Generation of magnetic field by activated cortex. (a) the activation can be located either in
sulci, resulting in tangential currents, or in gyri, resulting in radial currents. CSF - cerebrospinal fluid; (b)
cortical currents produce fields detectable outside the head.
The impressed current flows in the dendritic area of the pyramidal cells, roughly per-
pendicular to the cortex - see Figure 2.2.a. The cortex has a complex shape and contains
numerous sulci and gyri. As a result, the cortical current can be either tangential or ra-
dial to the outer surface of the brain. If the brain were modeled by a perfect sphere, then
due to the symmetry, only the tangential impressed currents would produce fields out-
side the sphere, while the radial currents would produce zero magnetic field [63] - see
Figure 2.3 or 2.2.b.
Current flow in a single cell cannot cause a sufficiently strong magnetic field to be
detectable outside the head. To produce measurable fields, it is necessary to have nearly
simultaneous activation of a large number of cells, typically in the range from 104 to
several times 105 [57]. However, in some cases, this activation of a large number of
cells can still be assumed spatially small and can be modeled by a point equivalent cur-
rent dipole [64]. It has been shown that the maximum current dipole density is an in-
variant with magnitude in the range from 0.5 to 2 nA·m/mm2 [60]. Consider an exam-
ple of the Auditory Evoked Fields (AEF) as in Figure 4.22. When such fields are inter-
preted in terms of equivalent current dipoles, they yield dipole magnitudes of 20 to 80
nA·m [61], which in the framework of the maximum estimated current dipole density
[60] corresponds to the activated area of the cortical tissue of the order of 1 cm2.
The equivalent current dipole is an approximation useful only in a situation where
the magnetic field patterns are clearly dipolar, as in the case of various evoked fields. In
general, the cortical currents are distributed, and a more sophisticated analysis of mag-
netic fields should be performed. The concept of current dipoles is, however, very sim-
ple and will be used for the purposes of this paper.
The magnetic field of an equivalent current dipole in a spherical medium can be
computed analytically [62, 63]. For the geometry in Figure 2.4.c, where the current di-
pole q is located at position a, and the measurement is performed at r, the field is given
by
Figure 2.3. Tangential and radial currents in perfectly spherical conductors. (a) tangential impressed cur-
rents produce external magnetic field; (b) radial impressed currents do not produce external magnetic field.
8
µo
B (r ) = ( Fq ×a − q × a ⋅r∇F) (2.1)
4πF 2
where
ro = r - a (2.2)
A further simplification of the expression for the current dipole field can be obtained
if it is assumed that the measured field is radial. This is indeed the case as the majority
of MEG detectors are configured to measure the radial or near-radial components of
magnetic fields (or their derivatives) - see Section 6. If the sensor array is small, as in
Figure 2.4.a, the approximation of radial detectors is quite good; if the sensor array is
large, as in the helmet-type systems in Figure 2.4.b, the approximation of a radial de-
tectors is not satisfied for all sensors (especially for the sensors close to the helmet
rim). Nevertheless, a useful insight into the character of the brain fields can be obtained
by examining the radial fields. Let u rad be a unit radial vector at the sensor location,
then the radial magnetic field due to a tangential current dipole can be calculated using
the Biot-Savart law
µo q× ro
Brad = urad ⋅ (2.5)
4πro3
The radial field in a plane perpendicular to the dipole and passing through the dipole
center is shown in Figures 2.5.a and b for shallow and deep current dipoles. The field
exhibits two extrema of opposing polarity, and the magnitudes of the extrema and their
spacing depend on the dipole depth. If it is assumed that the conducting medium fills
half space with a planar boundary, the dipole depth is related to the separation between
the dipoles by depth = separation/√2 [64, 65]. An example of a dipolar field pattern, cor-
responding to brain response to the median nerve stimulation is shown in Figure 2.5.c.
In this case the single dipole approximation is quite satisfactory.
So far, magnetic fields due to an equivalent current dipole have been discussed. Elec-
tric potentials (EEG) are related to magnetic fields (MEG) because both methods detect
identical current generators - see Figure 2.6. The radial magnetic fields are mostly caused
Figure 2.4. Geometry for calculation of the magnetic field of an equivalent current dipole. (a) small detec-
tor array, all sensors can be assumed radial; (b) helmet-type detector arrays, not all sensors can be assumed
radial; (c) geometry for calculation of magnetic field.
9
Figure 2.5. The pattern of radial magnetic fields generated by an equivalent current dipole. (a) for a shal-
low dipole, the field magnitude is large, and the field extrema are close together; (b) for a deep dipole, the
field magnitude is small, and the field extrema are more separated; (c) an example of a field pattern due to
median nerve stimulation, positions of magnetic sensors are indicated by dots. Aadapted from ref. [16].
by the impressed currents, while the electric fields are caused by the volume currents,
and the magnetic field and electric potential patterns on the scalp surface are orthogonal
[65]. Further, the equivalent current dipoles determined from the electric and magnetic
measurements should be located at the same point, but with opposite directions [66].
An experimental example of such orthogonal electric and magnetic patterns is
shown in Figure 2.7 (due to mechanical stimulation of the right index finger [66]). The
contour maps of the measured fields are orthogonal, and the dipole directions determined
from electrical and magnetic maps are opposite (but at the same location). Note that the
magnetic field maps are more localized than the maps of electric potentials.
In conclusion, it was shown that the observable magnetic fields (MEG) are due to an
almost simultaneous collective excitation of a large number of cells. The magnetic
fields are mostly due to intracellular currents, while the electric potentials (EEG) are due
to extracellular (volume) currents. MEG is predominantly sensitive to tangential current
generators, and for radial generators, the magnetic fields cancel due to axial symmetry.
EEG potentials are sensitive to both radial and tangential generators.
Figure 2.6. The generation of magnetic fields and electric potentials. The radial magnetic fields are caused
by intracellular currents and electric potentials are caused by extracellular currents. The directions of cur-
rents determined from EEG and MEG are opposite.
10
Figure 2.7. Comparison of EEG and MEG (unshielded, 3rd-order gradiometer) signals for the averaged re-
sponse to mechanical stimulation of the right index finger, 32 channels of EEG and 143 channels of MEG.
(a, c) superimposed EEG and MEG signals; (b) isocontour map of electrical field at the peak of p50 re-
sponse, shaded area represents negative potential in steps of 0.37 µV; (d) isocontour map of magnetic field
at the peak of p50m response, shaded area corresponds to in-going field, contour intervals 23 fT. Modified
from reference [66].
Figure 3.1. Basic elements of a SQUID magnetometer. (a) Josephson junction (thin film tunnel junction in
this example); (b) magnetometer comprising flux transformer, DC SQUID with two Josephson junctions,
and connections to the electronics.
11
Figure 3.2. Examples of flux transformers. (a) magnetometer; (b) 1st-order radial gradiometer; (c) 1st-or-
der planar gradiometer; (d) 2nd-order gradiometer; (e) 3rd-order gradiometer.
ers, as in Figure 3.1.b. In addition to providing the match to the measured fields, the
flux transformers also provide noiseless gain, and their frequency response is flat from
DC up to a maximum operating frequency (because the flux-transformer circuits are su-
perconducting).
The flux transformer usually consists of pickup coils, which are exposed to the
measured fields, leads and a coupling coil which inductively couples the flux-transformer
current to the SQUID. Depending on the measurement objective, the flux transformers
can have a variety of configurations, examples of some are shown in Figure 3.2 [70].
The simplest flux transformer is a magnetometer whose pick-up coil is a loop of wire
(one or several turns), as in Figure 3.2.a. If the gradient of the magnetic field is required,
the pick-up consists of two oppositely-wound coils coupled by a common lead to the
SQUID sensor. The currents induced in the two coils subtract, and these devices detect
the variation (or gradient) of the magnetic fields. Examples of radial and planar 1st-order
gradiometer flux transformers are shown in Figures 3.2.b and c. Higher-order gradiome-
ter flux transformers can be constructed by correctly combining larger number of coils;
examples of 2nd- and 3rd-order hardware gradiometer flux transformers are shown in
Figures 3.2.d and e.
The SQUID sensors with their associated flux transformers are the basic building
blocks of magnetometers - see Figure 3.3. The flux transformers and SQUIDs are im-
mersed in a cryogen within a cryogenic container (dewar). The cryogen is usually liquid
Figure 3.3. Block diagram of elements of a SQUID magnetometer. The cryogenic dewar must be electro-
magnetically transparent to eliminate distortions of measured fields.
12
Figure 3.4. Block diagram of a typical MEG system. The MEG dewar is mounted in a gantry capable of
seated and horizontal positions. The MEG and EEG detectors in a shielded room are interfaced to elec-
tronics and DSP preprocessed before the data are acquired by a computer. There are provisions for deliv-
ery of miscellaneous stimuli and for observation of and communication with the patient.
helium for low-temperature sensors and liquid nitrogen for high-temperature sensors.
The dewar is a vacuum vessel with a complex set of various heat-shield components in
its vacuum space. The signals from SQUID sensors are brought to room temperature by
a transmission line, and are amplified and processed by SQUID electronics before acqui-
sition by a computer [22, 23, 71]. Software is used to control the SQUID electronics
(tuning, diagnostics, channel configurations, etc.) and data collection (sampling parame-
ters, filtering, etc.). Data are then processed, displayed and archived.
The individual magnetometers then are used as building blocks for biomagnetome-
ters (or specifically MEG) systems - see Figure 3.4. In an MEG system, the dewar is
constructed in a helmet shape to cover as large an area of the cortex as possible [22, 23,
Figure 3.5. BTi Magnes® 2500 WH MEG system capable of recording in seated and fully-reclined position.
Courtesy of Biomagnetic Technologies, Inc.
13
Figure 3.6. CTF Whole-Cortex MEG system. (a) fully supine position; (c) upright position.
24]. The dewar is mounted in a gantry which allows either a seated or horizontal patient
position.
The dewar and the gantry may or may not be positioned in a shielded room. The
MEG measurement is usually combined with EEG, and both MEG and EEG signals are
brought out from the room to the processing electronics and data collection computer.
In addition to data collection, there are usually several other computer work stations at-
tached to the system to allow for data analysis while the MEG system is used for meas-
urement. The system also contains a stimulus computer which is synchronized with the
data acquisition and can drive various stimulus hardware. The installation is completed
with video camera for patient observation and an intercom.
Examples of the existing and newly-announced commercial whole-head MEG sys-
tems are shown in Figures 3.5 to 3.7. The systems utilize magnetometers, radial or
Figure 3.7. Neuromag Vectorview™ MEG system. (a) fully supine measurement position; (b) upright meas-
urement position. Courtesy of Neuromag Ltd.
14
planar gradiometers, as well as a reference system for noise cancellation. The relative
merits of various sensor configurations are discussed in Section 6.
The MEG systems contain a number of diverse technological elements: a shielded
room and noise cancellation, SQUID sensor, flux transformers, cryogenics, mechanical
supports, electronics and data acquisition, EEG system, peripheral systems, data analy-
sis and interpretation. Each element represents its own interesting technology. The rest
of the presentation in this chapter will concentrate in greater detail on noise cancellation
and shielded rooms, flux transformers, and issues associated with biomagnetic data ac-
quisition. The issues associated with various peripheral equipment and measuring prac-
tices will be dealt with in a less detailed manner.
Biomagnetometers (or MEG systems) are exposed to noise even when operated within
shielded rooms. Therefore, it is important to understand the character of the noise and
have good methods for its elimination. In this section, the noise will be first described
and then several methods for its elimination will be discussed: shielding, active noise
cancellation, synthetic noise cancellation (high-order gradiometers and adaptive meth-
ods), and signal space noise cancellation.
Many examples in this section will show high-order gradiometer results. It should
be noted that the formation of high-order gradiometers is a synthetic method, and the
gradient order of a particular sensor can be selected not only during the recording, but it
can be arbitrarily changed off-line. Therefore, for a given recording, it is possible to si-
multaneously display the results transformed into several different gradiometer orders.
The noise levels for both magnetometers and gradiometers will be given in field
units (T). For gradiometers, such units are obtained by multiplying the correct gradient
unit by appropriate baseline lengths (e.g., the 1st gradient unit is [T/m]; however, the
unit of 1st-order gradiometer output will be [T], which corresponds to the 1st gradient
unit multiplied by baseline length [m]).
MEG systems are sensitive to all magnetic fields present in their measuring space. Of
prime interest are the fields generated by the investigated brain region (wanted signal),
e.g., region "S" in Figure 4.1. However, the measured fields also contain unwanted con-
tributions from various environmental noise sources (e.g., moving cars, elevators, ma-
chinery, tools, or power-line signals), from parts of the brain which are not being inves-
tigated (e.g., extended background brain activity during investigation of evoked signals),
or from other physiological activities (muscle contractions, eye movement, heart beats
detected either directly or through ballistic effects, etc.). The objective of the noise can-
cellation is to separate the wanted brain signals from the contributions due to various
unwanted sources.
It is assumed in the present discussion that the noise is of magnetic origin, i.e., the
noise is introduced into the system only via the flux transformers - see Section 3. Noise
generated by non-magnetic means is assumed to be eliminated by proper design of the
biomagnetometer (e.g., any rf interference is eliminated by proper rf shielding within
the instrument), and the system electronics is assumed to be perfect (i.e., it is fast
15
Figure 4.1. Magnetic sensors are subjected not only to the measured MEG signal S, but also to unwanted
signals (environmental noise, signals from parts of brain not being measured and other body parts).
enough with sufficiently large dynamic range and good linearity, as described in Section
5). Also, it is assumed that the sensor mounting and the cryogenic vessel have low
magnetic contamination and are sufficiently rigid such that the vibrational noise is neg-
ligible and various system coefficients are stable over long periods of time.
To appreciate the magnitude of environmental noise relative to biomagnetic fields,
examples of both are displayed on the same diagram in Figure 4.2. For comparison,
sensitivities of flux-gate [72, 73], optically-pumped [74], and SQUID [67] magnetome-
ters also are shown in Figure 4.2, and it is obvious that only the SQUID magnetome-
ters are sensitive enough for biomagnetic applications.
Environmental noise depends on the type of environment, the time of the day and
Figure 4.2. Comparison of biomagnetic fields, environmental noise, and sensitivity in 1Hz bandwidth of
various types of magnetometers.
16
whether the system is shielded or unshielded. As a result, for each device and shielding
condition, the observed noise shows a variability over several orders of magnitude.
Noise cited in a range of papers has been summarised [75 - 81, 70], and its envelopes
are presented in Figure 4.3 for magnetometers and 1st-order gradiometers in shielded and
unshielded environments. In order to design a system with universally reliable perform-
ance, conditions close to the "worst case" in Figure 4.3 should be considered.
For comparison, noise in the 3rd-order gradiometer mode also is shown in Figure
4.3. The upper limit of the curve corresponds to the noise achievable in unshielded envi-
ronments, and the lower limit corresponds to the noise achievable within shielded
rooms. Also shown is a 5 fT rms/√Hz dashed line, which represents a required resolu-
tion of the MEG systems.
The "bumps" in the range from 10 to 30 Hz in Figure 4.3 correspond to vibrational
noise calculated assuming magnitudes of 10-5 rad for rotational and 10 µm for transla-
tional vibrations, gradiometer common-mode vector C ≈ 10 -2, baseline d = 5 cm, and
DC field and gradient magnitudes in an unshielded environment of BUnsh = 50 µT, G Unsh =
200 nT/m and within shielded rooms of BSh = 50 nT and GSh = 10 nT/m [70].
The sharp peaks at 50- to 60-Hz range in Figure 4.3 correspond to the largest ob-
served amplitudes of power-line noise, which for unshielded environments are BUnsh ≈
500 nT, GUnsh ≈ 1 nT/cm, and in shielded environments BSh ≈ 5 pT and GSh ≈ 40 fT/cm.
The power-line peaks are plotted as if their magnitudes were constant in a 1-Hz band-
width. The relationship between the field and 1st gradient for power lines can be ap-
proximated by assuming that the field of the power-line source is B = M/R n, where M is
dipole density in proper units, R is distance from the power-line source, and n = 2 for
parallel wires (n = 1 for a single wire and n = 3 for a magnetic dipole). Then neglecting
the vector and tensor character of the fields and gradients, the relationship between B and
the 1st gradient G(1) can be approximated by
Figure 4.3. Summary of observed noise for balanced 1st-order gradiometers and magnetometers in shielded
and unshielded environments. Shielded environments are assumed to be µ-metal rooms with modest low
frequency attenuation [86]. Bumps in the frequency range from 10 to 30 Hz represent vibrational noise, and
the peaks in the range from 50 to 60 Hz represent power-line signals. These peaks represent signal
amplitudes and are shown as if their amplitudes were constant in a 1-Hz bandwidth. Dashed line - required
system white-noise level of 5 fT rms/√Hz. Gradiometer noise corresponds to gradient noise multiplied by
appropriate baselines, i.e., G(1) ·d and G (3) ·d1 ·d2 ·d3 . (a) 1st gradient noise (measured by a 1st-order gra-
diometer with baseline of d = 5 cm and balanced to C ≤ 10-5 ); (b) magnetic field noise. Adapted from [109].
17
Figure 4.4. Interference signals originating within the human body. The primary sensors were hardware
1st-order gradiometers with 5-cm baseline. (a) spontaneous brain activity and the system noise measured in
an unshielded environment (urban laboratory). fs = 250 samples/sec, collection bandwidth = DC to 50 Hz,
frequency bin fb = 0.122 Hz, 3rd-order gradiometer, files: no subject - Oct 7:15 (93), eyes closed - Oct 7:17
(93), eyes open - Oct 7:18 (93). All traces are in the 3rd-order gradiometer mode; (b) heart beats recorded
by a 64-channel MEG system inside a shielded room, fs = 125 samples/sec, bandwidth = DC to 40 Hz, file:
Feb 9:6 (93), subject D.C., no averaging.
G (1)d d
=n (4.1)
B R
Magnetic shielding is the most straightforward, but not the least costly, method for re-
duction of environmental noise. Over the years a variety of shields (shielded rooms)
have been used for biomagnetic applications - see Figure 4.5.a. The simplest shielding
18
Figure 4.5. Performance of various shielded rooms. (a) noise attenuation factors for Al room, standard µ-
metal room, high attenuation µ-metal room, whole body high-temperature superconducting shield, 3rd-order
gradiometer, and 3rd-order gradiometer and combined standard µ-metal room (also, see this figure (b) at
frequency < 1 Hz); (b) noise within a standard µ-metal shielded room (in Vienna), magnetic field, 1st- and
3rd-order synthetic gradiometers.
1st and 3rd gradient noise are shown within a standard-attenuation µ-metal room [81].
The magnetic field noise for frequencies below about 40 Hz, and 1st gradient noise for
frequencies below about 4 Hz, are larger than 5 fT rms/√Hz. Since the majority of ap-
plications require low frequencies (1 Hz or lower), it is obvious that the shielded-room
attenuation must be supplemented by additional noise cancellation.
Figure 4.6. Active noise cancellation. (a) coil system in an unshielded environment; (b) coil system com-
bined with a shielded room.
20
tometers, then the noise of the references is added to the MEG sensor noise [96] (e.g., if
the total MEG sensor noise is required to be 5 fT rms/√Hz and if the MEG sensor noise
without active compensation is 4 fT rms/√Hz, then the noise of the compensation refer-
ence detector should be ≤ 3 fT rms/√Hz). If the MEG sensors are gradiometers, then the
noise contribution due to the reference is more complicated and will depend on the field
uniformity produced by the compensation coils.
Consider a sensor and a number of separate references located at different positions from
the sensor. In such a system, synthetic noise cancellation will be understood to mean
subtraction of the reference outputs from the sensor output, using subtraction coeffi-
cients constructed to accomplish a specified task, e.g., to minimize the norm of the re-
sulting signal, or to construct a higher-order gradiometer, etc. Synthetic noise cancella-
tion is quite flexible because results with different characteristics can be obtained from
the same sensor and reference data. This is different from hardware noise cancellation,
where the coefficients of the subtraction are predetermined by the hardware construction.
For example, a 1st-order hardware radial gradiometer consists of two coils wound in op-
posite directions, the coil closer to the head may be considered to be a sensor, and the
more distant coil, the reference. If the gradiometer is perfect, then the subtraction coeffi-
cient is determined by the gradiometer construction and is equal to 1.
First, a general discussion of synthetic noise cancellation will be presented and the
differences between synthetic gradiometers and adaptive noise cancellation will be illus-
trated by an example of a magnetometer-based system. Then, the synthetic higher-order
gradiometers will be examined in greater detail, gradiometer balancing will be discussed,
and examples of noise cancellation by gradiometers will be shown. The discussion will
be concluded by demonstration of adaptive noise cancellation for the frequency-independ-
ent and frequency-dependent situations.
Figure 4.7. Synthetic noise-cancellation systems. (a) magnetometer-based system; (b) hardware 1st-order
radial gradiometer-based system; (c) hardware 1st-order planar gradiometer-based system.
21
Output of the synthetic cancellation system, s, can be expressed in terms of the sen-
sor output, σ, and reference outputs, ri , i = 1, ... R as
R
s = σ − ∑ ξi ri (4.2)
i=1
where ξi are the subtraction coefficients characteristic of the synthetic process, and R is
the number of references.
The character of the synthetic noise cancellation will be illustrated by an example of
a magnetometer-based system with a simple 3-component magnetometer reference, as in
Figure 4.8.a. Denote the magnetic field and 1st gradient tensor generated by the noise
sources at the origin by B o and Go - the field is a vector and the 1st gradient is a 9-com-
ponent tensor. If the magnetic sensor is located at the origin and oriented along a direc-
tion p (unit vector), and the reference is positioned at a distance d from the magnetic
sensor center, and if 2nd and higher gradients are neglected, then the magnetic field at the
reference center is given by
B = Bo + Go ⋅d (4.3)
Without loss of generality, it can be assumed that the gains of all reference magne-
tometers and also the sensing magnetometer are identical (and equal to 1), and that the
reference magnetometers are perfectly orthogonal and are oriented along the coordinate
axes. Then the three outputs from the reference magnetometers form a reference vector,
b ref , and the reference and sensor outputs are given by
σ = B o ⋅p (4.4)
bref = B = Bo + Go ⋅ d (4.5)
The output from the synthetic noise-cancellation process is then given by Eq.4.2 as
s = σ −ξ ⋅ bref = Bo ⋅ p − ξ ⋅( B o + Go ⋅ d ) (4.6)
Figure 4.8. Magnetometer-based synthetic noise-cancellation system with vector magnetometer as a refer-
ence. (a) general detection configuration; (b) system exposed to a field consisting of a number of noise
sources (dipoles M 1 , M 2 , and M 3 in the present diagram; however, the sources are not limited to dipoles
and there can be a number of sources different than three).
22
The output s in Eq.4.6 represents the device response to the noise after the noise has
been cancelled - at this time only the noise is considered; no signals are applied. Eq.4.6
can be used immediately to determine "true" gradiometer coefficients. Recall, that the
1st-order gradiometer is designed to cancel the magnetic field. Therefore, putting Go = 0
and requiring s = 0 in Eq.4.6, the coefficients corresponding to the "true" gradiometer,
ξgrad , are given by
ξgrad = p (4.7)
For a perfect gradiometer, the vector of the subtraction coefficients is equal to the
orientation of the magnetometer sensor. This result is intuitively obvious, as the gra-
diometer subtraction coefficients will project the reference outputs to the vector p and
will effectively create a synthetic reference which is parallel to the sensor coil.
Assume that at a given time noise is generated by a collection of noise sources as in
Figure 4.8.b. Sources need not be dipolar; dipoles were shown only for simplicity.
Eq.4.6 may be rewritten for a time instant tk
{ }
s(t k ) =σ (t k ) −ξ ⋅ bref (t k ) = Bo (t k )⋅ p − ξ ⋅ B o (t k ) + G o (t k ) ⋅d (4.8)
Also, define the norm of a vector as its magnitude squared. An example of the vector S
norm is
K
∑ s ( tk )
2
S = S⋅S = (4.10)
k=1
The objective of the synthetic noise cancellation is to minimize the norm ||S ||, or,
using Eqs.4.8 and 4.10, to minimize
2
K 3
S = ∑ Σk − ∑ ξiB refik (4.11)
k =1 i =1
Further, if it is assumed that K is large and that the components of magnetic fields and
gradients are independent, then the relationships between various vectors in Eqs.4.9
simplify to
Boi ·Boj = δij ||Boi || (4.12.a)
23
[ ( )]
3
S = ∑ p2i Boi − 2ξ ipi B oi +ξ 2i Boi + Gi (4.13)
i=1
For large K, the minimization of Eq.4.11 (or 4.13) relative to the coefficients ξi , yields
pi B oi
ξi = , i = 1, 2, 3 (4.14)
B oi + Gi
The coefficients in Eq.4.14 are different from the "true" gradiometer coefficients in
Eq.4.7. The general solution in Eq.4.14 is equivalent to the gradiometer only if the
norm of the gradient vector is zero, ||Gi || = 0, or if the minimization is performed in a
uniform, gradient-free field (as in a good quality Helmholtz coil). Then the "true" gra-
diometer subtraction coefficients depend only on the geometry of the sensor (the orienta-
tion vector p, Eq.4.7) and are universal, i.e., they do not have to be changed when the
system is operated in environments with variable noise character (i.e., variable field and
gradient composition).
In all other cases when the gradients are present during the minimization, the sub-
traction coefficients will be called the "adaptive coefficients." The adaptive coefficients
are different from the true gradiometer coefficients because they contain information
about the gradients present during the minimization. Thus, in the general case, the adap-
tive coefficients are not universal and have to be changed when the environmental char-
acter changes.
The difference between the "true" gradiometer and adaptive systems is shown graphi-
cally in Figure 4.9. In Figure 4.9.a, a general system consisting of a sensing coil and a
reference is shown; the reference is assumed to be a 3-component magnetometer. In
Figure 4.9.b a configuration corresponding to a "true" gradiometer is shown (Eq.4.7);
the reference is projected to look like the sensing coil (with the same orientation and
area), but displaced by the baseline d. Figure 4.9.c corresponds to the general adaptation
case. The projection of the reference is not parallel to the sensing coil (dashed line), but
the difference is not large, especially when the gradient terms ||Gi || are small. Finally,
the degenerate adaptation cases are shown in Figures 4.9.d 1 and d 2. It was assumed dur-
ing the derivation of Eq.4.14 that the number of measured noise states is large and that
the relationships in Eqs.4.12 are valid. If, however, the number of noise states is less
24
Figure 4.9. Magnetometer-based system, geometrical representation of the difference between various
types of subtraction coefficients. (a) general magnetometer-based system; (b) equivalent diagram for a true
gradiometer system; (c) adaptive system with adaptation performed with more than three noise states; (d1 ,
d2 ) degenerate adaptation with less than three noise states.
than 4, Eq.4.11 can be solved exactly to yield ||S || = 0. If the number of noise states is
3, there is only one solution, and if it is less than 3, there is an infinite number of solu-
tions. In these cases, the environmental fields can be cancelled perfectly by different ref-
erence configurations, e.g., a large coil nearly parallel to the applied field (Figure 4.9.d 1)
or a small coil nearly perpendicular to the field (Figure 4.9.d2). Such degenerate situa-
tions are obtained if short measurements are performed, or if the applied noise does not
exercise enough degrees of freedom.
The minimized magnitudes of the norm ||S || for gradiometer and adaptive systems
are obtained when the subtraction coefficients from Eqs.4.7 and 4.14 are substituted into
Eq.4.13
3 B G
S adapt
min = ∑ p i2 B oi + Gi (4.15.a)
i =1 oi i
3
∑ p i2 Gi
grad
S min = (4.15.b)
i =1
In the gradiometer case, the coefficients were determined to completely cancel the
field, and therefore the residual ||S || contains gradient components projected to the direc-
tions of the baseline and the sensor-coil orientation. In the adaptive case the minimized
||S || contains a mixture of gradients and fields. All quantities in Eqs.4.15 are positive,
and for a given component i, the adaptive term is smaller than the corresponding gra-
diometer term. Thus, the adaptive process yields better noise cancellation than the gra-
diometers, but only during the time period for which the noise character is not changing
(norms of Gi 's and Boi 's remain constant). When the noise character is changing rapidly,
the gradiometers perform much better than the adaptive process - depending on the time
of the day and location, the periods during which the noise character remains constant
can be as short as a few seconds [70]. This conclusion also is true for higher-order gra-
diometers. In addition, it also is found that the higher-order gradiometers perform better
than the lower-order adaptive systems.
Consider a simplified case where the magnetometer sensor is oriented along the x3
direction, p = (0, 0, 1), and introduce the ratio r i = ||Gi ||/||Boi ||. Then Eqs.4.15 simplify to
25
r3
S adapt
min = Bo3 (4.16.a)
1+ r3
grad
S min = Bo3 r3 = G 3 (4.16.b)
The amplitude of the mean residual noise is proportional to √||S || and is shown as a
function of √r3 in Figure 4.10. For small r3's the gradiometer and adaptive residual noise
are comparable. For large r3's the adaptive residual noise is independent of r3 and is
roughly equal to √||Bo3||. This indicates that the adaptive coefficients are dominated by
the environmental gradients and are structured to cancel them; the residual noise is
mostly given by the magnetic fields (if r3 → ∞, then ||S||min adapt → ||Bo3|| in Eq.4.16.a).
The true gradiometer response is monotonically increasing with increasing r3, as shown
by Eq.4.16.b. The crossover between the small and large r3 happens for r3 ≈ 1 or for
||Bo3|| ≈ ||G3||. At this point, the residual noise of the adaptive process is smaller than
that of the gradiometer by a factor of 1/√2.
Next, the equality ||Bo3|| ≈ ||G3|| will be examined, and physical conditions for which
r3 < 1 will be determined. Consider, for a moment, that the environmental noise is
caused by a dipole source moving on the surface of a sphere with radius R and origin at
the magnetic-field detector. The dipole has a constant moment M, and its orientations
and positions on the sphere surface are random. The magnetic field and gradient at the
sphere origin are given by [97]
M
Bo = βo (4.17.a)
R3
3M
Go = γo (4.17.b)
R4
where the vector βo and tensor γ o depend on dipole position and orientation and their
norms are of the order of 1. The norms ||Bo3|| and ||G3|| can be calculated by substituting
Eqs.4.17 to Eqs.4.9.c and e, and their ratio yields the coefficient r3 as
Figure 4.10. Magnetometer based system, mean residual noise for a "true gradiometer" and for an "adaptive
system."
26
2
3d
r3 = Ψ (4.18)
R
where Ψ is a ratio of the sum of squares of the components of βo and γ o, and its magni-
tude is of the order of 1:
2
d
{ }2
K K
Ψ = ∑ γo (t k ) ⋅ ∑ β o3 (t k ) (4.19)
k =1 d 3 k =1
n-th gradient. On the other hand, the adaptive systems minimize applied fields and gradi-
ents of all orders, but the field and lower gradients are not canceled as well as by gra-
diometers.
Figure 4.12. Response of spatial gradiometers to a dipole source. Dipole moment magnitude is M = 1017
fT·cm 3 (equivalent to a passenger car), gradiometer baseline unit = 5 cm, room shielding ratio = 70. The
hatched band corresponds to the required range of biomagnetometer resolution in a 1-Hz bandwidth.
28
Figure 4.13. Synthetic higher-order gradiometers. The primary sensors are low-order gradiometers, either
magnetometers (0th-order gradiometers) or hardware 1st-order gradiometers. The reference system con-
tains a sufficient number of magnetometers and gradiometers to allow synthetic formation of 2nd-, 3rd-, or
higher-order gradiometers.
Figure 4.14. Synthetic 1st-order gradiometer based on magnetometer sensor and a vector magnetometer
reference.
29
where α B is the reference gain, and Bk is one of the three orthogonal components of the
applied magnetic field, B. The reference magnetometer outputs, bk, form a vector of ref-
erence outputs, b. The sensing magnetometer output is given by the projection of the
applied magnetic field on the sensing-magnetometer-coil normal, p (as in Eq.4.4)
σ = α s(p·B) (4.21)
where α s is the sensing magnetometer gain. To form a software gradiometer, the refer-
ence magnetometer outputs are projected on the direction p, the projection is scaled by
the sensor-to-reference magnetometer gain ratio, and the result is subtracted from the
sensor output. Then, similar to Eq.4.5 (using the Taylor expansion for the magnetic
field at a location u, defining the gradiometer baseline as d = u 2 - u 1, and shifting the
origin to the center of the baseline), the synthetic 1st-order gradiometer output, g(1), can
be derived as [70]
αs
g (1) = σ − (p⋅ b) =α s p⋅G ⋅d (4.22)
αB
The output of the 1st-order gradiometer is equal to the projection of the 1st-gradient
tensor to the vectors p and d. Examples of ideal 1st-order gradiometers are shown in
Figure 4.15. Figure 4.15.a corresponds to an axial gradiometer where d = d(0, 0, 1), p 2
= -p 1 = (0, 0, 1) and g(1) = α G1p·G·d = α G1dG33. Figure 4.15.b corresponds to an off-axis
(planar) gradiometer where d = d(0, 1, 0), p 2 = -p 1 = (0, 0, 1) and g(1) = α G1dG23. In Fig-
ure 4.15.c the coils and the baseline are not aligned, d = d(0, 0, 1), p 2 = -p 1 = (0, 1/√2,
1/√2) and g(1) = α G1(G23 + G 33)/√2. In this last case, the gradiometer output is a linear
combination of the 1st-gradient tensor components.
Software formation of a 2nd-order gradiometer is illustrated in Figure 4.16 [70,
100]. Two ideal 1st-order gradiometers are located at positions u and u'. The
gradiometers have coil vectors p and p', and baselines d and d'.
The Taylor expansion is used to express the gradient at a point u in terms of the
gradient at the origin as G(u) = Go + G(2)·u, where Go and G(2) are the 1st and 2nd gra-
dient tensors measured at the origin. The outputs of the ideal 1st-order gradiometers can
then be written, using Eq.4.22, as
Figure 4.15. Examples of ideal 1st-order gradiometers. (a) axial (radial) gradiometer; (b) off-axis (planar)
gradiometer; (c) gradiometer with tilted coils.
30
Figure 4.16. Synthetic formation of a 2nd-order gradiometer using two 1st-order gradiometers.
The software 2nd-order gradiometer output, g(2), with gain equivalent to the gain of
gradiometer g, can be formed by scaling the output of g' by the ratio of the gains α G/α G'
and by the ratio of the baselines d/d', and by subtracting the scaled result from g. As-
suming that the vectors p and p', and d and d' are parallel, defining the 2nd-order gra-
diometer baseline as q = u - u' and shifting the origin to the center of the 2nd-order gra-
diometer baseline, the 2nd-order gradiometer output can be expressed as
αG d
g (2) = g − g' = αG p⋅ G(2) ⋅ q⋅ d (4.25)
αG' d'
The output of the synthetic 2nd-order gradiometer in Eq.4.25 is a projection of the 2nd
gradient tensor onto the gradiometer vectors p, q, and d, where p is a unit vector, and q
and d have lengths associated with them.
The two 1st-order gradiometers used for the synthesis of the 2nd-order gradiometer
can be constructed either in hardware, or either one or both can be formed synthetically.
If the sensing 1st-order gradiometer is synthetised from magnetometers, then the above
procedure describes the formation of a synthetic 2nd-order gradiometer based on a magne-
tometer sensor.
Examples of ideal 2nd-order gradiometers are shown in Figure 4.17. Figure 4.17.a
corresponds to a 2nd-order gradiometer where all three principal vectors are parallel, p ||
Figure 4.17. Examples of ideal 2nd-order gradiometers. (a) axial gradiometer G333 ; (b) gradiometer G 233 ;
(c) 2nd-order gradiometer with tilted baseline.
31
Figure 4.18. Sources of common-mode and eddy-current vectors. (a, b, c) common-mode vector sources;
(a) inequality of coil areas; (b) coil tilt; (c) presence of a superconducting object near the gradiometer; (d)
eddy-current vector source, presence of a normal-metal object near the gradiometer.
32
.
g (1) =α G1 C B ⋅B+ E⋅ B+ p⋅ G(1) ⋅d 1 (4.26)
.
g (2) =α G2 C B ⋅B + E ⋅ B+ CG1 ⋅y1d1 + p⋅ G(2) ⋅ d1 ⋅ d2 (4.27)
.
g (3) =α G3 C B ⋅B + E ⋅ B+ CG1 ⋅y1d1 + CG2 ⋅ y2 d1d2 +p ⋅G(3) ⋅ d1 ⋅d 2 ⋅d 3 (4.28)
where d1, d2, and d3 have been used to denote the baselines of the 1st-, 2nd- and 3rd-order
gradiometer; C B, C G1 and C G2 are 3-, 5- and 7-component common-mode vectors; y 1
and y 2 are 5- and 7-component vectors with elements equal to the linearly-independent
components of the 1st and 2nd gradient tensors; and G(1), G (2) and G(3) are the 1st, 2nd
and 3rd gradient tensors. The outputs of gradiometers with order higher than 3 can be
expressed similarly. For good noise cancellation the common-mode and eddy-current
vectors must be measured, and their effects eliminated from the gradiometer output [70].
Figure 4.19. Time plots of a typical channel of 143-channel MEG system collected in a shielded room
(Vienna). (a) over-night recording, about 9 hours, fs = 1.25 samples/sec, bandwidth = DC to 0.5 Hz, 1st-
and 3rd-order synthetic gradiometers on the same scale (noise spectrum for the time from the beginning of
the record to about 0:30 is shown in Figure 4.5.b); (b) expanded plot of 3rd-order synthetic gradiometer
from (a); (c) short, about 16 sec, day-time record, fs = 1250 samples/sec, bandwidth = DC to 300 Hz, 1st-
and 3rd-order synthetic gradiometers; (d) expanded plot of 3rd-order synthetic gradiometer from (c).
33
Figure 4.20. Noise spectra collected in unshielded environment (CTF laboratories), 143-channel system. |B|
- magnitude of magnetic field, g1 - 1st-order gradiometer, g 3 - 3rd-order gradiometer, dashed line - 5 fT
rms/√Hz. Peak-to-peak amplitude of the magnetic field magnitude at 60 Hz is about 260 nT peak-to-peak.
demonstrate the noise reduction as the gradiometer order is increased. A short, ≈ 16-sec
segment of data recorded in the same shielded room during the day is shown in Figure
4.19.c and d. Again, in the 1st-order gradiometer mode, the peak-to-peak noise is about
40 pT, and it reduces to less than 1 pT when the 3rd-order gradiometer is synthesized.
Figure 4.19 demonstrates that even within shielded rooms, the environmental noise is
high and can be successfully reduced by higher-order gradiometers.
The noise of higher-order gradiometers in unshielded environments is shown in Fig-
ure 4.20 in the frequency range from 0.1 to about 400 Hz. At lower frequencies, the
noise amplitude of the 3rd-order gradiometer is about 4 orders of magnitude smaller than
that of magnetometers, and is limited above ≈ 1 Hz by the noise floor of the SQUID
sensor to about 5 fT rms/√Hz. Generally, the unshielded environmental noise has been
reduced to below 10 fT rms/√Hz above ≈ 1 Hz, except for the environmental peaks due
to close-by activity or due to power lines. In this example, the magnitude of the mag-
netic field at 60 Hz is about 260 nT peak-to-peak. Considering the 3rd-order gradiometer
white noise level in a 1-Hz bandwidth of about 5 fT rms, q = 4, the dynamic range of
the environmental signal is D ≈ 8.3 x 10 8 or 178.4 dB (see Eq.5.1).
Further demonstration of the gradiometer performance is shown in Figure 4.21,
where the frequency content of the signal and noise are roughly the same, and the signal
cannot be separated from the noise by temporal filtering, yet it can be separated by spa-
tial filtering using gradiometers. In Figure 4.21.a, the MCG signal [100] was measured
in an unshielded environment. The detector was a synthetic gradiometer with hardware
1st-order gradiometer primary sensor and references capable of 2nd-order gradiometer
formation (a 7-channel system, operated with RF SQUIDs and with sensitivity of only
about 30 fT rms/√Hz). The increase of gradiometer order from 1st to 2nd removed al-
most all noise resulting in a clear MCG signal. In Figure 4.21.b an Auditory Evoked
34
Figure 4.21. Noise cancellation by high-order gradiometers, unshielded environment. (a) heart signals in
1st- and 2nd-order gradiometer mode, bandwidth = DC to 40 Hz (collected using a 7 channel RF SQUID
system, primary sensors were hardware 1st-order gradiometers with 5 cm baseline); (b) AEF experiment in
an unshielded environment, channel SR44, 1 kHz tone, 50 msec duration, right ear, fs = 625 samples/sec,
duration = 0.82 sec, bandwidth = DC to 70 Hz, 10 averages (collected using a 64 channel MEG system, [22,
23]).
Field (AEF) experiment was performed in the presence of large background magnetic
noise (a whole-cortex 64-channel MEG system, operated with DC SQUIDs and with
sensitivity of about 5 fT rms/√Hz). Increasing the gradiometer order from 1st to 3rd
makes the AEF signal clearly resolvable even though only 10 trials were averaged.
The last example of noise cancellation by higher-order gradiometers in an unshielded
environment using a 143-channel whole-cortex MEG system is shown in Figure 4.22
for an AEF experiment. The 3rd-order gradiometers makes the AEF signals clearly visi-
ble even in the unshielded environment, and the example shows that the same reference
system can produce a consistent performance over all channels.
Figure 4.22. AEF experiment in an unshielded environment using 143-channel MEG system, 1 kHz tone,
100 msec duration, both ears, inter-stimulus-interval = 1.5 to 2 sec (random), fs = 625 samples/sec, band-
width = DC to 40 Hz, 100 averages, file: Jan 1096:8. (a) 1st-order gradiometer; (b) 3rd-order gradiometer.
35
Figure 4.23. Comparison of noise cancellation by 3rd-order synthetic gradiometers and by adaptive sub-
traction. Unshielded environment, fs = 250 samples/sec, duration = 6.6 sec, bandwidth = DC to 3 Hz, file:
Feb 10:10 (95), magnet in position 1. (a) as collected, 1st-order hardware gradiometer, not balanced; (b)
3rd-order synthetic gradiometer; (c) adaptive noise cancellation using coefficients determined from the
data with magnet in position 1 (self-adaptation) and reference channels as for synthetic 2nd-order gra-
diometer; (d) adaptive noise cancellation using coefficients determined from data with magnet in position 2
and reference channels as for synthetic 2nd-order gradiometer; (e) adaptive noise cancellation using coef-
ficients determined from the data with magnet in position 1 (self-adaptation) and reference channels as for
synthetic 3rd-order gradiometer; (f) adaptive noise cancellation using coefficients determined from data
with magnet in position 2 and reference channels as for synthetic 3rd-order gradiometer.
36
Figure 4.24. Frequency-dependent linear input-output model. The inputs (references R) can be mutually
correlated. The sensor output, S, is modeled by summing the references (multiplied by appropriate fre-
quency-dependent transfer functions) and the random noise N.
37
tion. The problem is solved by the ordered, conditioned input-output model [103]. In
this model, the references are ordered by the magnitude of the ordinary coherence func-
tion between each reference and the sensor output, and the linear effects of the references
R 1, R 2, .... R k-1 are removed from Rk by optimum linear least squares prediction tech-
niques.
A comparison between frequency-dependent and frequency-independent adaptive noise
cancellation is shown in Figure 4.25. In this example, the sensor is a magnetometer,
and the references are 3 orthogonal magnetometers. The AEF signals were measured in a
whole-body superconducting shield [91] with relatively large residual DC fields locked
in the shield. Without any noise cancellation, the vibrations caused large noise, which
completely obscured the AEF signals, Figure 4.25.a.
When frequency-dependent adaptive noise cancellation was applied, as shown in Fig-
ure 4.25.b, practically all vibrational noise was removed, and the AEF signal was de-
tected - the magnetometer was not positioned at the AEF signal maximum, and there-
fore the signal was rather small. However, when only the frequency-independent adaptive
method was used, as shown in Figure 4.25.c, only a small part of the vibrational noise
was removed, and the AEF signal was not resolved. In this example, the frequency-de-
pendent adaptive technique was essential for good noise elimination.
In some cases, the frequency-dependent adaptive technique is more successful than
the frequency-independent technique. However, the changing noise character represents
the same problem for both techniques, and in both cases the re-determination of the co-
efficients is necessary if the noise character changes.
These methods rely on the observation that different signal or noise sources are repre-
sented by vectors pointing in different directions in the sensor space - sensor space is a
multidimensional space with dimension equal to the number of channels, N, where each
coordinate represents measurement by one channel. Various sources then can be sepa-
rated from each other by utilizing their different directionality in the sensor space.
One such method, called Signal-Space Projection (SSP) [104, 105] defines a source
as a set of current elements with identical amplitudes as a function of time. The signal
vector associated with the source is then called a component vector, and has a fixed ori-
entation in the sensor space, only its amplitude is changing with time. A set of k (k ≤
N) component vectors is selected, and using the Decomposition Theorem [106], the
sensor space is divided into two subspaces using the projection operators P|| and P ⊥: ap-
plication of the operator P|| on the measurement selects a subspace spanned by the se-
lected component vectors ( k-dimensional subspace), and the application of the operator
P ⊥ selects the other subspace, whose elements cannot be produced by any combination
of the selected component vectors (N-k dimensional subspace). The projection operators
then are used either to isolate signals from the desired subspace, or to exclude from the
signal known unwanted sources. The ability of the method to separate various sources is
a function of the angle between the source vectors in the sensor space; if the angle is
small, the sources cannot be well separated, while if the angle approaches π/2, a good
separation can be obtained.
The method is illustrated in a 2-dimensional sensor space (or 2-channel system) in
Figure 4.26, where the measurement vector m is assumed to be composed of noise n
and signal s . In Figure 4.26.a the noise is reduced by applying the operator P|| (deter-
mined relative to the signal subspace) and constructing a component of m in the direc-
tion of the "wanted" signal. Alternately, the same result can be obtained by applying P⊥
(relative to the signal subspace) and subtracting the result from the measurement. In
Figure 4.26.b the noise is eliminated by application of the operator P⊥ (determined rela-
tive to the noise subspace). This procedure is equivalent to using P|| (relative to the
noise subspace) and subtracting the result from the measurement. The simple graphical
representation in Figure 4.26 also indicates that if the angle between the noise and sig-
nal vectors is small, good separation cannot be achieved.
The selected component vectors can be estimated without using a model directly
Figure 4.26. Signal space projection, measurement m = (m 1 , m2 ), m = n + s. (a) selection of "wanted" sig-
nal by operator P|| (relative to the signal subspace); (b) rejection of "unwanted" noise by operator P⊥ (rela-
tive to the noise subspace).
39
from the measured signal m, e.g., by selecting a time when the source of interest is
known to be strong in comparison with other components of m. Or the selection can be
based on information from previous studies, or the constancy of the direction of the
measured vector, or the modeling of the measured activity, or other information.
Synthetic Aperture Magnetometry (SAM) [107, 108] maximizes the sensitivity of
the sensor array at a specified point in the brain, and minimizes the sensitivity to activi-
ties at other localities, and thus provides spatially-selective noise reduction. SAM cre-
ates a virtual sensor at the location of interest by means of a projection of the measure-
ments, and the results are normalized to yield an rms value of the activity at the selected
location. SAM's good spatial selectivity allows observation of the brain's activity with-
out the need to average the data. The principle of the method also can be illustrated
geometrically in a simplified 2-dimensional sensor space, as in Figure 4.27.
For a selected location in the brain, the measurement m is assumed to be a sum of
signal s generated by the current at the selected location and a noise n originating from
a different source (either other locations in the brain or in the body, or an environmental
noise). A forward solution model is used to determine the direction s in the sensor space
corresponding to the selected point in the 3D (real) space, as in Figure 4.27 (where for
simplicity, the possibility of different current orientations at the selected real-space loca-
tion is neglected in this discussion). When set to the highest spatial resolution, the
method rejects the noise n by projecting the measurement vector m through the opera-
tor P ⊥ (determined relative to the noise subspace), which in Figure 4.27 corresponds to
the projection into a direction d orthogonal to the noise n (1). Then the result is pro-
jected to the forward solution direction s (2) and scaled to assure unity gain (3). When
the noise vector n is close to the forward solution s , the amplitude projected through (1)
and (2) is small, and subsequent scaling may unacceptably amplify the uncorrelated sen-
sor noise. A trade-off between the spatial selectivity and the resulting noise [106, 108]
alleviates this problem and allows a smooth transition between the highest spatial reso-
lution (accompanied by higher projected noise) and the lowest spatial resolution (with
low projected noise). In the limit of the lowest spatial resolution the SAM method is
similar to the SSP method.
magnitude of the measured signal. For the purposes of the present analysis, the system
white-noise level will be determined by considering a simple spherical conductor with a
point current dipole and requiring that the accuracy of the dipole localization is better
than 0.1 cm when the peak signal amplitude is 100 fT. Further, it will be assumed that
only the sensor random noise is present, the measurement bandwidth is 100 Hz, the
number of averages is 100, and there are 150 channels distributed over a hemispherical
sensor shell with radius of 10.7 cm. These assumptions result in the required system
white-noise level of nw = 5 fT rms/√Hz - see Eq.6.1.
The noise to which the biomagnetic systems are exposed depends on the type of en-
vironment, the time of the day, and whether the system is operated shielded [86] or un-
shielded [22, 23]. In shielded environments in Figure 4.3, the 1st gradient noise is ≤ 5
fT/√Hz for frequencies greater than 1 to 4 Hz, while magnetic field noise is below that
level for frequencies greater than 9 to 70 Hz. In both cases additional noise cancellation
is required, and especially the magnetometers alone cannot be used for any meaningful
measurement. When unshielded, the noise of both device types is even larger and their
operation without noise cancellation is unthinkable.
The noise signals in Figure 4.3 can be used to calculate the signal dynamic range
and slew rate. Dynamic range may be defined as the ratio of peak-to-peak (pp) signal
amplitude to the system resolution. For a magnetometer-based system, the resolution
∂B can be related to the white noise in a 1-Hz bandwidth, δB = n w√1Hz, by ∂B = δB/2 ,
q
where q is the number of bits by which the resolution is different from the noise in a 1-
Hz bandwidth (e.g. q = 4 in [22, 23]). If Bo is the noise amplitude, the dynamic range
for magnetometers is given by
2Bo 2Bo q
DB = = 2 (5.1)
∂B δB
For gradiometers, the dynamic range has a two-fold origin. First, the common-mode
signals - see Section 4.4.3 - leak the magnetic field into the gradiometer, and the corre-
sponding gradiometer dynamic range due to the common mode effects is
2CBo q
DC
G = 2 (5.2)
δGd
where C is the common-mode vector, d is the gradiometer baseline, and δGd is the gra-
diometer noise in a 1-Hz bandwidth (in fT rms/√Hz). For the power-line noise, the gra-
diometer dynamic range due to the common-mode vector can be related to the magne-
tometer dynamic range in Eq.5.1 through the distance to the power-line source, R, by
Eq.4.1. Second, the dynamic range due to the applied gradient Go (T/m) is
2G o q
D grad = 2 (5.3)
G δG
where δG is the gradient noise in a 1-Hz bandwidth (in T/m rms/√Hz). The dynamic
range calculation for low-frequency noise is slightly more complicated. The low-fre-
quency noise pp amplitude must be determined by integration over the bandwidth of in-
terest. Assume that the low-frequency noise rms value (per √Hz) is
A
n low = (5.4)
fk
42
where k is the "slope" of the low-frequency noise (dependent on the type of environment
[70]), and the dimension of the constant A is fT rms·Hzk-0.5 . Integration of the noise
power (or the square of Eq.5.4) should be carried from the lowest frequency of interest,
f1, to f 1 + ∆f, where ∆f is the measurement bandwidth. However, the low-frequency
noise amplitude decreases rapidly with increasing frequency, and the low-frequency noise
contributions at high frequencies are negligible. Therefore, the integration can be per-
formed in the limit ∆f → ∞ to yield the total rms low-frequency noise amplitude
A f1
n rms = (5.5)
f1k 2k −1
Assume that the low-frequency noise value, nlow 1 (rms/√Hz) is determined at fre-
quency f1. Then substituting nlow 1 and f1 into Eq.5.4, the constant A can be calculated.
Further, if the pp and rms noise values are related by a factor of 5, then the pp ampli-
tude of the low-frequency noise is obtained as
f1
n pp = 5n low1 (5.6)
2k −1
The peak-to-peak amplitude in Eq.5.6 now can be used to calculate the dynamic range
of the low-frequency noise measured with the lowest frequency of interest f1, as
5n Blow1 f1
D Blowf = 2q (5.7)
δB 2k −1
5n Glow1 f1
D Glowf = 2q (5.8)
δGd 2k −1
The nB low 1 and nG low 1 in Eqs.5.7 and 5.8 correspond to the magnetometer and gradiome-
ter low-frequency noise values at frequency f1. The observed values of the constant k in
shielded rooms are usually in the range from about 2 to 3. The value k = 2 can be used
in all cases, as the results for k = 3 differ from k = 2 by less than 1 bit.
For a magnetometer, the slew rate is defined as S = 2πfBo (T/sec), where f is the
signal frequency. The slew rate in flux units is obtained from S by multiplication by
the gain α = δΦ/δB, where δΦ is the flux noise in Φ o rms/√Hz. Then, the flux slew
rate (in Φ o/sec) can be expressed in terms of the dynamic range as
A result similar to Eq.5.9 also is valid for gradiometers, provided that the gradiometer
values for δΦ, D and q are used.
The dynamic range of the noise in Figure 4.3 as a function of frequency is shown by
the hatched and crosshatched areas in Figure 5.2. Also shown are lines of constant slew
rate calculated from Eq.5.9. For shielded systems, all noise sources exhibit negligible
slew rates (< 10 Φ o/sec). However, the dynamic range of low-frequency noise is large,
up to about 21.5 bits (129.4 dB) for gradiometers and up to 26.5 bits (159.5 dB) for
magnetometers. For unshielded systems, both slew rate and dynamic range are large.
Slew rate is dominant for power-line noise (≈ 2x10 4 Φ o/sec for gradiometers and ≈ 10 6
43
Figure 5.2. Dynamic range as a function of frequency for observed noise in Figure 4.3 for gradiometer-
and magnetometer-based systems. The baseline of the hardware 1st-order gradiometer sensors is d = 5 cm.
The dynamic range of MEG signals is shown by dashed lines and shaded areas labeled by 10 fT, 100 fT,
and 1 pT. Dash-dot line is white noise in 1-Hz bandwidth, δB = 5 fT rms, q = 4. Vibrational noise as de-
scribed in connection with Figure 4.3, and field power-line amplitudes of 200 nT in unshielded and 2 pT in
shielded environments were assumed. Modified from reference [109].
Φ o/sec for magnetometers), and the dynamic ranges for the power-line and low-frequency
noise are comparable, being slightly larger for the power-line noise, up to 26.5 bits
(159.5 dB) for gradiometers and up to 31 bits (186.6 dB) for magnetometers.
Successful noise cancellation (as in Sections 4.4 and 4.5) and any other data process-
ing are possible only if the system linearity is adequate. For magnetometers, the re-
quired linearity may be defined as a ratio of the white noise in a 1-Hz bandwidth, δB, to
the p-p applied signal, and it can be expressed in terms of D and q as
δB 2q
L= = (5.10)
2Bo D
The same relationship between the L, D and q, as in Eq.5.10, also may be used for gra-
diometers.
In addition to the linearity, the delay between channels also must be sufficiently
small. The delay tolerance may be estimated from a simplified model of common-mode
vector balancing in the presence of delay between channels [70] and by requiring that the
resulting error is less that the noise in a 1-Hz bandwidth - see Figure 5.3. Consider a
simplified model of a magnetometer and a gradiometer in a uniform periodic magnetic
field. The magnetometer output is equal to the applied magnetic field, B = Bo sin(ω t)
and the gradiometer output is solely determined by the common mode vector Co and it is
time shifted by δt relative to magnetometer, G = CoBosin(ω t + ωδt). The gradiometer
output also can be expressed as
.
G = C ⋅ B+ E ⋅B (5.11)
where
C = C o cos ωδt (5.12.a)
sinωδ t
E = C o δt (5.12.b)
ωδt
44
First, assume that ωδt << 1, such that C and E are frequency independent constants
and the common-mode term in Eq.5.11 is balanced out. Then the eddy-current term in
Eq.5.11 represents the delay error. Requiring that the error be smaller than the white-
noise in 1-Hz bandwidth, δB, the condition for acceptable delay δt is
L 2q
δt ≤δt (1) = = (5.13)
πfCo πfC oD
Note that for magnetometers, Co = 1, and the δt(1) is much smaller than for gradiome-
ters. As an example, consider the power line frequency f = 60 Hz. A magnetometer-
based system with C o = 1 and L = 2 -27 would require delay of 40 psec or less, while a
gradiometer-based system with Co = 5 x 10 -3 and L = 2 -22.5 could tolerate delay as large
as 170 nsec.
The eddy-current contribution in Eq.5.11 could also be balanced out by differentiat-
ing the measured fields and subtracting the eddy current term from the gradiometer out-
put [70]. However, simple digital differentiators produces large errors and more accurate
differentiators are computationally expensive or require high sampling rates [126]. As-
suming that the differentiation problems can be overcome, then the common-mode and
eddy-current terms in Eq.5.11 could both be cancelled by constant C and E and the re-
sulting error would be caused only by the frequency dependence of the vectors C and E.
Assuming that ωδt is sufficiently small, such that expansion of Eqs.5.12 to quadratic
terms in ωδt is sufficient, and requiring that the residual error be less than δB, the con-
dition for acceptable delay becomes
2 L δt(1)
δt ≤δt (2) = = (5.14)
ω Co πf
The time δt(2) is longer than the time δt(1), provided that δt(1) < (πf)-1. If for unshielded
magnetometer-based system δt(1) = 40 psec and f = 60 Hz, then δt(2) = 461 nsec. The de-
lay would become completely unimportant if frequency dependent common-mode and
eddy-current balancing was implemented (e.g. by using coherence methods [103]). Such
methods are known to work well, but are computationally expensive.
The maximum dynamic range, slew rate and minimum acceptable inter-channel
delay are summarized in Figure 5.4. The main results are: the dynamic range of the low-
frequency noise is large (even for shielded systems), the slew rate magnitudes are large
only for power-line noise in unshielded environments, and the time-delays between
channels must be most accurately controlled for unshielded magnetometer-based sys-
tems. The time-delay is one of the more difficult parameters to satisfy. If dealys of
about > 100 nsec are considered to be readily achievable in practice, then the delay δt(1)
is outside this range for the unshielded power-line noise for both magnetometer- and
45
Figure 5.4. Summary of dynamic range, slew rate and acceptable inter-channel delay for 1st-order gra-
diometer- and magnetometer-based systems. Co = 5 x 10 -3 (gradiometers), Co = 1 (magnetometers), q = 4.
Frequencies for calculation of acceptable delay: low-frequency noise f = 0.1 Hz, power-line signals f = 60
Hz. Dynamic range values: gradiometers - shielded low-frequency noise = 21.5 bits, shielded power-line
signals = 8 bits, unshielded low-frequency noise = 25 bits, unshielded power-line signals = 26.5 bits; mag-
netometers - shielded low-frequency noise = 26.5 bits, shielded power-line signals = 14 bits, unshielded
low-frequency noise = 30.5 bits, unshielded power-line signals = 31 bits. (a) dynamic range; (b) slew rate;
(c) acceptable delay δt(1) - Eq.5.13, assuming that the eddy-current term is not balanced; (d) acceptable
delay δt(2) - Eq.5.14, assuming that accurate differentiation is available and both common-mode and eddy-
current terms are balanced. Modified from reference [109].
Figure 5.5. Operation of SQUID feedback loop (CTF Systems Inc.). (a) SQUID transfer function with pe-
riodicity of 1 Φo and ± 1 Φo locking range; (b) schematic diagram of SQUID feedback loop. The loop is
completed digitally to assure optimum channel-to-channel matching. Outputs from the SQUID MEG and
EEG channels are subjected to identical DSP processing. Modified from reference [109].
The operation of the digital loop is shown in more detail in Figure 5.6. In Figure
5.6.a the input sinusoidal signal is shown with a vertical scale in Φ o. Whenever an in-
teger multiple of Φ o is traversed, the loop lock is released, the loop is internally reset to
zero, and either a plus or minus count is added to the counter. The internal loop signal
is shown by the solid line, and the counter is shown by the gray line in Figure 5.6.b.
The internal loop signal is recorded with a resolution of 20 bits per 1 Φ o, and the
counter has a range of 12 bits, yielding a data word with a 32-bit range.
Biomagnetic systems also must be capable of handling large numbers of channels.
The presently-operated commercial whole-cortex MEG systems contain well over 100
SQUID channels - see Section 3 - and in addition, CTF's versions also are equipped
with references, ranging from a few to as many as 29 channels, auxiliary A/D and D/A
channels, and in many cases with up to 64 EEG channels.
An example of the CTF electronics architecture for a large number of channels is
shown in Fig. 5.7 [71]. The electronics is organized into banks of plug-in units, one
unit in each bank devoted to data communications, and twenty units for up to 20
SQUID channels. Variable numbers of banks can be connected via the internal bus to al-
low for variable numbers of SQUID channels. Usually, some plug-in units also are re-
quired to perform A/D and D/A functions. There is also a bank which allows connection
of EEG channels and a bank with DSP processors for real-time data processing before
transmission to a host computer. A block diagram of the complete electronic system is
shown in Figure 5.8.
Figure 5.6. Operation of digital SQUID electronics (CTF Systems Inc.). (a) input sinusoidal signal; (b) solid
line - signal within the locking range of ± 1Φo , gray line - counter signal (counts of flux transitions along the
transfer function); (c) 20-bit digitization of ± 1 Φo range and 12-bit counter, merged to yield 32-bit data
word.
47
Figure 5.7. Architecture of a biomagnetic data collection system including optional EEG (or ECG) channels
(CTF Systems Inc.). Modified from reference [109].
Figure 5.8. Block diagram of the electronics and data collection for an MEG/EEG system (CTF Systems
Inc.).
48
ments, as well as large numbers of channels, was described. It was shown that magne-
tometer-based systems are subject to far more stringent requirements than gradiometer-
based systems and that for magnetometer-based systems some specifications due to the
observed environmental noise might be difficult to meet.
Failure to meet the specifications derived in this section would have the following
consequences: (a) dynamic range for low-frequency noise - the system could still be op-
erated, but it would have to be reset whenever the limit of the dynamic range was
reached (or, in other words, the dynamic range would have to be extended); (b) dynamic
range for power-line signals - the system would not be functional because the overflows
caused by the dynamic range at the power-line frequencies cannot be handled well in
practice; (c) linearity requirement (both for low-frequency and power-line noise) - the
data quality would be degraded and accurate noise cancellation or other data processing
would not be possible; (d) delay between channels - good noise cancellation would not
be possible (the inter-channel time delays could potentially be corrected if the delay were
constant and could be measured accurately).
Note that the system requirements also apply to all signal processing functions,
e.g., filtering, balancing, etc. The linearity and inter-channel matching requirements
drive the choice of digital feedback loops and digital filters, as the errors between ana-
logue circuits can easily exceed the specified tolerances.
The flux transformers of primary biomagnetic sensors can have a variety of forms:
magnetometers, 1st-order hardware gradiometers, and in some cases even higher-order
gradiometers. Magnetic field is a vector, and therefore, magnetometers can be oriented in
three orthogonal directions. First gradient is a 9-component tensor with 5 linearly inde-
pendent components [70] and usually only the following components of this tensor are
used for biomagnetic detection: radial gradient of radial magnetic field, radial gradient of
tangential magnetic field, and tangential gradient of radial magnetic field. Tensors of
higher gradients have 3 + 2k independent components, where k is the gradient order.
Graphical representation of some primary sensors is shown in Figure 6.1.
MEG and biomagnetism in general are usually discussed in terms of magnetic fields.
Yet, in most systems the detection flux transformers are gradiometers. When a uniform
magnetic field is applied to a gradiometer, Figure 6.2.a, the currents generated in the
two gradiometer coils are equal, and because the coils are connected in opposing direc-
Figure 6.1. Examples of some possible configurations of primary sensor flux transformers.
49
Figure 6.2. Measurement of biomagnetic fields by gradiometers. (a) uniform magnetic field produces zero
gradiometer output; (b) gradiometer can be positioned either in near field or far field.
tions, they cancel, and the net output is zero. However, in biomagnetic practice, the
fields are not uniform. This is illustrated in Figure 6.2.b, where the source is located at
the origin. For a gradiometer positioned in the near field, the variation of the gradient
over the gradiometer dimension (baseline) is large, the coil closer to the source sees a
much larger field than the more distant coil, and the gradiometer output is roughly pro-
portional to the magnetic field at the closest coil (in reality, the effect of the more dis-
tant coil is not neglected and is always included in any data analysis). In contrast, a gra-
diometer positioned in the far field is exposed to a nearly uniform gradient, and its out-
put is proportional to the spatial derivative of the magnetic field. If the primary sensors
are planar gradiometers, the detected signal is a tangential difference of fields (approxi-
mately equal to the tangential gradient), but even that information can be converted into
a radial field [119].
In this section, various primary biomagnetic sensor configurations will be
discussed. First, the method of sensor comparison will be reviewed in Section 6.1,
radial and vector magnetometers will be compared in Section 6.2, and in Section 6.3,
radial gradiometers will be optimized (baseline length optimization). Radial and planar
gradiometers will be compared in Section 6.4, and the results of the sensor discussion
will be summarized in Section 6.5.
Sensors will be compared by evaluating their signal-to-noise (S/N) ratio and their capa-
bility to accurately reconstruct one or several equivalent current dipoles in a spherical
conducting medium. The reconstruction accuracy will be determined by Monte-Carlo
simulations in the presence of either random or correlated noise, Figure 6.3. Generally,
the reconstructed dipole positions produce an ellipsoidal cluster [111] with different
standard deviations (σ1, σ2, σ3) along the three orthogonal directions. In the present dis-
cussion, the rms value of the three standard deviations, σ = √[(σ12+ σ22+ σ32)/3], will be
used as a measure of the dipole reconstruction accuracy.
The Monte-Carlo simulations were performed as a function of the number of chan-
nels, N, and for different rms noise levels. The N channels were distributed over N sites
for magnetometers and radial gradiometers, over N/2 sites for planar gradiometers (2 or-
thogonal sensors per site), and over N/3 sites for vector magnetometers (3 orthogonal
sensors per site). The sites were assumed to be uniformly distributed over a hemispheri-
cal sensor shell, Figure 6.4.a. The resulting σ as a function of the number of channels
50
Figure 6.3. Biomagnetic sensors are exposed to random and correlated noise. (a) random noise sources
usually originate in the SQUID sensors or electronics; (b) correlated noise can originate either within the
brain or can be caused by environmental sources (e.g., moving dipoles, current wires, etc.).
is shown schematically in Figures 6.4.b and d for random and correlated noise. For ran-
dom noise, the σ is decreasing with increasing number of channels as 1/√N, and the ran-
dom noise is spatially averaged. For correlated noise, the resulting σ becomes independ-
ent of the number of channels for large channel density. For correlated brain noise [24]
the limiting σ is reached for N > 100 channels, or for mean spacing between the chan-
nels of about 3 cm. This limiting channel density qualitatively corresponds to inter-
channel spacing of the order of the noise correlation distance. Further increase of the
channel density does not improve the S/N ratio, and the limiting error is reached.
The simulated results scale accurately with noise, and the standard deviation σ nor-
malized by noise depends only on the number of channels, as in Figures 6.4.c and e.
The scaling by noise also can be derived from expansion of the dipole equations in the
vicinity of the true dipole position [112]. It has been found by examining the simula-
tion results that the dependence of the standard deviation σ on the dipole depth and the
sensor type can be explained approximately by inverse dependence on the peak signal
[113], Sig peak . Then the standard deviations for random, σr, and correlated, σc , noise can
be expressed as
Figure 6.4. Dipole reconstruction accuracy (standard deviation, σ, of the cluster of reconstructed dipoles).
(a) geometry for simulation of the effect of noise on dipole reconstruction accuracy, r = 10.7 cm; (b, c)
random noise; (d, e) correlated noise; (b, d) σ as a function of the number of channels, N, for noise levels
of 15, 30 and 45 fT rms. Note the saturated (limiting) σ for larger number of channels (N > 100, or inter-
sensor spacing of about 3 cm) for correlated noise; (c, e) σ/noiserms as a function of N.
51
Figure 6.5. Demonstration of the effect of various normalizations on the dipole reconstruction error for
random noise and radial magnetometers for dipole distances from the sphere center of a = 5 and 7 cm.
Random noise levels are 15, 30 and 45 fT rms, and the channels are uniformly distributed over the hemi-
spherical sensor shell with radius r = 10.7 cm, a = 5 and 7 cm, dipole magnitude q = 10 nA·m.
d cnoise rms dc
σc ≈ = (6.2)
Sig peak ( S/N )
where it has been assumed that for the correlated noise the number of channels is N >
100 (for which σc reaches the limiting error in Figure 6.4.e). The universal constants βr
and dc are βr ≈ 0.44 unitless and dc ≈ 2.5 cm (or for a hemispherical sensor shell dr ≈ 12
cm), n is the channel density (cm -2), and the signal-to-noise ratio, S/N, means the ratio
of the peak signal, Sigpeak , to the rms noise, noise rms. The accuracy of Eq.6.1 is illus-
trated in Figure 6.5 for radial magnetometers. In Figure 6.5.a the standard deviation, as
Figure 6.6. Dipole localization accuracy for random noise normalized by noise and by peak signal as a
function of the number of channels and different sensor types. Dipole distances from the sphere center are
a = 5 and 7 cm, sensor shell is hemispherical with radius r = 10.7 cm, and random noise levels are 15, 30
and 45 fT rms. Sensor types are: radial magnetometers, vector magnetometers, radial gradiometers (d = 5
cm), and planar gradiometers (d = 1.65 cm). The N channels are distributed over N sites for radial magne-
tometers and gradiometers, over N/3 sites (3 orthogonal sensors per site) for vector magnetometers, and
over N/2 sites (2 orthogonal sensors per site) for planar gradiometers and the site distribution is uniform.
52
simulated, shows the dependence on noise and dipole distance from the sphere center, a.
In Figure 6.5.b, the standard deviation normalized by noise is shown. The plots, for a
larger number of channels, are independent of the noise level and depend only on dipole
position, a. In Figure 6.5.c the complete normalization of σ by noise and by peak sig-
nal (quantity βr/√n) is shown. For N > 80 (which corresponds to inter-sensor spacing of
about 3.4 cm) the curves for all noise values and dipole positions coincide.
The accuracy of Eq.6.1 for different types of sensors (e.g., radial magnetometers,
vector magnetometers, radial gradiometers and planar gradiometers) is shown in Figure
6.6. For N > 80 the plots for all noise levels, dipole positions a = 5 and 7 cm and all
sensor types are very close together, and indicate that Eq.6.1 is reasonably universal.
Similar results were obtained also for Eq.6.2.
For random noise, the standard deviation normalized by noise for radial and vector mag-
netometers was computed as a function of the number of channels by simulation and
also by using expansion of dipole equations [112]. The dipole source was located on the
sensor array axis at two positions, a = 5 cm (deep below the sphere surface) and a = 7
cm (shallow, or close to the sphere surface). The results are plotted in Figure 6.7 and
show that for both investigated dipole positions, the radial magnetometers give better
results (i.e., smaller standard deviation σ).
The comparison of dipole reconstruction accuracy when the dipole is not on the axis
of the sensor array is shown in Figure 6.8 for random and correlated noise. For random
noise, Figure 6.8.b, the radial magnetometers give better results (smaller σ) for dipoles
positioned even outside the sensor array (up to about 2.5 rad from the vertex for a hemi-
spherical sensor shell). For correlated noise, Figure 6.8.c, the radial and vector magne-
tometers give similar results for dipoles located well within the sensor array (up to
Figure 6.7. Comparison of standard deviation normalized by rms noise magnitude for radial and vector
magnetometers as a function of the number of channels for Monte-Carlo simulation and direct calculation
by linearisation of dipole equations. Random noise, 15, 30 and 45 fT rms, a = 7 and 5 cm, hemispherical
sensor shell r = 10.7 cm, dipole moment q = 10 nA·m, one radial magnetometer per site, three vector mag-
netometers per site, sites uniformly distributed over the sensor shell area. Solid and dashed lines - calcula-
tion from linearisation of dipole equations for radial and vector magnetometers, symbols - Monte-Carlo
simulation.
53
Figure 6.8. Comparison of radial and vector magnetometers at the sensor array edge, plot of σ√N for ran-
dom and σ for correlated noise as a function of the angle of dipole position from vertex. Each value of σ
was determined by 200 Monte-Carlo simulations and the correlated noise was simulated using 1,000 ran-
domly distributed, random amplitude dipoles in the model sphere with radius rbrain = 8 cm. Hemispherical
sensor shell, r = 10.7 cm, a = 7 cm, position 1 corresponds to 18 deg from vertex, and all subsequent posi-
tions are in 18 deg increments, position 5 corresponds to the dipole at the sensor array edge. Number of ra-
dial magnetometers used N = 97, 142, 190, 289, 586, and the number of vector magnetometers used 96, 141,
291, 570 (number of vector magnetometer sites 32, 47, 97, 190), lines in (b) and (c) correspond to constant
number of channels, solid lines - radial magnetometers, dashed lines - vector magnetometers. Thin vertical
dashed line corresponds to sensor array edge. (a) geometry for the calculation; (b) results for random noise
of 30 fT rms; (c) results for correlated noise of 30.4 fT rms for radial magnetometers and 22.5 fT rms for
vector magnetometers (mean rms value over all three vector components), the two different noise values
correspond to the same random distribution of dipoles.
about 1 rad from the vertex), but for the dipoles close to or outside the sensor array
edge, the vector magnetometers give better results (smaller σ). For correlated noise, the
vector magnetometers are better at "extrapolating" when the dipole is far from the sensor
array axis.
The results in Figure 6.8 are independent of the number of channels (each line corre-
sponds to a constant number of channels, and the lines for different numbers of channels
yield roughly the same results).
It was shown in Section 4 that, even within shielded rooms, the detection of biomag-
netic fields is not possible with magnetometers, and noise cancellation using references
must be performed. The references are used to remove noise from magnetometers and
can lead to 1st- or higher-order noise cancellation. Similarly, the primary sensors could
be 1st-order hardware gradiometers, and references are used to provide higher-order noise
cancellation. In both cases, the noise cancellation is performed with either gradiometer
or adaptive coefficients. Both magnetometer and hardware 1st-order gradiometer sensors,
after the first level of noise cancellation, represent a 1st-order system (either gradiometer
or adaptive), and this section will discuss the optimization of the primary 1st-order base-
line, defined as in Figure 6.9.
The baseline optimization will be developed first for hardware gradiometers, and then
it will be extended to synthetic gradiometers and adaptive systems. The baseline will be
optimized by considering an interplay between the detected signal strength and the device
sensitivity to environmental noise.
It was shown in Section 4, that even within shielded rooms, the environmental
noise is not negligible and affects the SQUID detectors. The detected noise is clearly of
54
Figure 6.9. Distinction between magnetometer- and gradiometer-based systems, and between systems with
different baseline lengths. (a) magnetometer-based system with sensors close to a helmet surface and ref-
erence at some distance away; the baselines are quite long, and both the 1st-order and high-order baselines
are different for different channels; (b, c) hardware 1st-order gradiometer-based systems, the 1st-order
gradiometer baseline is the same for all channels, only high-order gradiometer baselines are different for
different channels; (b) longer 1st-order baseline; (c) shorter 1st-order baseline.
environmental origin, because it depends on the time of the day and can be reduced by
high-order gradiometers (Figures 4.5.b and 4.19). A simple representation of SQUID de-
tector noise is shown in Fig.6.10.a. The SQUID noise consists of white-noise (of in-
strumental origin) and of low-frequency noise (of environmental origin), Eq.5.4. The
dependence of the low-frequency noise on baseline length - see Section 4.4 - can be ex-
pressed as
A = A od/d o (6.3)
where the constant A corresponds to baseline d, and constant Ao, to baseline do. The on-
set of low-frequency noise can be defined as a point where the white noise is equal to the
low-frequency noise, nw = n low. If f o is the low-frequency noise onset corresponding to
baseline do and fod to baseline d, then
Ao d A d
f od = k =k = fo k (6.4.a)
nw do nw do
and
Ao
fo = k (6.4.b)
nw
Using Eqs.5.4, 6.3 and 6.4, the low-frequency noise may be expressed as
d fo k
n low = n w (6.5)
do f
The total noise detected by the SQUID flux transformer is a combination of white and
low-frequency noise, nT 2 = n low2 + n w2, and
n2 d 2 f 2k
n 2T = w 1+ o (6.6)
N ave do f
55
Figure 6.10. Noise of a SQUID detector exposed to low-frequency environmental noise. fo = 2 Hz, d o = 5
cm, nw = 5 fT rms/√Hz, N ave = 1, k = 1, ∆f = 100 Hz. (a) representation of low-frequency and white noise
on log-log plot; (b) total noise as a function of baseline length, f1 = 0.1 Hz; (c) total noise as a function of f1
for d = 5 and 20 cm.
where it has been assumed that the measured signal was averaged Nave times. Usually,
the measurement is performed in a frequency range from f1 to f 1 + ∆f. The total rms
noise in this bandwidth is obtained by integration of Eq.6.6
f 1 +∆ f nw
2
f1 d f o
2k f 2k −1
n rms = ∫
2
n T ⋅ df = ∆f + 1− 1
(6.7)
f1 N ave ( − 1) d o f1
2k f1 +∆f
The detected noise now depends on the noise parameters nw, f o, f 1, ∆f, k, and base-
lines, d and do. The behavior of Eq.6.7 as a function of baseline, d, and the lowest fre-
quency of interest, f1, is shown graphically in Figures 6.10.b and c. The total noise is
monotonically increasing as a function of baseline and is decreasing as the lowest fre-
quency of interest is increasing. If the lowest frequency of interest is larger than the on-
set of low frequency noise, f1 > f od, then the system operation is completely in the
white-noise region, Eq.6.7 simplifies to nrms = n w√(∆f/Nave ) and there is no dependency
of the detected noise on b or f1 (such situation is usually not observed in practice).
The dependence of signal strength on baseline can be determined by considering the
peak signal magnitude of a single current dipole source in a spherical medium, as in
Figure 6.11.a [63]. The peak signal was calculated as a function of radial gradiometer
baseline, and the results were normalized by the peak magnetometer signal strength. The
results are shown for different dipole distances from the sphere center in Figure 6.11.b.
As the baseline increases, the normalized gradiometer signal strength approaches 1
(magnetometer signal strength). The signal strength increase with baseline is fastest for
Figure 6.11. Determination of radial gradiometer peak signal as a function of baseline. (a) calculation ge-
ometry; (b) gradiometer peak signal normalized by magnetometer (d→ ∞) peak signal. Adapted from ref-
erence [114].
56
Figure 6.12. Comparison of dipole reconstruction accuracy, σ, for random and correlated brain noise as a
function of baseline length. Sensor shell radius r = 11 cm, dipole moment magnitude q = 20 nA·m, nw = 5 fT
rms/√Hz, a = 4, 5.5, and 7 cm. (a) random noise, number of sensors N = 150, bandwidth = 100 Hz, Nave = 1,
dr = 12 cm; (b) correlated noise, bandwidth = 100 Hz, Nave = 100, d c = 2.5 cm, brain noise is given as a
function of baseline as in Figure 6.23 at a distance from the head of 1.9 cm.
shallow dipoles and slowest for deepest dipoles. The curve marked by → 0 in Figure
6.11.b corresponds to the limit where the current dipole approaches the sphere center.
Alternatively, the dependence of signal strength on baseline can be expressed in
terms of the dipole reconstruction accuracy, σ - see Section 6.1. For random noise, σ as
a function of baseline for different dipole distances from the sphere center a = 4, 5.5, and
7 cm is shown in Figure 6.12.a. As the baseline increases, the signal strength also in-
creases (as in Figure 6.11.b), and the dipole reconstruction accuracy improves, i.e., σ
decreases. For correlated noise, Figure 6.12.b, σ is nearly independent of the baseline, as
the calculations for the correlated noise were performed in the region of "limiting error"
in Figure 6.4.e.
The signal strength also has been represented by the rms value of the signal com-
puted over all channels. This representation was found to be close to the peak signal re-
sults and since the peak signal calculations are simpler, they were used for the present
demonstration of the baseline optimization.
The principle of baseline optimization is shown in Figure 6.13 [114]. The detected
noise and signal magnitudes both increase with increasing baseline, however, because of
their functional differences, the S/N ratio peaks at a certain "optimal" baseline.
Examples of the actual S/N ratio curves are shown in Figure 6.14 for shallow and
deep dipoles for different onsets of low-frequency noise, fo. As fo decreases, the positions
of the S/N ratio maxima and therefore the optimal baselines increase. For no low-fre-
quency noise (fo = 0 Hz), the optimal baseline is infinitely long. The optimization
curves are only weakly dependent on the dipole depth, and their shapes are independent
of the number of averages, Nave .
Figure 6.13. Principle of baseline optimization for hardware 1st-order gradiometers. (a) dependence of
signal strength on baseline; (b) dependence of detected noise on baseline; (c) dependence of S/N ratio on
baseline. Adapted from reference [114].
57
Figure 6.14. Plots of S/N ratio as a function of baseline length. do = 5 cm, k = 1, nw = 5 fT rms/√Hz, f1 = 0.1
Hz, ∆f = 100 Hz, Nave = 1, r = 10.7 cm, q = 10 nA·m. (a) shallow dipole, a = 7 cm; (b) deep dipole, a = 5
cm.
The general baseline optimization method is shown in Figure 6.15. The S/N ratio
surface as a function of fo and d is constructed for given noise parameters and the required
collection frequencies f1 and ∆f (Figure 6.15.c). Then, for a given fo, the maximum S/N
ratio is found, and the corresponding optimal baseline is determined. This procedure
yields short baselines for large values of fo and longer baselines for small fo's.
The above baseline optimization was developed for 1st-order hardware radial gra-
diometers. Hardware gradiometers always reduce the detected signal magnitude, while the
synthetic gradiometers may reduce or increase it. A simple model of this behavior is
shown in Figure 6.16 where the synthetic gradiometers consist of magnetometer sensors
and a vector magnetometer reference located at a fixed position, e.g., above the region of
negative signal, as in Figure 6.16.b. The reference also will detect the negative signal
and depending on whether it is subtracted from the magnetometers in the negative or
positive regions, the total synthetic gradiometer output will be either reduced or in-
creased. The above explanation is grossly simplified, as in reality the effective reference
orientation will be different for different magnetometers. However, the character of the
synthetic gradiometer behavior is correctly captured by this simple model.
Therefore, for synthetic 1st-order gradiometers, depending on the source position
relative to the reference, the signal strength is the same or larger than for the hardware
gradiometers (and the hardware gradiometers represent the worst-case limit of the syn-
thetic gradiometers). Since the white-noise levels and the environmental noise rejection
Figure 6.15. Method of baseline optimization. (a) small onset of low-frequency noise, fo ; (b) large onset of
low-frequency noise, fo ; (c) plot of S/N ratio as a function of low-frequency noise onset, fo , and baseline, d.
Solid line indicates the maximum of S/N ratio. Adapted from reference [114].
58
Figure 6.16. Simplified model demonstrating the effect of a synthetic gradiometer on the detected signal. (a)
geometry of a 1st-order synthetic gradiometer; (b) magnetometer (solid line) and synthetic gradiometer
(dashed line) signals. Adapted from reference [114].
are similar for both types of gradiometers, baseline optimization for synthetic gradiome-
ters leads to the same or shorter baselines than for the hardware gradiometers. Since the
optimization is most important in situations where the signals are weakest, the syn-
thetic gradiometers should be optimized in their worst-case limit, which is the hardware
gradiometer. Similar arguments also apply to adaptive systems.
Optimization of the primary sensor baseline for higher-order synthetic gradiometers
is similar to the procedures for the 1st-order gradiometers. The effect of higher-order gra-
diometer references can generally be neglected [80], and these gradiometers are then op-
timized as 1st-order gradiometers, except that the noise parameters fo, k, n w correspond
to the noise detected by the high-order gradiometers.
The noise character, as discussed above, also was verified experimentally. A standard
64-channel MEG system [80] with 1st-order hardware gradiometers as primary sensors (d
= 5 cm) has been equipped with high-sensitivity reference vector magnetometers and
four additional magnetometer-based synthetic noise cancellation channels (as in Figure
4.7.a) with baselines of d = 10.7, 18.3, 24.4, and 32.9 cm. These additional channels
were balanced as synthetic 1st- or 2nd-order gradiometers (fixed coefficients). Their level
of balance was evaluated using a rotating magnet [115] and it was found to be similar to
that of the standard gradiometer-based channels. Noise spectra of the synthetic gradiome-
ters with different baselines, as measured in a shielded room, are shown in Figure 6.17.
The environmental low-frequency noise in Figure 6.17 increases with increasing
baseline. The noise parameters were extracted at 0.5 Hz and are shown in Figure 6.18 as
Figure 6.17. Spectra of hardware and synthetic gradiometers with different primary baselines, collected in
a shielded room. Spectra of 6 gradiometer-based channels with the primary hardware baseline d = 5 cm
and two magnetometer-based channels with primary synthetic gradiometer baselines of d = 18.3 and 32.9
cm are shown. (a) 1st-order hardware and synthetic gradiometers; (b) 2nd-order synthetic gradiometers.
Adapted from reference [114].
59
Figure 6.18. Low-frequency noise parameters measured at frequency f = 0.5 Hz from noise spectra as in
Figure 6.16. (a, b) low-frequency noise slope k; (c, d) constant A (fT·Hzk-0.5 ); (e, f) low-frequency noise
onset f od (Hz). Adapted from reference [114].
Baseline optimization for the Amsterdam site (last row in Table 6.1) is shown in
Figure 6.19. The optimal baseline is plotted as a function of fo and f1, and the hatched
band corresponds to the appropriate range of observed fo's. If ∆f = 100 Hz and the lowest
frequency of interest were f1 ≈ 1 Hz, then the optimum baseline would be in the range of
about 5 - 8 cm and 7 - 10 cm for 1st- and 2nd-order gradiometers, respectively. How-
ever, because of the need to correctly determine the data offset, the lowest frequency of
interest, even for evoked responses, is usually less than 1 Hz. Then considering f1 in the
60
Figure 6.19. Plot of optimum baseline as a function of low-frequency noise onset fo and the minimum fre-
quency of interest, f1 . r = 10.7 cm, a = 4 cm, q = 50 nA·m, nw = 5 fT rms/√Hz, d o = 5 cm, ∆f = 20 Hz, N ave
= 1. Regions 1: d > 15 cm, 2: 8 cm < d ≤ 15 cm, 3: 5 cm < d ≤ 8 cm, 4: d ≤ 5 cm. (a) k = 1.85, 2nd-order
gradiometer; (b) k = 2.6, 1st-order gradiometer. Adapted from reference [114].
range from 0.2 to 0.5 Hz, the optimal primary sensor baseline is about 1.5 - 3.5 cm and
2 - 6 cm for 1st- and 2nd-order gradiometers, respectively. Increasing the gradiometer or-
der increases the optimal baseline, but even for 2nd-order gradiometers, the optimal
baseline remains relatively short. For 1st-order gradiometers, the short optimal baseline
also was suggested by ter Brake [116].
The baseline optimization for other MEG systems in Table 6.1 results in similarly
short baselines. The optimization for 3rd-order synthetic gradiometers would lead to
somewhat longer primary baselines (because the environmental noise is significantly re-
duced). However, such systems have not been discussed because long primary baselines
would cause the 3rd-order gradiometers to have increased white-noise levels.
Comment on the selection of the lowest frequency of interest: The lowest frequency
of interest depends on the measurement paradigm and will differ widely in different stud-
ies. However, the measurement of evoked signals is probably one of the more univer-
sally performed experiments and can serve as a model for discussion of the lowest fre-
quency of interest. The width of the evoked signal is about 50 msec, Figure 6.20 (e.g.,
[22, 23]), and it corresponds to relatively high frequencies of about 3 Hz. However, it is
not sufficient to determine the "high frequency" peak with good S/N ratio, it is also
necessary to determine the signal offset level with good S/N ratio. The offset is deter-
mined from the pre-stimulus time interval, which is usually in the range from 0.2 to
0.5 sec and therefore corresponds to a frequency range from 0.3 to 0.8 Hz. Thus, even
Figure 6.20. Stylized MEG evoked signal. Signal width is about 50 msec (≈ 3 Hz), and signal offset is de-
termined on a time interval ∆t, roughly equal to the pre-stimulus interval.
61
Comparison of radial and planar gradiometer signal strengths as a function of the source
depth is shown in Figure 6.21. An equivalent current dipole was positioned at a distance
a from the center of a conducting sphere and magnetometer, radial and planar gradiome-
ters were each located at a position of their maximum signal strength, Figure 6.21.a. In
order to remove the steep dependence on the depth, the results were normalized by the
peak magnetometer signal. Unless otherwise stated, all calculations were performed for
"point" detectors, i.e., the radii of all flux transformer coils were assumed to be zero.
The plot of the normalized peak signals as a function of dipole depth is shown in
Figure 6.21.b. The radial magnetometer corresponds to a horizontal line at the value of
1 (because the results were normalized by magnetometer peak signal). For radial gra-
diometers, as the baseline length decreases, the signal strength also decreases (similar to
Figure 6.11.b). The planar gradiometers show a strong signal for superficial sources and
fast signal decrease when the current source is deeper. This rapid loss of signal strength
also is reflected in the dipole reconstruction accuracy, σ, in Figure 6.22, when random
noise is considered. For the most frequently-used radial and planar gradiometers, σ is
smaller (better reconstruction) for radial than for the planar gradiometers.
Figure 6.21. Comparison of detected signal strengths for various types of sensors. (a) geometry of the cal-
culation, for each dipole position, each sensor is assumed to be located at a position of its maximum signal, r
= 11 cm, rhead =9.1 cm; (b) peak signals normalized by magnetometer peak signal as a function of normal-
ized dipole position, a/r. Dashed line - magnetometer, solid lines - radial gradiometers with baselines of 8, 5,
3, and 1.65 cm, chain line - planar gradiometer with baseline of 1.65 cm.
62
Figure 6.22. Comparison of dipole localization accuracy, σ/noiserms , for radial and planar gradiometers and
for random noise. Sensor shell radius r = 10.7 cm, dipole magnitude, q = 10 nA·m, radial gradiometer base-
line d = 5 cm, planar gradiometer baseline d = 1.65 cm, radial gradiometers are uniformly distributed on the
sensor shell surface at N sites, planar gradiometers are uniformly distributed at N/2 sites (two orthogonal
channels per site), noise values 15, 30, and 45 fT rms. Solid lines - radial gradiometers, dashed lines - planar
gradiometers. (a) a = 7 cm, shallow dipole; (b) a = 5 cm, deep dipole.
A similar result also was obtained when the sensor array information content [117]
was considered. The channel capacity of 122-channel planar gradiometer system was
compared with an array of a 122-channel radial gradiometer system. It was found that the
channel capacity grows approximately linearly with the number of channels, with the
planar 122-channel array giving 10% less bits than the radial array [24].
When correlated brain noise is considered, it is found that the noise magnitude de-
pends on the sensor type (radial or planar) and on the gradiometer baseline. Results from
[24], recomputed for a larger range of baselines, are shown in Figure 6.23, where the
correlated brain noise is plotted as a function of distance from the surface of the head.
The magnetometer (radial gradiometer with d → ∞) detects the largest noise, and as the
radial gradiometer baseline length decreases, the detected brain noise also decreases. The
brain noise rms value detected by the planar gradiometer with d = 1.65 cm is about 1/2
of that detected by the radial gradiometer with baseline d = 5 cm.
The differences between the dipole reconstruction accuracies, σ, for random and
correlated noise and for different types of detectors are shown in Figure 6.24. The
Figure 6.23. Simulated magnitudes of the correlated brain noise for magnetometers and radial and planar
gradiometers with different baselines. rhead = 9.1 cm, rbrain = 8 cm, position of the measurement point varied
in the range of r = 10.7 to 12.7 cm. Results in this figure were computed as in [24] and expanded to a larger
range of baselines. (a) geometry of computation; (b) simulated brain noise as a function of distance from
the scalp surface. Solid lines - radial gradiometers, dashed line - planar gradiometer.
63
Figure 6.24. Standard deviation of the dipole reconstruction cluster as a function of dipole position for ran-
dom and correlated noise and for magnetometer, radial gradiometers with 5 and 1.65 cm baselines, and
planar gradiometer with 1.65 cm baseline. Sensor shell radius r = 11 cm, rhead = 9.1 cm, r brain = 8 cm. Mag-
netometer and radial gradiometer coil areas 1.77 x 1.77 cm2 = 3.14 cm 2 , planar gradiometer coil areas 1.65
x 3.3 cm2 = 5.44 cm 2 . (a) geometry for computation; (b) random noise, nw = 5 fT rms/√Hz, bandwidth = 100
Hz, N ave = 1, noise rms = 50 fT rms, N = 150 channels, dr = 12 cm; (c) correlated noise, bandwidth = 100 Hz,
Nave = 100, d c = 2.5 cm, and brain noise is dependent on the type of flux transformer, as in Figure 6.23.
Considering the distance from the head of 1.9 cm, nbrain ≈ 26 fT rms/√Hz for radial gradiometer with d = 5
cm and n brain ≈ 12.5 fT rms/√Hz for planar gradiometer with d = 1.65 cm.
Figure 6.25. Ratio of signal strengths of radial to planar gradiometers for quadrupole sources. Radial gra-
diometer baseline d = 5 cm, planar gradiometer baseline d = 1.65 cm, each device is positioned at a point of
its maximum signal. Distances between the quadrupole sources e = 0.5, 1, 2, 4, and 6 cm, dipole moments q
= 10 nA·m, radius of sensor shell r = 10.7 cm. Thin horizontal line corresponds to the case when maximum
signal strengths of radial and planar gradiometers are equal, thick solid line corresponds to a single dipole
ratio. (a) two antiparallel dipoles "in-line"; (b) two antiparallel dipoles "side-by-side".
Figure 6.26. Difference between radial and planar gradiometers for linearly-extended sources as a func-
tion of source length. Radius of sensor shell r = 10.7 cm, integrated dipole moment = 10 nA·m, drad = 5 cm,
dpln = 1.65 cm. Source is a segment of an arc filled with an uniform density of dipoles perpendicular to the
arc plane. (a) geometry of the calculation; (b) maximum signal as a function of the source length for dipole
positions a = 4 (deep) to 8 (shallow) cm. Solid lines - radial gradiometers, dashed lines - planar gradiome-
ters; (c) ratio of radial to planar gradiometer maximum signal strength for a = 4 to 8 cm.
65
Figure 6.27. Relative S/N ratio for radial and planar gradiometers as a function of arc length and depth be-
low the scalp for correlated noise. Radius of sensor shell r = 10.7 cm, drad = 1.65, 3, 5, 8 cm and magne-
tometer (d → ∞), d pln = 1.65 cm, all sensors are assumed to have "point" coils. The curves are normalized
by radial gradiometer with baseline drad = 5 cm for all depths. (a) short source with length of 0.5 cm, similar
result as presented in [24]; (b) source length 6 cm; (c) source length 10 cm.
and is similar to that obtained for an equivalent current dipole in [24] or in Figure
6.24.c. However, as the source length increases, the advantage of the planar gradiometer
decreases. For a source length of 6 cm and superficial sources, the planar gradiometer is
comparable to the radial gradiometer, and for sources deeper than about 2.8 cm, the pla-
nar gradiometer becomes worse, Figure 6.27.b. For longer sources, the performance of
radial gradiometers with any baseline is better than that of planar gradiometers for all
source depths, Figure 6.27.c.
In summary: the signal strength detected by planar gradiometers is smaller than that
of radial gradiometers or magnetometers (except for very superficial sources, where the
planar gradiometer signal strength is larger). For situations where spontaneous brain
signals are investigated and the noise is the random instrumental noise, the radial gra-
diometers give better results, and in situations where the evoked signals are studied and
averaging is performed, the planar gradiometers give better results. For quadrupole
sources, the exact configuration of the source determines whether planar or radial gra-
diometers are favored. For extended sources, the planar gradiometers are at a disadvan-
tage, and even in the case of correlated noise, for extended sources the radial gradiometers
give better results than the planar gradiometers.
It was assumed in this section that all sensor arrays have the same number of channels
regardless of the sensor type. However, the sensor arrays may have different numbers of
sensor sites, with multiple sensors at one site, e.g., a 150-channel system populated by
radial gradiometers will have 150 sites, by planar gradiometers 75 sites, and by vector
magnetometers 50 sites.
Radial and vector magnetometers: The radial magnetometers always give better re-
sults when random noise is considered. For correlated noise, radial and vector magne-
tometers are comparable when the sources are well within the sensor array, while vector
magnetometers give better results when the sources are close to or outside the sensor ar-
ray edge. Use of vector magnetometers (which also measure tangential field compo-
nents) increases the sensitivity to volume currents and thus to the details of the model.
66
Radial gradiometers (or magnetometers with remote reference): Increasing the base-
line increases the magnitude of the detected signal and also increases the sensitivity to
environmental noise. As a result, there is an optimum baseline at which the S/N ratio
is largest. Analysis of the noise measured at the existing shielded MEG sites shows that
the optimum baseline is relatively short (1.5 to 10 cm). In addition, long baseline sys-
tems are more susceptible to vibrational noise.
Radial and planar gradiometers: The relative merits of the two sensor types depends
on the detection situation. The radial gradiometers are favored for sources at all depths if
random noise is considered, while planar gradiometers are favored if correlated noise and
not very deep sources are considered. The advantage of planar gradiometers in the case of
correlated noise quickly decreases or is completely eliminated if the sources are extended.
Thus, if the sources are well localized, relatively superficial, and the background activity
of the brain is of no interest, planar gradiometers give better results, while radial gra-
diometers give better results in all other cases, i.e., if the background activity of the
brain is of interest, or if the sources are not well localized, or if the sources are deep.
Similar conclusions are reached if instead of the dipole reconstruction accuracy the rela-
tive merits of planar and radial gradiometers are evaluated on the basis of angular resolu-
tion in the sensor space of two nearby sources in the presence of noise.
In this section, MEG system considerations not directly concerned with the SQUID
magnetometer design will be discussed briefly: number of channels, head position loca-
tion, co-registration with other modalities (e.g., MRI), EEG electrodes, patient han-
dling, and stimulation equipment.
It was assumed in the previous discussions that there is a certain number of channels
distributed around the head, but no criteria for channel spacing have been given (except
in Section 6.1 where the sensor spacing limit was introduced for the sake of mathemati-
cal simplification of the dipole reconstruction expressions in the case of correlated
noise). The number of channels can be determined from knowledge of the head area cov-
ered by the sensors, Figure 7.1, and the sensor spacing. One possible way to determine
the sensor spacing is to require that it be compatible with the maximum signal spatial
frequency for which the signal power spectrum is above the system noise [118, 119]. A
Figure 7.1. Determination of the inter-channel spacing. (a, b) example of head coverage for CTF MEG
system, approximate surface area at the sensor level is 1160 cm2 ; (c) geometry for calculation of inter-
channel spacing; sensor is separated from the head by vacuum gap of the dewar.
67
simple analysis of this problem for a current dipole source leads to a rule of thumb that
the sensor spacing should be roughly equal to the source depth below the sensors when
the power S/N ratio is less than about 30.
If the mean channel separation is e, then the area per sensor is e2 for sensors arranged
in a square grid and (e2√3)/2 for a hexagonal grid. The source depth below the sensor
shell consists of the vacuum gap in the dewar and the actual depth below the surface of
the head, Figure 7.1.c. Using the above "rule of thumb," the channel spacing should be
e = gap + depth, and the required numbers of channels, N, are shown in Table 7.1.
Table 7.1. Number of channels, N, for different source depths. Surface area of the
sensor array at the sensor level is assumed to be 1,160 cm2 , vacuum gap = 1.5 cm.
depth (cm) e (cm) N (hexagonal lattice)
2 3.5 109
1.5 3 149
1 2.5 214
If, however, sources more complex than a single current dipole are used for estima-
tion of the highest spatial frequency, or if different criteria are used (e.g., resolution of
synthetic aperture magnetometry in Section 4.5), then the number of required channels
will be larger than indicated in Table 7.1.
A photograph of a 143-channel sensor layout with mean sensor spacing of about 3.2
cm is shown in Figure 7.2. The primary sensors are hardware 1st-order radial gradiome-
ters with 5-cm baseline, wound on the cylinders perpendicular to the structural shell.
Even though the subject's head is inserted in the dewar helmet, there is still some
freedom to move it, and accurate measurement of the head position relative to the sen-
sors is necessary. To accomplish this task, various 3D digitizing methods may be used
[120], or the MEG system itself can be utilized for the head position determination. An
example of such scheme is shown in Figure 7.3, where small coils are attached at spe-
cific anatomical locations on the head. The coils are activated by computer-controlled
signal generators, their magnetic field is detected by the reference system and converted
into positional information. Alternatively, the positional information also can be de-
rived from the sensor signals.
Figure 7.2. Photograph of 1st-order gradiometer sensors in the helmet area, 143 channels.
68
Figure 7.3. System for determination of head position used in CTF's MEG system. Small coils are placed at
the nasion and preauricular points. Coils are energized by a computer-controlled current source, and sig-
nals are detected by the reference system and converted into coil positions.
The measuring procedure itself is probably the most accurate step in the determina-
tion of the head position. The actual repeatability of the positioning measurement de-
termined using a static phantom was found to be about 0.3 ± 0.1 mm in unshielded en-
vironments, and the absolute accuracy (given by the system calibration) is better than 2
mm. Larger errors are caused by inaccurate placement of the marker coils and by head
motion during the measurement. It is estimated that the coils can be positioned on the
head surface with accuracy of no better than about 2 to 3 mm, and the head motion is
typically also of the order of 1 to 3 mm.
One goal of MEG is the localization of functional brain activity relative to anatomi-
cal structures obtained typically from MRI. Such localization requires the knowledge of
a transformation matrix (rotation, scaling, translation) between the two modalities.
There are different methods for determination of such transformations. The most com-
mon and simplest method uses a small number of anatomical marks that can be meas-
ured using both techniques (e.g., markers at the same locations as used for the MEG
head position determination). This method is not too accurate because the placement er-
rors of MEG and MRI markers are compounding. To reduce these problems, the partial-
or whole-head surface is digitized in the MEG system of coordinates by a device
mounted on the dewar, e.g., [120], or by magnetic sensors themselves, as in Figure 7.3
(using a method similar to that for measuring the EEG electrode positions - see below).
In the MRI system of coordinates, the surface of the head is reconstructed from the MRI
information. The transformation between the two systems is then determined by mini-
mizing the difference between the two surfaces [121-123].
When EEG is used simultaneously with MEG, care must be taken to select elec-
trodes with a low-level magnetic signature. The electrode positions can be determined ei-
ther by a 3D digitizer, or magnetically as in Figure 7.3. In this case, 3 marker coils are
placed at the anatomical landmarks, the head is withdrawn from the helmet such that its
surface can be comfortably reached by hand, and a 4th energized coil is used to touch the
electrodes and determine their positions in the MEG frame of reference. The currents in
the EEG electrode wires due to EEG potentials usually do not interfere with magnetic
measurements. Consider that the EEG electrode wires form a loop with radius r. The
69
Figure 7.4. Schematic diagram of various stimulation methods. (a, b) visual stimulators; (c) electrical
stimulator; (d) mechanical tactile stimulator; (e) switch for subject's response, e.g. voluntary motion; (f)
auditory stimulation.
magnetic field at the center of the loop is B = µ oI/(2r), where I is the current in the loop.
If it is required that B < 5 fT, and if r = 5 cm, then the current in the loop should satisfy
I < 0.4 nA. If the corresponding EEG voltage was assumed to be about 200 µV, then
the EEG amplifier impedance should be > 0.5 MΩ . The EEG amplifiers usually satisfy
this criterion, and the currents caused by the EEG signals are negligible from the MEG
point of view.
Before the MEG measurement, the subject must be properly prepared. All poten-
tially magnetic articles must be removed (correction glasses, various metal clips, zip-
pers, buttons, watches, etc.) and the patient should be checked for magnetic contamina-
tion (dental braces, contaminated hair, etc.). Such precautions may be relaxed if effective
noise cancellation is employed (e.g., synthetic aperture magnetometry, Section 4.5, al-
lows a successful MEG measurement even when subjects with dental braces are asked to
speak aloud). Where applicable, stimulus equipment is attached to the subject, and the
subject is measured either in a seated or a reclining position, and in some cases, the head
is fixed in the helmet by additional padding or inflatable bladders. All equipment and
components near the sensitive end of the MEG system must have a sufficiently low-
level magnetic signature in order to avoid magnetic noise and/or artifact generation.
MEG experiments are carried out with a wide range of stimulation equipment, Fig-
ure 7.4 (e.g., visual stimulation using projection screens and/or computer monitors,
electrical stimulation, tactile stimulation, auditory stimulation, various switches for de-
tection of subject's response, etc.). The wires of the electrical stimulators should be
tightly twisted when they are close to the helmet region. The computer monitors, if
used, must be tested for magnetic noise and synchronous artifacts - some monitors could
be as close as 1.5 m from the magnetic sensors if 3rd-order gradiometer noise cancella-
tion is used. All other magnetic components (speakers, switches, etc.) must be located
sufficiently far from the sensitive end of the MEG system.
70
The discussion of MEG systems will be concluded by a few examples of the end results
of the measurements. First two examples illustrate the use of MEG for pre-surgical
mapping. Pre-surgical mapping is used to determine important centers in the brain,
which in its diseased state may be distorted and may be difficult to recognize by a sur-
geon. The first example in Figure 8.1 is of Auditory Evoked Fields (AEF), where the
subject is stimulated by sound to either one or both ears, as in Figure 7.4.f. The time
traces of AEF signals were already shown in Figure 4.22.b, and the two dipoles overlaid
on an MRI image are shown for a healthy brain in Figure 8.1.a [124]. The dipoles are
located in the Sylvian fissure. The round dot indicates the estimated dipole position, and
the short line, its orientation. In Figure 8.1.b [16] the AEF in a patient with glioma is
shown. The closeness of the auditory cortex to the tumor suggested that only a limited
resection should be done in order to preserve a possible language area behind the primary
auditory cortex. One month after the surgery the verbal IQ returned to within a normal
range.
Figure 8.1. Auditory Evoked Fields. (a) normal male subject, estimated dipole positions for contra-lateral
ear stimuli; (b) patient with glioma, N100 dipole indicated primary auditory cortex adjacent to the tumor.
Adapted from references [16 and 124].
Figure 8.2. Median nerve stimulated Somatosensory Evoked Fields in the case of right frontal meningioma.
(a) pre-operative SEF indicates shifted right central sulcus due to tumor mass effect; (b) post-operative SEF
indicates normalized right central sulcus. Adapted from reference [124].
Figure 8.3. Temporal lobe epilepsy, complex partial seizures. (a) time traces of two of each left frontal,
right frontal and central MEG channels and F8 EEG channel. An interictal spike is shown by an arrow; (b)
traces for all channels, dots indicate time traces in (a) and arrow indicates the interictal spike. Adapted
from [81].
9. Conclusions
Figure 8.4. Differential current density map of a cortical response to language stimuli of a typical subject
overlaid on the MRI. Three orthogonal views for a single run in 15- to 30-Hz frequency band. Region of
focal suppression of source activity (white areas) appears in the left lateral frontal cortex. The peak inten-
sity of the suppression is 6.4 nA·m, and the area shown has been thresholded at half amplitude. Adapted
from [125].
shielding method used, residual noise always is present and must be eliminated by some
noise cancellation method (perhaps with the exception of the whole-body supercon-
ducting shield with very low residual magnetic fields, Section 4.2).
Synthetic noise cancellation assumes the existence of reference channels in addition
to the sensor channels. The reference channels are placed at some distance from the
measured tissue, and they mostly detect environmental noise. The references and sensors
are combined synthetically to eliminate noise from the sensing channels. The coeffi-
cients of this synthetic system can be determined to represent either higher-order gra-
diometers, or simply correspond to an adaptive process. There is a clear distinction be-
tween the two types of noise cancellation. Gradiometers of order n are designed to cancel
field and all gradients of orders up to n-1, and are sensitive to the gradient of order n (and
also affected by higher gradients). On the other hand, adaptive systems minimize the
outputs for all gradient orders. Adaptive systems of order n may not completely cancel
all gradients of orders less than n, but this incomplete cancellation of lower-order gradi-
ents is compensated by additional partial cancellation of gradiometers of order n and
higher. Adaptive systems perform better than gradiometer systems during the time pe-
riod for which the noise character remains constant. However, their performance quickly
deteriorates when the noise character changes. In practice, it was found that the time pe-
riod during which noise remains constant can be as short as several seconds.
Gradiometer systems are characterized by universal or "fixed" coefficients, i.e., the
coefficients are independent of the noise character and, therefore, of the time of the day,
system orientation or location. On the other hand, the adaptive coefficients are not uni-
versal and cannot be transported to different environments.
For all practical measurement situations, the output of both gradiometer and
adaptive systems is proportional to their dimension or baseline. In a simple 1st-order
system the baseline is the distance between the sensing coil and the reference system.
The sensor space filtering methods rely on the observation that different signal or
noise sources are represented by vectors pointing in different directions in a multidimen-
sional space, where the number of dimensions is equal to the number of channels. The
noise cancellation further assumes that the noise and signal vectors are not collinear and
that they can be separated. These methods can provide cancellation of far-field environ-
mental noise as well as noise from physically near sources (e.g., the activity in one part
73
of the brain can be separated from the activity in other parts of the brain, or heart sig-
nals can be separated from brain signals, etc.). Simultaneous use of the sensor space fil-
tering and synthetic noise cancellation methods can be beneficial; the synthetic noise
cancellation can be used to remove the noise and reduce the data dynamic range before
the sensor space filtering methods are applied.
Sensing coils, or flux transformers, are exposed to the measured fields and noise, and
their configuration determines the characteristics of the biomagnetometer system. Sev-
eral flux-transformer types were examined and compared using the equivalent current di-
pole reconstruction accuracy as a measure of the flux-transformer performance:
Magnetometers: Radial and vector magnetometers were compared. It was shown
that, generally, radial magnetometers are a better choice than vector magnetometers.
Radial magnetometers always give better results when random noise is considered. For
correlated noise, the radial and vector magnetometers are equivalent when the sources are
well within the sensor array, and the vector magnetometers are better when the sources
are close to or outside the sensor array edge. However, if the vector magnetometers were
used, their tangential components would introduce greater sensitivity to volume cur-
rents, and through it, to the details of the model.
Radial gradiometers (hardware or synthetic): The dependence of S/N ratio on the
baseline was examined. It was concluded that the best S/N ratio exists for a certain op-
timum baseline length, given by an interplay between the detected signal strength and
the sensitivity to environmental noise. The conditions for the optimum baseline were
calculated and investigated experimentally. At all experimentally investigated MEG
sites, the optimum baseline was found to be quite short (about 1.5 to 10 cm). In com-
parison with magnetometers, gradiometers with near optimum baseline are a better
choice.
Planar gradiometers: The relative merit of radial and planar gradiometers depends on
the type of measurement performed. If general brain activity is of interest, or if the
sources are not well localized (i.e., are extended) or the sources are deep, radial gradiome-
ters perform better. On the other hand, if general brain activity is of no interest, or if the
sources are known to be well localized, or are shallow, planar gradiometers perform bet-
ter.
In addition to the optimization of the sensing-coil configurations and the SQUID
electronics, biomagnetometers also contain other important elements, e.g., head posi-
tioning system, EEG electrodes, patient support system, miscellaneous stimulation
equipment, etc. For successful and low-noise operation of biomagnetometers, all pe-
ripheral equipment must be magnetically clean and must be harmoniously integrated
with the SQUID detectors.
Acknowledgment
I would like to express my gratitude to Drs. D. Cheyne and A.A. Fife for reviewing the
manuscript and for their constructive comments. I would like to thank my colleagues at
CTF Systems Inc., K. Betts, J. Box, M.B. Burbank, T. Cheung, D. Cheyne, A. A.
Fife, S. Govorkov, G. Haid, V. Haid, C. Hunter, P.R. Kubik, S. Lee, W. Ludwig, J.
McCubbin, J. McKay, D. McKenzie, D. Nonis, E. Reichl, S. Robinson, C. Schroyen,
P. Spear, B. Taylor, M. Tillotson and S. White for their enthusiastic support, help with
experiments, analysis and simulations; Dr. W. Ludwig for suggesting to investigate de-
74
tection of quadrupoles; Dr. H. Weinberg and Dr. D. Cheyne for guidance and help with
human experiments; Dr. H. Wilson, EDRD, Victoria, B.C. for help and discussion re-
garding conventional magnetometers; Dr. D. Cheyne of CTF and Dr. N. Nakasato of
Kohnan Hospital, Sendai, Japan for suggesting several figures in Section 2; and Dr.
K.C. Squires, Mr. G. Haid, and Dr. A. Ahonen for providing Figures 3.5 to 3.7.
References
[1]Waller, A.D. (1887) A demonstration on man of electromotive changes accompanying the heart's beat,
J. Physiol. 8, 229-234.
[2]Einthoven, W. (1903) Ein neues Galvanometer, Ann. Phys. 12 (4), 1059-1071; Einthoven, W. (1908)
Weiteres uber das Elektrokardiogramm, Pfluegers Arch. 122, 517-584.
[3]Baule, G.M. and McFee, R. (1963) Detection of the magnetic field of the heart, Am. Heart. J. 66, 95-96.
[4]Berger, H. (1929) Uber das Electrenkephalogramm des Menschen, Arch. Psychiat. Nervenkr. 7, 527-
570.
[5]Cohen, D. (1968) Magnetoencephalography: evidence of magnetic fields produced by alpha-rhythm
current, Science 161, 784-786.
[6]Josephson, B.D. (1962) Possible new effects in superconductive tunneling, Phys. Lett. 1, 251-253.
[7]Jaklevic, R., Lambe, R.C., Silver, A.H., and Mercereau, J.E. (1964) Quantum interference effects in
Josephson tunneling, Phys. Rev. Lett. 12, 159-160.
[8]Zimmerman, J.E. and Silver, A.H. (1964) Quantum effects in type II superconductors, Phys. Lett. 10,
47-48.
[9]Silver, A.H. and Zimmerman, J.E. (1967) Quantum states and transitions in weakly connected super-
conducting rings, Phys. Rev. 157, 317-341.
[10]Zimmerman, J.E., Thiene, P., and Harding, J.T. (1970) Design and operation of stable rf-biased super-
conducting point-contact quantum devices, and a note on the properties of perfectly clean metal con-
tacts, J. Appl. Phys. 41, 1572-1580.
[11]Mercereau, J.E. (1970) Superconducting magnetometers, Rev. Phys. Appl. 5, 13-20.
[12]Cohen, D., Edelsack, E.A. and Zimmerman, J.E. (1970) Magnetocardiograms taken inside a shielded
room with a superconducting point-contact magnetometer, Appl. Phys. Lett. 16, 278-280.
[13]Cohen, D. (1972) Magnetoencephalography: Detection of the brain's electrical activity with a super-
conducting magnetometer, Science 175, 664-666.
[14]Nakaya, Y. and Mori, H. (1992) Magnetocardiography, Clin. Phys. Physiol. Meas. 13, 191-229.
[15]Wiskwo, J.P.Jr. (1995) SQUID Magnetometers for Biomagnetism and Nondestructive Testing: Impor-
tant Questions and Initial Answers, IEEE Trans. Appl. Supercond. 5, 74-120.
[16]Nakasato, N., Seki, K., Kawamura, T., Ohtomo, S., Kanno, A., Fujita, S., Hatanaka, K., Fujiwara, S.,
Kayama, T., Takahashi, A., Jokura, H., Kumabe, T., Ikeda, H., Mizoi, K. and Yoshimoto, T. (1996)
Cortical Mapping Using an MRI-Linked Whole Head MEG System and Presurgical Decision Making,
Electroencephalogr. Clin. Neurophysiol. Suppl. 47, 333-341.
[17]Nakasato, N., Levesque, M.F., Barth, D.S., Baumgartner, C., Rogers, R.L. and Sutherling, W.W.
(1994) Comparisons of MEG, EEG, and ECoG source localization in neocortical partial epilepsy in
humans, Electroencephalogr. Clin. Neurophysiol. 171, 171-176.
[18]Vieth, J. B., Kober, H., Stippich, C., Kassubek, J.R., and Hopfengaertner, R. (1997) Time Course of
Abnormal MEG Activity Associated with Transient Ischemic Attack, in C. Aine et al (eds.),
Biomag96: Advances in Biomagnetism Research, Springer-Verlag, in press.
[19]Vieth, J., Kober, H., Grummich, P., Pongratz, H., Ulbricht, D., Brigel, C., and Daun, A. (1995) Slow
wave and beta wave activity associated with white matter structural brain lesions, localized by the di-
pole density plot, in C. Baumgartner et al (eds.), Biomagnetism: Fundamental Research and Clinical
Applications, Elsevier Science, IOS Press, pp.50-54.
[20]Lewine, J.D., Davis, L.E., Hart, B.L., Davis, J.T., Orrison, W.W.Jr. (1997) Neuromagnetic Evaluation
of Ischemic Lesions, in C. Aine et al (eds.), Biomag96: Advances in Biomagnetism Research,
Springer-Verlag, in press.
[21]Lewine, J.D., Orrison, W.W.Jr., Astur, R.S., Davis, L.E., Knight, J.E., Maclin, E.L., and Reeve, A.
(1995) Explorations of pathophysiological spontaneous activity by magnetic source imaging, in C.
Baumgartner et al (eds.), Biomagnetism: Fundamental Research and Clinical Applications, Elsevier
Science, IOS Press, pp.55-59.
[22]Cheyne, D., Vrba, J., Crisp, D., Betts, K., Burbank, M., Cheung, T., Fife, A.A., Haid, G., Kubik, P.,Lee,
S., McCubbin, J., McKay, J.,McKenzie, D., Spear, P., Taylor, B., Tillotson, M., Weinberg, H., Basar,
E., and Tsutada, T. (1992) Use of an unshielded, 64-channel whole-cortex MEG system in the study of
normal and pathological brain function. Proc. Satellite Symposium on Neuroscience and Technology,
Proc. IEEE-EMBS, Lyon, France.
75
[23]Vrba, J., Betts, K., Burbank, M.B., Cheung, T., Fife, A.A., Haid, G., Kubik, P.R., Lee, S., McCubbin, J.,
McKay, J., McKenzie, D., Spear, P., Taylor, B., Tillotson, M., Cheyne, D., and Weinberg, H. (1993)
Whole Cortex, 64 Channel SQUID Biomagnetic System, IEEE Trans. Appl. Supercond. 3, 1878-1882.
[24]Knuutila, K.E.T., Ahonen, A.I., Hamalainen, M.S., Kajola, M.J., Laine, P.P., Lounasmaa, O.V.,
Parkkonen, L.T., Simola, J.T.A., and Tesche, C.D. (1993) A 122-channel whole-cortex SQUID system
for measuring the brain's magnetic fields, IEEE Trans. Mag. 29, 3315-3320.
[25]Biomagnetic Technologies, Inc. 9727 Pacific Heights Blvd., San Diego, CA 92121-3719, USA.
[26]CTF Systems Inc., 15-1750 McLean Ave, Port Coquitlam, B.C., Canada, V3C 1M9.
[27]Neuromag Ltd., Elimaenkatu 22 A, P.O. Box 68, FIN-00511 Helsinki, Finland.
[28]Pflieger, M.E., Simpson, G.V., Ahlfors, S.P. and Ilmoniemi, R.J. (1997) Superadditive Information from
Simultaneous MEG/EEG data, in C. Aine et al (eds.), Biomag96: Advances in Biomagnetism Research,
Springer-Verlag, in press.
[29]Cohen, D. and Cuffin, B.N. (1987) A method of combining MEG and EEG to determine the sources,
Phys. Med. Biol. 32, 85-89.
[30]ter Brake, H.J. M, Flokstra, J., Jaszczuk, W., Stammis, R., van Ancum, G.K., and Martinez, A. (1991)
The UT 19-channel dcSQUID based neuromagnetometer, Clin. Phys. Physiol. Meas. 12B, 45-50.
[31]Becker, W., Dickmann, V., Jurgens, R. and Kornhuber, C. (1993) First experiences with a multichan-
nel software gradiometer recording normal and tangential components of MEG, Physiol. Meas. 14,
A45-A50.
[32]Dieckmann, V., Jurgens R., Becker, W., Elias, H., Ludwig, W. and Vodel,W. (1996) RF-SQUID to
DC-SQUID Upgrade of a 28-Channel Magnetoencephalography (MEG) System, Meas. Sci. Technol.
7, 844-852.
[33]Matlashov, A.N., Slobodchikov, V.Y., Bakharev, A.A., Zhuravlev, Yu., and Bondarenko, N. (1995)
Biomagnetic Multichannel system built with 19 cryogenic probes, in C. Baumgartner et al (eds.),
Biomagnetism: Fundamental Research and Clinical Applications , Elsevier Science, IOS Press, pp.493-
496.
[34]Drung, D., Absmann, Curio, G., Mackert, B.-M., Matthies, K.-P., Matz, H., Peters, M., Scheer, H.-J.
and Koch, H. (1995) The PTB 83-SQUID System for Biomagnetic Applications in a Clinic, IEEE
Trans. Appl. Supercond. 5, 2112-2117.
[35]Dossel, O., David, B., Fuchs, M. Kruger, J., Ludeke, K.-M., and Wischmann, H.-A. (1993) A 31-
channel SQUID system for biomagnetic imaging, Appl. Supercond. 3, 1813-1825.
[36]Yoshida, Y., Arakawa, A., Kondo, Y., Kajihara, S., Tomita, S., Tomita, T., Matsuda, N., and
Takanashi, Y. (1997) A 129 Channel Vector Neuromagnetic Imaging System, in C. Aine et al (eds.),
Biomag96: Advances in Biomagnetism Research, Springer-Verlag, in press.
[37]Ueda, M., Kandori, A., Ogata, H., Takada, Y., Komuro, T., Kazami, K., Ito, T. and Kado, H. (1995)
Development of a biomagnetic measurement system for brain research, IEEE Trans. Appl. Supercond.
5, 2465-2469.
[38]Sata, K., Fujimoto, S., Fukui, N., Haraguchi, E., Kido, T., Nishiguchi, N., and Kang, Y.M. (1997) De-
velopment of a 61-Channel MEG System Cooled by a GM/JT Cryocooler, in C. Aine et al (eds.),
Biomag96: Advances in Biomagnetism Research, Springer-Verlag, in press.
[39]Fujimoto, S., Sata, K., Fukui, N., Haraguchi, E., Kido, T., Nishiguchi, K., and Kang, Y.M. (1997) A 32-
channel MCG System Cooled by a GM/JT Cryocooler, in C. Aine et al (eds.), Biomag96: Advances in
Biomagnetism Research, Springer-Verlag, in press.
[40]Itozaki, H., Kugai, H., Nagaishi, T., Toyoda, H., Hirano, T., Haruta, Y., Tanaka, S., and Kado, H.
(1997) 32 Channel High Tc SQUID System for Biomagnetism, in C. Aine et al (eds.), Biomag96: Ad-
vances in Biomagnetism Research, Springer-Verlag, in press.
[41]Talairach, J. and Tournoux, P. (1988) Co-Planar Stereotactic Atlas of the Human Brain, G. Thieme
Medical Publishers, Stuttgart, New York. Or an electronic version CD-rom, digitized number of paper
brain atlases published by Thieme-Verlag: Talairach, J. and Tournoux, P. (1988) Co-Planar Stereotac-
tic Atlas of the Human Brain; Schaltenbrand, G., Hassler, R., and Wahren, W. (1977) Atlas of Stereo-
taxy of the Human Brain; selected patterns from Ono M., Kubik M.D., and Abernathy C.D. (1990)
Atlas of the Cerebral Sulci. CieMed, Center for Information-enhanced Medicine, jointly set up in 1994
by Johns Hopkins University and the Institute of Systems Science of the National University of Singa-
pore, Heng Mui Keng Terrace, Kent Ridge, 0511, Singapore 119597, Republic of Singapore,
http://ciemed.iss.nus.sg/research/brain/brain_cd.html.
[42]Hiriyannaiah, H.P. (1997) X-ray Computed Tomography for Medical Imaging, IEEE Signal Process-
ing 14, No.2, 42-59.
[43]Wright, G.A. (1997) Magnetic Resonance Imaging, IEEE Signal Processing 14, No.1, 56 - 66.
[44]Deans, S.R. and Roderick, S. (1983) The Radon Transform and some of its applications, Wiley, New
York.
[45]Smith, S.L. (1985) Magnetic Resonance Imaging, Anal. Chem. 57, A595-A607.
[46]Kumar, A., Welti, D., and Ernst, R.R. (1975), NMR Fourier Zeugmatography, J. Magn. Reson. 18, 69-
83.
[47]Lauterbur, P.G. (1973), Image Formation by Induced Local Interactions: Examples Employing Nu-
clear Magnetic Resonance, Nature 242, 190-191.
[48]Ollinger, J.M. and Fessler, J.A. (1997) Positron-Emission Tomography, IEEE Signal Processing 14,
No.1, 43-55.
76
[49]Gilardi, M.C., Rizzo, G., Lucignani, G., and Fazio, F. (1996) Integrating Competing Technologies with
MEG, in H. Weinstock (ed.), SQUID Sensors: Fundamentals, Fabrication and Applications, NATO
ASI Series E: Applied Sciences, Vol. 329, Kluwer Academic Publishers, Dordrecht, pp.491-516.
[50]Leichner, P.K., Morgan, H.T., Holdeman, K.P., Valentino, F., Lexa, R., Kelly, R.F., Hawkins, W.G.
and Dalrymple, G.V. (1995) SPECT imaging of Fluorine-18, J. Nucl. Med. 36, 1472-1475.
[51]Forzaneh, F., Riederer, S.J., and Pelc, N.J. (1990) Analysis of T2 Limitations and Off-Resonance Ef-
fects on Spatial Resolution and Artifacts in Echo-Planar Imaging, Magn. Reson. Med. 14, 123-139.
[52]Bruder, H., Fischer, H., Reinfelder, H.E., and Schmitt, F. (1992) Image Reconstruction for Echo Planar
Imaging with Nonequidistant k-Space Sampling, Magn. Reson. Med. 23, 311-323.
[53]Bandettini, P.A., Jesmanowicz, A., Wong, E.C., and Hyde, J.S. (1993) Processing Strategies for Time-
Course Data Sets in Functional MRI of the Human Brain, Magn. Reson. Med. 30, 161-173.
[54]George, J.S., Schmidt, D.M., Mosher, J.C., Aine, C.J., Ranken, D.M., Wood, C.C., Lewine, J.D.,
Sanders, J.A., and Belliveau, J.W. (1997) Dynamic Neuroimaging by MEG, Constrained by MRI and
fMRI, in C. Aine et al (eds.), Biomag96: Advances in Biomagnetism Research, Springer-Verlag, in
press.
[55]Ioannides, A.A., Taylor, J.G., and Muller-Gartner, H.W. (1997) MFT based investigation of MEG
congruence with fMRI, in C. Aine et al (eds.), Biomag96: Advances in Biomagnetism Research,
Springer-Verlag, in press.
[56]Partridge, L.D. and Partridge, L.D. (1993) The Nervous System, Its Function and Its Interaction with
the World, A Bradford Book, The MIT Press, Cambridge, Massachusetts.
[57]Wikswo, J.P.Jr. (1989) Biomagnetic Sources and their Models, in S.J. Williamson et al (eds.), Advances
in Biomagnetism, Plenum Press, New York and London, pp.1-18.
[58]Taccardi, B. (1982) Electrophysiology of excitable cells and tissues, with special consideration of the
heart muscle, in S.J. Williamson, G.L. Romani, L.Kaufman, and I.Modena (eds.), NATO ASI Biomag-
netism, An Interdisciplinary Approach , Series A: Life sciences, Vol.66, Plenum Press, New York and
London, pp.141-171.
[59]Carpenter M.B. (1985) Core Text of Neuroanatomy, Williams & Wilkins, Baltimore, London, Sydney.
[60]Okada, Y.C., Papuashvili, N., and Xu, C. (1997) Maximum current dipole moment density as an im-
portant physiological constraint in MEG inverse solutions, in C. Aine et al (eds.), Biomag96: Advances
in Biomagnetism Research, Springer-Verlag, in press.
[61]Nakasato, N., Fujita, S., Seki, K., Kawamura, T., Matani, A., Tamura, I., Fujiwara, S., and Yoshimoto,
T. (1995) Functional localization of bilateral auditory cortices using an MRI-linked whole head mag-
netoencephalography (MEG) system, Electroencephalogr. Clin. Neurophysiol. 94, 183-190.
[62]Grynszpan F. and Geselowitz, D.B. (1973) Model studies of the magnetocardiogram, Biophysical
Journal 13, 911-925.
[63]Sarvas, J. (1987) Basic Mathematical and Electromagnetic Concepts of the Biomagnetic Inverse
Problem, Phys. Med. Biol. 32, 11-22.
[64]Williamson, S.J. and Kaufman, L. (1981) Biomagnetism, J. Mag. Mag. Materials 22, 129-201.
[65]Williamson, S.J. and Kaufman, L. (1981) Biomagnetic Fields of the Cerebral Cortex, in S.N. Erne et al
(eds.), Biomagnetism, Walter de Gruyter, Berlin and New York, pp.353-402.
[66]Cheyne, D., Roberts, L.E., Gaetz, W., Bosnyak, D., Weinberg, H., Johnson, B., Nahmias, C., and
Deecke, L. (1997) EEG and MEG Source Analysis of Somatosensory Evoked Responses to Mechani-
cal Stimulation, in C. Aine et al (eds.), Biomag96: Advances in Biomagnetism Research, Springer-
Verlag, in press.
[67]Clarke, J. (1993) SQUIDs: theory and practice, in H. Weinstock and R.W. Ralston (eds.), The New
Superconducting Electronics, Kluwer Academic Publishers, Dordrecht, pp.123-180.
[68]Clarke, J. (1996) SQUID Fundamentals, in H. Weinstock (ed.), SQUID Sensors: Fundamentals,
Fabrication and Applications, NATO ASI Series E: Applied Sciences, Vol. 329, Kluwer Academic
Publishers, Dordrecht, pp.1-62.
[69]Braginski, A.I. (1996) Fabrication of high-temperature SQUID magnetometers, in H. Weinstock (ed.),
SQUID Sensors: Fundamentals, Fabrication and Applications, NATO ASI Series E: Applied Sciences,
Vol. 329, Kluwer Academic Publishers, Dordrecht, pp.235-288.
[70]Vrba, J. (1996) SQUID gradiometers in real environments, in H. Weinstock (ed.), SQUID Sensors:
Fundamentals, Fabrication and Applications, NATO ASI Series E: Applied Sciences, Vol. 329, Kluwer
Academic Publishers, Dordrecht, pp.117-178.
[71]McKay, J., Vrba, J., Betts, K., Burbank, M.B., Lee, S., Mori, K., Nonis, D., Spear, P., and Uriel, Y.
(1993) Implementation of a Multi-Channel Biomagnetic Measurement System Using DSP Technology,
Proceedings of 1993 Canadian Conference on Electrical and Computer Engineering, vol. II, pp.1090-
1093.
[72]Ripka, P. (1992) Review of fluxgate sensors, Sensors and Actuators, A33, 129-141.
[73]Nielsen, O.V., Petersen, J.R., Primdahl, F., Brauer, P.,Hernandot, B., Fernandez, A., Merayo, J.M.G.,
and Ripka, P. (1995) Development, construction and analysis of the "Oersted" fluxgate magnetometer,
Meas. Sci. Technol. 6, 1099-1115.
[74]Le Garrec, A., Leger, J.M., Pureur, D., Douay, M., Bernage, P., Niay, P., and Delevaque, E. (1994) A
tunable fiber laser for application to helium optically pumped magnetometer, Journal de Physique IV,
4, C4:695-698.
77
[75]Carelli, P., Modena, I., and Romani, G.L. (1982) Detection coils, in S.J. Williamson, G.L. Romani,
L.Kaufman, and I.Modena (eds.), NATO ASI Biomagnetism, An Interdisciplinary Approach, Series A:
Life sciences, Vol.66, Plenum Press, New York and London, pp.85-95.
[76]Katila, T. (1989) Principles and applications of SQUID sensors, in S.J. Williamson et al (eds.),
Advances in Biomagnetism, Plenum Press, New York and London, pp.19-32.
[77]Katila, T. (1981) Instrumentation for biomedical applications," in Biomagnetism, Proceedings of the
Third International Workshop on Biomagnetism, in S.N. Erne et al (eds.), Biomagnetism, Walter de
Gruyter, Berlin and New York, pp.3-31.
[78]Kelha, V.O. (1981) Construction and performance of the Otaniemi magnetically shielded room, in S.N.
Erne et al (eds.), Biomagnetism, Walter de Gruyter, Berlin and New York, pp.33-50.
[79]Sullivan, G.W. and Flynn, E.R. (1987) Performance of the Los Alamos shielded room, in K. Atsumi et
al (eds.), Biomagnetism’87, Tokyo Denki University Press, Tokyo, pp.486-489.
[80]Vrba, J., Betts, K., Burbank, M., Cheung, T., Cheyne, D., Fife, A.A., Haid, G., Kubik, P.R., Lee, S.,
McCubbin, J., McKay, J., McKenzie, D., Mori, K., Spear, P., Taylor, B., Tillotson, M., and Xu, G.
(1995) Whole cortex 64 channel system for shielded and unshielded environments, in C. Baumgartner
et al (eds.), Biomagnetism: Fundamental Research and Clinical Applications, Elsevier Science, IOS
Press, pp.521-525.
[81]Vrba, J., Angus, V., Betts, K., Burbank, M.B., Cheung, T., Fife, A.A., Haid, G., Kubik, P.R., Lee, S.,
Ludwig, W., McCubbin, J., McKay, J., McKenzie, D., Robinson, S.E., Smith, M., Spear, P., Taylor, B.,
Tillotson, M., Cheyne, D., and Weinberg, H. (1997) 143 channel whole-cortex MEG system, in C.
Aine et al (eds.), Biomag96: Advances in Biomagnetism Research, Springer-Verlag, in press.
[82]Knuutila, J. and Hamalainen, M. (1987) Characterisation of brain noise using a high sensitivity 7-chan-
nel magnetometer, in K. Atsumi et al (eds.), Biomagnetism’87, Tokyo Denki University Press, Tokyo,
pp.186-189.
[83]Huotilainen, M., Ilmoniemi, R.J., Tiitinen, H. Lavikainen, J., Alho, K., Kajola, M., and Naatanen, R.
(1995) The projection method in removing eye-blink artefacts from multichannel MEG measurements,
in C. Baumgartner et al (eds.), Biomagnetism: Fundamental Research and Clinical Applications,
Elsevier Science, IOS Press, pp.363-367.
[84]Zimmerman, J. E. (1977) SQUID instruments and shielding for low level magnetic measurements, J.
Appl. Phys. 48, 702-710.
[85]Stroink, G., Blackford, B., Brown, B. and Horacek, M. (1981) Aluminum Shielded Room for Biomag-
netic Measurements. Rev. Sci. Instrum. 52(3), 463-468.
[86]Vacuumschmelze GmbH, Hanau, Germany; Shielded Room model AK-3.
[87]Amuneal Manufacturing Corp., 4737 Darrah Street, Philadelphia, PA 19124, USA.
[88]Tokin Corporation, 6-7-1 Koriyama Tihakuku, Sendai-City, Miyagi-pref, 982, Japan.
[89]Sullivan, G. W. and Flynn, E. R. (1987) Performance of the Los Alamos Shielded Room, in K. Atsumi
et al (eds.), Biomagnetism’87, Tokyo Denki University Press, Tokyo, pp.486-489.
[90]Erne, S. N., Hahlbohm, H.-D., Scheer, H. and Trontelj, Z. (1981) The Berlin Magnetically Shielded
Room (BMSR) Section B - Performances, in S.N. Erne et al (eds.), Biomagnetism, Walter de Gruyter,
Berlin and New York, pp.79-87.
[91]Matsuba, H., Shintomi, K., Yahara, A., Imai, K., Irisawa, D., and Matsuda, M. (1997) High Tc super-
conducting shielded biomagnetometer system, in C. Aine et al (eds.), Biomag96: Advances in Biomag-
netism Research, Springer-Verlag, in press.
[92]Matsumoto, K., Yamagishi, Y., Wakusawa, A., Noda, T., Fujioka, K., and Kuraoka, Y. (1992) SQUID
based active shield for biomagnetic measurement, in M. Hoke et al (eds.), Biomagnetism: Clinical
Aspects, Elsevier Science Publishers B.V., Excerpta Medica, Amsterdam, London, NewYork, Tokyo,
pp.857-861.
[93]Malmivuo, J., Lekkala, J., Kontro, P., Suomaa, I., and Vihinen, H. (1987) Improvement of the proper-
ties of an eddy current magnetic shield with active compensation, J. Phys. E: Sci. Instrum. 20, 151-
164.
[94]Kelha, V.O., Pukki, J.M., Peltonen, R.S., Penttinen, V.J., Ilmoniemi, R.J., and Heino, J.J. (1982) Design,
construction, and performance of a large volume magnetic shield, IEEE Trans. Magn. 18, 260-270.
[95]ter Brake, H.J.M., Wieringa, H.J., and Rogalla, H. (1991) Improvement of the performance of a µ-
metal magnetically shielded room by means of active compensation, Meas. Sci. Techn. 2, 596-601.
[96]ter Brake, H.J.M., Huonker, R., and Rogalla, H. (1993) New results in active noise compensation for
magnetically shielded rooms, Meas. Sci. Techn. 4, 1370-1375.
[97]Vrba, J. and McCubbin, J. (1983) First-gradient balancing of higher-order gradiometers, Il Nuovo
Cimento 2D, 142-152.
[98] Vrba, J., Fife, A.A., and Burbank, M.B. (1982) Spatial discrimination in SQUID gradiometers and
third-order gradiometer performance, Can. J. Phys. 60, 1060-1073.
[99]Vrba, J., McCubbin, J., Lee, S., Fife, A.A., and Burbank, M.B. (1989) Noise cancellation in biomag-
netometers, in S.J. Williamson et al (eds.), Advances in Biomagnetism, Plenum Press, New York and
London, pp.733-736.
[100]Vrba, J., Haid, G., Lee, S., Taylor, B., Fife, A.A., Kubik, P., McCubbin, J., and Burbank, M.B. (1991)
Biomagnetometers for unshielded and well shielded environments, Clin. Phys. Physiol. Meas. 12 B,
81-86.
78
[101]Bruno, A. C., Dolce, C. S., Soares, S. D. and Ribeiro, P. C. (1989) Spatial Fourier Technique for Cali-
brating Gradiometers, in S.J. Williamson et al (eds.), Advances in Biomagnetism, Plenum Press, New
York and London, pp.709-712.
[102]Drung, D. (1991) Performance of an Electronic Gradiometer in Noisy Environments, in H. Koch and
H. Lubbig (eds.), SQUID’91, Superconducting Devices and their Applications, Berlin, June 18-21,
1991, Springer Proceedings in Physics, 64, pp.542-546.
[103]Bendat, J. S. and Piersol, A. G. (1986) Random Data, John Willey & Sons, New York, Chichester,
Brisbane, Toronto, Singapore.
[104]Uusitalo, M.A. and Ilmoniemi, R.J. (1997) Signal-Space Projection method for separating MEG or
EEG into components, Med. & Biol. Eng. & Comp. 35, 135-140.
[105]Huotilainen, M., Ilmoniemi, R.J., Tiitinen, H., Lavikainen, J., Alho, K., Kajola, M., Simola, J., and
Naatanen, R. (1995) The projection method in removing eye-blink artefacts from multichannel MEG
measurements, in C. Baumgartner et al (eds.), Biomagnetism: Fundamental Research and Clinical
Applications, Elsevier Science, IOS Press, pp.363-367.
[106]Parker, L.R. (1994) Geophysical inverse theory, Princeton University Press, Princeton, NJ.
[107]Robinson, S.E. and Rose, D.F. (1992) Current source image estimation by spatially filtered MEG, in
M. Hoke et al (eds.), Biomagnetism: Clinical Aspects, Elsevier Science Publishers B.V., Excerpta
Medica, Amsterdam, London, NewYork, Tokyo, pp.761-765.
[108]Robinson, S.E. (1989) Theory and properties of lead field synthesis analysis, in S.J. Williamson et al
(eds.), Advances in Biomagnetism, Plenum Press, New York and London, pp.599-602.
[109]Vrba, J. and McKay, J. (1997) Character and acquisition of multichannel biomagnetic data, abstract
submitted to ISEC'97, June 25-28, Berlin, Germany.
[110]Vrba, J., Fife, A.A., Burbank, M.B. (1981) Digital SQUID electronics in geophysical applications, in
H. Weinstock and W.C. Overton, (eds.), SQUID Applications to Geophysics, Tulsa, Oklahoma: The
Society of Exploration Geophysicists, pp.31-34.
[111]Lutkenhoner, B. (1996) Current dipole localization with an ideal magnetometer system, IEEE Trans.
Biomed. Eng. 43, 1049-1061.
[112]Matsuba, H., Vrba, J., and Cheung, T. (1997) Current Dipole Localization Error as a Function of
System Noise and the Number of Sensors, in C. Aine et al (eds.), Biomag96: Advances in Biomag-
netism Research, Springer-Verlag, in press.
[113]Vrba, J., Cheung, T., and Robinson, S. (1996) Simulation of dipole reconstruction accuracy, CI-683-
1096, unpublished CTF internal report.
[114]Vrba, J. (1997) Baseline optimization for noise cancellation systems. Submitted to IEEE-EMBS,
Chicago, Oct 30 to Nov 2.
[115]Vrba, J., Taylor, B., Cheung, T., Fife, A.A., Haid, G., Kubik, P.R., Lee, S., McCubbin, J., and
Burbank, M. (1995) Noise cancellation by a whole-cortex SQUID MEG system, IEEE Trans. Appl.
Supercond. 5, 2118-2123.
[116]ter Brake, H.J.M., Sensing-coil optimization based on signal-to-noise ratio, private communication.
The paper was presented at the New York Biomagnetism Conference 1989, but was not published.
[117]Kemppainen, P.K. and Ilmoniemi, R.J. (1989) Channel capacity of multichannel magnetometers, in
S.J. Williamson et al (eds.), Advances in Biomagnetism, Plenum Press, New York and London,
pp.635-638.
[118]Kuc, R. (1996) Magnetometer spacing criterion for biomagnetic source current imaging, IEEE Trans.
Biomed. Eng. 43, 1125-1127.
[119]Ahonen, A.I., Hamalainen, M.S., Ilmoniemi, R.J., Kajola, M.J., Knuutila, J.E.T., Simola, J.T., and
Vilkman, V.A. (1993) Sampling theory for neuromagnetic detector arrays, IEEE Trans. Biomed. Eng.
40, 859-869.
[120]Polhemus Inc., 1 Hercules drive, PO Box 560, Colchester, VT, 05446, USA.
[121]Bamidis, P.D. and Ioannides, A.A. (1997) Combination of point and surface matching techniques for
accurate registration of MEG and MRI, in Aine, C., Okada, Y., Stroink, G., Swithenby, S., and Wood,
C. (eds.), Biomag96: Advances in Biomagnetism Research, Springer-Verlag, in press.
[122]Abraham-Fuchs, K., Lindner, L., Wegener, P., Nestel, F., and Schneider, S. (1991) Fusion of biomag-
netism with MRI or CT images by contour-fitting, Biomed. Eng. 36 (Suppl), 88-89.
[123]Kober, H., Grummich, P., and Vieth, J. (1995) Fit of the digitized head surface with the surface re-
constructed from MRI-tomography, in C. Baumgartner et al (eds.), Biomagnetism: Fundamental
Research and Clinical Applications, Elsevier Science, IOS Press, pp.309-312.
[124]Nakasato, N., Fujita, S., Matani, A., Tamura, I., Fujiwara, S., and Yoshimoto, T. (1995) Clinical appli-
cation of the whole head MEG: auditory evoked response in patients with intracranial structural le-
sions, in C. Baumgartner et al (eds.), Biomagnetism: Fundamental Research and Clinical
Applications, Elsevier Science, IOS Press, pp.186-190.
[125]Robinson, S.E., Weinberg, H., Cheyne, D., Vrba, J., and Jantzen, K.J. (1997) Functional imaging of
cerebellar activity during a simple differential reading task by whole-head
magentoencephalography, 3rd International Conference on Functional Mapping of the Human Brain,
Copenhagen, May 19-23, 1997.
[126]Rabiner, L.R. and Gold, B. (1975) Theory and application of digital signal processing, Prentice-Hall,
Inc., Englewood Cliffs, New Jersey.
79