Physics in Medicine & Biology

You may also like

- Photoactive layer formation in the dark for
Excitation spectroscopy in multispectral optical high performance of air-processable
organic photovoltaics
fluorescence tomography: methodology, feasibility Akihiro Maeda, Ruiyuan Liu, Kilho Yu et al.

- An improved micro-thermo-mechanics
and computer simulation studies model for shape memory alloys: analysis
and applications
Tareq Manzoor, Muhammad Zafar,
To cite this article: Abhijit J Chaudhari et al 2009 Phys. Med. Biol. 54 4687 Sanaullah Manzoor et al.

- Tomographic reconstruction of the

runaway distribution function in TCV using
multispectral synchrotron images
T.A. Wijkamp, A. Perek, J. Decker et al.
View the article online for updates and enhancements.

This content was downloaded from IP address 195.43.22.135 on 04/08/2022 at 15:59

IOP PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 54 (2009) 4687–4704 doi:10.1088/0031-9155/54/15/004

Excitation spectroscopy in multispectral optical

fluorescence tomography: methodology, feasibility
and computer simulation studies
Abhijit J Chaudhari1,2, Sangtae Ahn3, Richard Levenson4,
Ramsey D Badawi2, Simon R Cherry1 and Richard M Leahy3
1 Department of Biomedical Engineering, University of California-Davis, Davis, CA 95616, USA
2 Department of Radiology, UC Davis Medical Center, Sacramento, CA 95817, USA
3 Signal and Image Processing Institute, University of Southern California, Los Angeles,

CA 90089, USA
4 Cambridge Research Instruments, Inc., Woburn, MA 01801, USA

E-mail: [email protected] and [email protected]

Received 12 February 2009, in final form 11 June 2009

Published 10 July 2009
Online at stacks.iop.org/PMB/54/4687

Abstract
Molecular probes used for in vivo optical fluorescence tomography (OFT)
studies in small animals are typically chosen such that their emission spectra lie
in the 680–850 nm wavelength range. This is because tissue attenuation in this
spectral band is relatively low, allowing optical photons even from deep sites
in tissue to reach the animal surface and consequently be detected by a CCD
camera. The wavelength dependence of tissue optical properties within the
680–850 nm band can be exploited for emitted light by measuring fluorescent
data via multispectral approaches and incorporating the spectral dependence of
these optical properties into the OFT inverse problem—that of reconstructing
underlying 3D fluorescent probe distributions from optical data collected on
the animal surface. However, in the aforementioned spectral band, due to
only small variations in the tissue optical properties, multispectral emission
data, though superior for image reconstruction compared to achromatic data,
tend to be somewhat redundant. A different spectral approach for OFT is
to capitalize on the larger variations in the optical properties of tissue for
excitation photons than for the emission photons by using excitation at multiple
wavelengths as a means of decoding source depth in tissue. The full potential
of spectral approaches in OFT can be realized by a synergistic combination
of these two approaches, that is, exciting the underlying fluorescent probe at
multiple wavelengths and measuring emission data multispectrally. In this
paper, we describe a method that incorporates both excitation and emission
spectral information into the OFT inverse problem. We describe a linear
algebraic formulation of the multiple wavelength illumination-multispectral
detection forward model for OFT and compare it to models that use only
excitation at multiple wavelengths or those that use only multispectral detection
0031-9155/09/154687+18$30.00 © 2009 Institute of Physics and Engineering in Medicine Printed in the UK 4687
4688 A J Chaudhari et al

techniques. This study is carried out in a realistic inhomogeneous mouse

atlas using singular value decomposition and analysis of reconstructed spatial
resolution versus noise. For simplicity, quantitative results have been shown
for one representative fluorescent probe (Alexa 700R
) and effects due to tissue
autofluorescence have not been taken into account. We also demonstrate the
performance of our method for 3D reconstruction of tumors in a simulated
mouse model of metastatic human hepatocellular carcinoma.
(Some figures in this article are in colour only in the electronic version)

1. Introduction

For optical fluorescence tomography (OFT) in small animals, a large number of near-
infrared (NIR) photons emitted by fluorescent sources even at deep sites in tissue reach
the animal surface due to short optical path lengths and low tissue attenuation (Chance 1991,
Weissleder et al 1999, Ntziachristos and Weissleder 2001). Consequently, 3D tomographic
reconstruction of fluorescent source distribution using non-ionizing NIR radiation becomes
possible (Weissleder and Ntziachristos 2003, Cherry 2004, Ntziachristos et al 2005). A large
number of novel fluorescent dyes and proteins probing cellular and sub-cellular processes
have become available (Neefjes and Dantuma 2004, Giepmans et al 2006, Adams et al 2007).
These advancements coupled with the development of 3D tomographic small animal imaging
systems by academicians (Ntziachristos and Weissleder 2002, Patwardhan et al 2005, Joshi
et al 2006a, Hassan et al 2007, Kumar et al 2008) and by industries, e.g. IVIS 3D R
(Caliper
Life Sciences, Hopkinton, MA), eXplore Optix MX (Advanced Research Technologies-GE
Healthcare), FMTTM 2500 (VisEn Medical, Inc., Bedford, MA), etc, have facilitated both
applications of OFT for early detection and rapid screening in mouse models of disease
as well as for detailed investigation of molecular function (Massoud and Gambhir 2003,
Ntziachristos 2006, Weber et al 2008, Luker and Luker 2008).
In this paper, we focus on continuous-wave (CW) domain illumination and detection
techniques for OFT (Ntziachristos et al 2005) with non-contact tomographic measurements
(Schulz et al 2004, Joshi et al 2006b). Light propagation for OFT in the CW domain can be
modeled by three steps:
(A) excitation photon transport from the external illumination source to the animal surface
and then to the site of the underlying fluorescent probe,
(B) propagation of emitted photons from the fluorescent source to the animal surface,
(C) mapping of the photon fluence from the animal surface to the CCD camera.
The coupled radiative mapping in steps (A) and (B), whose mathematical description
in biological tissue can be given by the radiative transport equation (RTE), constitutes
the OFT ‘forward’ model (Arridge et al 1993, Klose and Hielscher 2003). OFT studies
based on the solution of the RTE have been carried out (Klose et al 2005, Joshi et al
2008). For practical biomedical imaging applications, however, the RTE-based solution is
computationally expensive (Ntziachristos 2006). The steady-state diffusion equation (DE)
with appropriate boundary conditions may be used as a practical approximation to the
RTE (Patterson et al 1992, Arridge and Hebden 1997, Paithankar et al 1997). However,
concerns regarding the applicability of this approximation to small animal volumes where
source–detector separation distances are small and void, and non-diffusive regions are
Excitation spectroscopy in multispectral fluorescence tomography 4689

present, exist (Hielscher 2005). Ren et al (2007) compared RTE-based and diffusion-
approximation-based optical tomographic reconstructions in small domains. They found that
RTE-based solutions were more quantitatively accurate, while the diffusion-approximation-
based solutions were more practical computationally. Combined radiosity-diffusion models
have also been proposed in the literature (Dehghani et al 2000). The issue of how well the
DE approximates the RTE is not considered in our paper. Owing to the large number of
image reconstructions carried out in our paper, we have used the diffusion approximation for
generating the forward models and, thus, our results are relevant when this approximation is
sufficiently accurate.
The surface geometry and internal anatomy of the animal need to be estimated before
the OFT forward model can be accurately solved. Direct methods for this estimation involve
scanning the animal using an anatomical imaging modality, e.g. computed tomography (CT) or
magnetic resonance imaging (MRI), before or after the optical scan without moving the animal
(Ntziachristos et al 2002, Chaudhari et al 2005, Lv et al 2006, Guven et al 2007). Alternatively,
the animal’s surface topography can be estimated by all-optical methods, for example by using
photogrammetry systems (Ripoll et al 2003), by utilizing systems that project structured light
on the animal surface (Rice et al 2006), by shadowgrammetric techniques (Meyer et al 2007)
or via 3D volume carving (Lasser et al 2008). Once the surface topography is known, the
internal organ boundaries may be estimated using a deformable mouse atlas (Wang et al 2006,
Chaudhari et al 2007). Alternatively, data normalization techniques may be used to reduce the
impact of tissue heterogeneity (Soubret et al 2005, Axelsson et al 2007). Step (C) is typically
modeled using the estimated surface geometry with free-space optics formulae and assuming
that the mouse surface is Lambertian (Ripoll and Ntziachristos 2004).
In the NIR region, low attenuation of emitted photons is a consequence of a low optical
absorption coefficient μa . The reduced scattering coefficient μs is high compared to μa in this
region, but is lower than its own value for blue light. Since both μa and μs are functions of
wavelength λ (Cheong et al 1990), for step (B), tissue acts like a spectral filter for emitted
photons causing the wavelength spectra measured at the tissue surface for a given optical
source to vary depending on its depth in tissue (Swartling et al 2005 Zacharakis et al 2005
Zavattini et al 2006). This property has been exploited in bioluminescence tomography (BLT)
by the use of multispectral or hyperspectral data measurement techniques that lead to better
conditioning of the forward models and yield reasonably accurate 3D source reconstruction
results for in vivo studies compared to those generated using achromatic or monochromatic
data sets (Chaudhari et al 2005, Cong and Wang 2006, Dehghani et al 2006, Kuo et al 2007).
The imaging window used in these BLT studies was restricted to 550–660 nm corresponding
to the emission band for Firefly Luciferase (FFL). These studies concluded that variations in
tissue μa and μs for individual wavelength bins within the emission band of interest provided a
fair amount of independent information that helped in reducing the ill-posedness of the inverse
problem. For OFT, fluorescent probes are chosen to have their emission spectrum in the 680–
850 nm band where the tissue attenuation is a lot lower than in the 550–660 nm range (Pinaud
et al 2006). Thus, more fluorescent light photons will reach the CCD camera compared to
BLT for sources with equivalent strengths (Ntziachristos et al 2005). While multispectral
fluorescent emission data were also found to be helpful in reducing the ill-conditioning of
the OFT inverse problem (Zacharakis et al 2005, Swartling et al 2005, Zavattini et al 2006,
Chaudhari 2006), the full potential of spectral approaches has not yet been realized. This
is because in the 680–850 nm wavelength band, the variation in μa and μs with wavelength
is small (Weissleder and Ntziachristos 2003, Bargo et al 2005). As a consequence, spectral
emission data acquired in this band tend to be somewhat redundant.
4690 A J Chaudhari et al

Table 1. Forward models used for the singular value decomposition study, spatial resolution
versus variance analyses and imaging comparisons. These models were chosen to allow for
detailed investigation of both excitation and emission domains.

Model name Abbreviation

Single wavelength illumination-achromatic detection SWI-AD
Single wavelength illumination-multispectral detection SWI-MD
Multiple wavelength illumination-achromatic detection MWI-AD
Multiple wavelength illumination-multispectral detection MWI-MD

Step (A) for OFT offers a unique degree of freedom unavailable in BLT. The external
illumination source may be used to elicit different mappings from the fluorescent source to the
animal surface and can provide a way of reducing the ill-posedness of the OFT inverse problem
compared to that for BLT (Chang et al 1997, Ntziachristos et al 2005). Proposed approaches
have investigated parameters based on the nature of the illumination source (e.g. point-like or
distributed) and the spatial distribution of the incident illumination light in the imaging field
of view for improving the conditioning of the OFT forward problem (Ntziachristos et al 2004,
Godavarty et al 2004, Pogue et al 2004, Roy et al 2006, Joshi et al 2006b).
Step (A) is also a function of wavelength. Optical property variability for tissue is higher
for excitation light compared to emission light (Cheong et al 1990). This is because the
excitation spectrum is blue-shifted compared to the emission spectrum. Thus, illumination
at different wavelengths within the excitation band of a fluorophore should correspond to
different light propagation functions in tissue at those wavelengths. This multiple wavelength
approach for illumination is complementary to multispectral detection. Therefore, the ill-
posedness of the OFT inverse problem may potentially be reduced if spectral information
from both the excitation and emission domains is combined.
In this paper, we propose a method for OFT that incorporates both excitation and emission
spectral information about the underlying fluorescent probe into the tomographic inverse
problem. This paper is organized as follows. We first present a linear algebraic formulation
of forward models that allows taking into account excitation at multiple wavelengths and
multispectral detection of emission data. Benefits of utilizing both excitation and emission
domains for OFT are presented in a singular value decomposition (SVD) study in the 3D
Digimouse atlas (Dogdas et al 2007), where the conditioning of the forward models listed
in table 1 is compared. A regularized pseudo-inverse-based estimator is used for image
reconstruction. A comparative study of the resulting spatial resolution and variance for this
estimator for the different forward models is also presented. We also show results from a
performance evaluation study that demonstrates the application of the proposed method to the
reconstruction of primary and secondary tumors in a computer-simulated orthotopic mouse
model of metastatic human hepatocellular carcinoma (HCC) (Sun et al 1996).

2. Methods

2.1. Forward models for OFT

For OFT, the forward problem is to predict the photon fluence measured at the CCD camera
plane due to an underlying fluorescent probe distribution inside the 3D animal volume. This
model should also account for the excitation of the fluorophore by an external illumination
source. Steps (A) and (B) described in section 1 can be modeled using the following coupled
Excitation spectroscopy in multispectral fluorescence tomography 4691

equations based on the steady-state DE with a Robin-type boundary condition (Arridge et al

1993, Patterson and Pogue 1994, Reynolds et al 1997, Milstein et al 2003):

g(r, λi ) = Di p(r, λi ), (1)

q(r, λj ) = Dj h(r, λj ), (2)

where D = − { · κ(r, λ)  −μa (r, λ)c} represents the diffusion operator as a function of
wavelength λ and of 3D position r, and Di and Dj represent the diffusion operators at the
excitation wavelength λi and emission wavelength λj respectively. The term g denotes the
photon flux on the animal surface due to the external illumination, p and q denote the incident
and emitted flux for a fluorophore in the interior of the animal volume and h denotes the photon
flux inside the volume as well as on the surface. The emitted photon flux q = ηp, where η
represents the fluorescence yield. In the equation for D, the diffusion coefficient κ(r) is given
by the equation
c
κ(r) = , (3)
3[μa (r, λ) + μ s (r, λ)]
where c represents the velocity of light. Numerically, (1) and (2) can be solved by using
either a discretized representation of the animal volume and a finite-element-based diffusion
equation solver (Arridge et al 1993, Paulsen and Jiang 1995, Joshi et al 2004, Cong and
Wang 2005) or analytic approximations of the diffusion equation with appropriate boundary
conditions (Haskell et al 1994, Rice et al 2001, Soubret et al 2005, Axelsson et al 2007).
Since the focus of this paper is spectral domain analysis, it is not our goal to evaluate the
impact of internal anatomical variability on reconstructed images. Thus, we demonstrate our
method using the Digimouse atlas, where internal anatomy is known and optical properties
can be assigned to the organs based on published data (Alexandrakis et al 2005). We solve
both (1) and (2) using a tetrahedral representation of the animal volume and the finite element
method (FEM) as outlined in Arridge (1999).
Let n represent the total number of possible source locations inside the animal volume
and m denote the total number of nodes on the animal surface. Two sets of measurements
are carried out. First, a subset of the m surface nodes (w in number) are used for excitation
and the other (m − w) nodes are used for detection. Let A1 (λi ) ∈ Rn×w denote the mapping
from the w surface nodes to n internal points computed using (1) and B1 (λj ) ∈ R(m−w)×n
denote the mapping from the n internal location to (m − w) surface points computed using
(2). The subscripts i and j for λ are used to indicate the excitation and emission wavelengths,
respectively. If we assume that w nodes are illuminated by an excitation source such that
ζd , d = 1, 2, . . . , w, denotes the source strength at the w nodes, then the total photon flux
at the n internal nodes due to this surface illumination at wavelength λi is given by a linear
combination of the columns of A1 (λi ), i.e.
 w 

a1 (λi ) = ζd Ad1 (λi ) , (4)
d=1

with a1 (λi ) ∈ Rn and Ad1 representing the dth column of A1 (λi ). The combined mapping from
the surface excitation to the internal points and then from the internal points to the surface can
be represented as
X1 (λi , λj ) = ηB1 (λj ) diag{a1 (λi )} (5)
with X1 (λi , λj ) ∈ R (m−w)×n
denoting the mapping with excitation wavelength λi and emission
wavelength λj . The second set of measurements assumes that the remainder (m − w) nodes
4692 A J Chaudhari et al

(a) (b)

Figure 1. Optical setup used for simulation experiments. (a) The first configuration with
illumination on the top surface and detection from the bottom surface and (b) the second
configuration with illumination on the bottom surface and detection from the top and sides.
Multiple views are necessary for tomography and are acquired by the two-mirror setup (Chaudhari
2006).

are used for excitation and w nodes are used as detectors. The corresponding excitation and
emission models at λi and λj are represented by A2 (λi ) ∈ Rn×(m−w) and B2 (λj ) ∈ Rw×n ,
respectively. Let X2 (λi , λj ) ∈ Rw×n represent the combined mapping analogous to (5) for
this second configuration. Thus, the OFT forward model for the complete animal volume
for a single illumination wavelength λi and a single emission wavelength λj can be written
as
 T
Xij = XT1 (λi , λj ) XT2 (λi , λj ) , Xij ∈ Rm×n . (6)
The optical setup for the two configurations considered for this paper is shown in figure 1.
The mouse surface is assumed to have a total of m surface nodes. One geometrical subset
corresponds to uniform illumination (ζd = constant for all d) on the top surface at w nodes
and detection from the bottom at (m − w) nodes as indicated in figure 1(a), while the
second corresponds to uniform illumination on the bottom surface at (m − w) nodes and
data acquisition from the top and sides at w nodes. Here, we have considered only two
geometrical subsets of the nodes for simplicity. Our mathematical formulation, however, can
easily be extended to any number of spatial subsets where (6) would transform to
 T
Xij = XT1 (λi , λj ) XT2 (λi , λj ) · · · XTf (λi , λj ) , Xij ∈ Rm×n , (7)
where f represents the total number of spatial subsets that span the whole animal surface.
Assume that the fluorophore is excited at y wavelengths over the absorption band and
optical data are measured for z spectral bins over the emission wavelength band of interest.
Let s = [s(λ1 ) s(λ2 ) . . . s(λy )]T , s ∈ Ry , and t = [t (λ1 ) t (λ2 ) . . . t (λz )]T , t ∈ Rz ,
represent the absorption and emission spectra respectively of the fluorophore. Then, the four
forward models whose description is given in table 1 are calculated as indicated in table 2.
The formulae for the assembly of the SWI-AD and SWI-MD forward models assume that the
excitation wavelength is λi .
Forward model dimensions are listed in table 2, column 3. The MWI-MD forward model
is y-times larger than the SWI-MD model and z-times larger than the MWI-AD model. In
Excitation spectroscopy in multispectral fluorescence tomography 4693

Table 2. Formulae for the computation of the optical forward model

Model Derived formulae Size


z
SWI-AD XSWI−AD
i = s(λi ) j =1 t λj Xij Rm×n
T
SWI-MD XSWI−MD
i = s(λi )t (λ1 )XTi1 s(λi )t (λ2 )XTi2 · · · s(λi )t (λz )XTiz Rzm×n
 T  T  T T
MWI-AD XMWI−AD = XSWI−AD
1 XSWI−AD
2 ··· XSWI−AD
y Rym×n
 T  T  T T
MWI-MD XMWI−MD = XSWI−MD
1 XSWI−MD
2 ··· XSWI−MD
y Ryzm×n

all cases, the data b can be represented as a linear transformation on the underlying source
distribution q (Arridge 1999, Boas et al 2001, Ntziachristos et al 2002):
b = Xq, (8)
where we have dropped the model descriptor subscript on X. The size of b for each model
will be the row dimension of that model.

2.2. Inverse methods

The OFT inverse problem is to solve (8) for the unknown source distribution q for given data
b. We focus on least-squares estimation with L2 -norm regularization (Tikhonov and Arsenin
1977):
1 α
q̂ = arg min (q), (q) = b − Xq2 + q2 , (9)
q 2 2
where α denotes the regularization parameter that controls the trade-off between spatial
resolution and variance. The solution of (9) can be written as in Dahlquist and Bjorck
(2008):
−1
q̂ = XT X + αI XT b = VKVT XT b, (10)
where we have expanded X using the SVD as X = UVT , I is the identity matrix and
 
1
K = diag , K ∈ Rn×n , p = 1, 2, . . . , n.
σp2 + α

Here, σp2 denotes the square of the pth singular value. We can interpret XT b as a back-
projection of the data into image space, which is then filtered by the VKVT operator to give
us our estimate q̂. We have deliberately written (10) in terms of the singular values (σp ) and
the right singular component (V) of X. This is because the computation of the full SVD of X
is expensive, especially for the MWI-MD model. However, XT X = V 2 VT ∈ Rn×n , where
 2 ∈ Rn×n is a diagonal matrix containing the squared singular values of X. Therefore, the
solution in (10) can be found by computing only the right singular vectors of X or equivalently
the eigenvectors of XT X. Since typically n is much smaller than the number of rows in X,
this is a much less demanding computation than the full SVD of X. The major benefit of
the regularized pseudo-inverse method for our case, rather than more complex nonlinear or
constrained estimators, is that it is linear, i.e. the response of the corresponding estimator to
arbitrary source distributions can be modeled based on its point spread function alone. Thus,
a direct comparison between the forward models becomes relatively straightforward.
4694 A J Chaudhari et al

100
Absorption
90 Emission

80

70

Normalized intensity
60

50

40

30

20

10

0
500 550 600 650 700 750 800 850
Wavelength (nm)

(a) (b)

Figure 2. Assignment of optical properties and the absorption and emission spectra of Alexa 700
R
.
(a) Horizontal cut through the 3D tetrahedral mesh representation of the Digimouse atlas showing
the assignment of the absorption coefficient (mm−1 ) (left) and the optical diffusion coefficient
(mm−1 ) (right) to internal organs at wavelength 760 nm and (b) the characteristic absorption and
emission spectra of the Alexa 700 R
dye. For experiments in this paper, the 580–680 nm band was
used for excitation while the 700–800 nm band was used for multispectral data acquisition.

2.3. Computation of the forward models

The 3D labeled Digimouse mouse atlas5 was used for all studies. The volumetric tessellation
consisted of 306 773 tetrahedral elements connected at 58 244 nodes. Optical properties as a
function of wavelength were assigned to all 17 organs based on published data (Alexandrakis
et al 2005). An example of the organ-wise distribution of optical properties is shown in
figure 2(a). The fluorescent probe was assumed to be the Alexa 700 R
dye whose normalized
absorption and emission spectra are shown in figure 2(b). This dye has been used in vivo
in small animals for imaging metastatic disease (Koenig et al 2008). The forward models
were computed for the optical setup in figure 1. In this setup, the mouse surface was first
illuminated uniformly from the top and the photon fluence due to surface illumination was
computed at n = 8903 internal source nodes (spatial sampling of 1.2 mm) by a fast FEM-based
solver (Chaudhari 2006, Ahn et al 2008a). Then, a mapping from this source grid to w =
1455 nodes on the bottom surface was calculated. This procedure was repeated assuming that
the bottom surface of the mouse was illuminated uniformly and that data at (m − w) = 1778
nodes on the top surface were collected. This two-step process was carried out for y = 6
illumination wavelength bins (580–680 nm in steps of 20 nm) and for z = 6 emission spectral
bins (700–800 nm in steps of 20 nm). The SWI-AD and SWI-MD forward models were
assembled assuming that the illumination was at λi = 680 nm. The MWI-MD model used
data at all six illumination wavelengths and in all six emission bands. While computing the
SWI-AD model, we increased the detector spatial sampling by six times so that this model has
the same row dimension as the SWI-MD or the MWI-AD model. All forward models were
computed using a custom FEM code written in MATLAB R 6
.

5 http://neuroimage.usc.edu/Digimouse.html.
6 The MathWorks, Inc, Natick, MA, USA.
Excitation spectroscopy in multispectral fluorescence tomography 4695

2.4. Singular value decomposition study

For each of the forward models, we computed the singular values and the right singular vectors
(V) of X. The singular value spectra were plotted for each model. Comparison of only singular
values for OFT may at times be misleading since they only reflect the power along the SVD
bases, but do not give any information about the spatial distribution of the corresponding
basis vectors (Chaudhari 2006). Therefore, we have also performed a spatial resolution versus
variance analysis to further investigate our model comparisons.

2.5. Analysis of reconstructed spatial resolution versus estimator variance

In imaging systems, there inherently exists a trade-off between spatial resolution recovered
in the image reconstruction process and the statistical estimator variance, that is, sharpened
(higher) spatial resolution comes at the expense of also having higher variance (noise) in the
reconstructed image (Alessio and Kinahan 2006). Resolution here is defined as the level of
reproduction of spatial detail by the imaging system, while the variance is a measure of how
much the estimate varies around its average value. Thus, for a fixed value of variance, an
estimator providing the highest spatial resolution is superior. On the other hand, for a fixed
value of spatial resolution, the estimator corresponding to the least variance will correspond to
a more stable solution. We derive estimators from the computed forward models. Resolution-
variance analysis then provides a means for quantitatively comparing these estimators and, in
turn, the forward models.
We model the optical data in (8) as
b = Xq + n, (11)
where we assume the noise n to be Gaussian and white, such that E(n) = 0 and
cov(n) = E(nnT ) = λ2 I (Roy et al 2003). Since E(q̂) = (XT X + αI)−1 XT Xq, the point
spread function at voxel j (PSFj ) would be given by
 
−1 T
σp2
PSFj = (X X + αI) X Xej = Vdiag
T
VT ej , (12)
σp2 + α

where ej is an elementary vector with value 1 at voxel j and p = 1, 2, . . . , n. The spatial

resolution for a source at the j th voxel can be defined as the full width at half-maximum
(FWHM) of the PSF. The total variance, summed over all voxels, is given by
   
n
σp2
var(q̂p ) = λ Tr Vdiag
2
VT . (13)
σ 2+α 2
p=1 p

We chose 100 representative source locations that were at least 3 mm deep in tissue. The source
distribution was reconstructed using (10), and the corresponding spatial resolution (averaged
over the axial, sagittal and coronal directions) and total variance summed over all voxels were
computed. The regularization parameter α was varied from 10−6 to 10−13 in steps of 10−1 .
The spatial resolutions computed for all 100 sources for a fixed regularization parameter were
averaged. For realistic simulations, we added zero-mean Gaussian noise to the data such that
the signal-to-noise ratio (SNR) for MWI-MD was 100 (20 dB). This value was chosen based
on the performance of currently available systems (Ntziachristos and Weissleder 2002). For
a fair comparison, since it takes six times longer to acquire data for MWI-MD compared to
SWI-MD or MWI-AD (assuming that the scan times for each spectral bin for excitation and
emission are the same), we assume that data for the latter two were measured in a six times
4696 A J Chaudhari et al

0
10
SWI−AD
SWI−MD
MWI−AD

Normalized singular values

−5 MWI−MD
10

−10
10

−15
10

−20
10
0 1000 2000 3000 4000 5000 6000 7000 8000
Singular value index

Figure 3. Comparison of singular value spectra for the forward models. For the SWI-AD model,
we have used six times more surface detectors for a fair comparison.

longer scan. This assumption then decreased the noise variance λ2 for the SWI-MD and
MWI-AD models by a factor of 6. This same assumption also decreased the noise variance
λ2 by a factor of 36 for the SWI-AD case. With these consideration, the SNRs for the models
became 100 or 20 dB (MWI-MD), 600 or 27.8 dB (SWI-MD and MWI-AD), and 3600 or
35.5 dB (SWI-AD). We also assumed that the signal strength remained the same over the
duration of the scan.

2.6. Imaging comparisons in a realistic mouse model of human hepatocellular carcinoma

For creating metastatic mouse models of HCC, histologically intact specimens from patients
suffering from HCC are implanted directly into the liver of nude mice (Sun et al 1996). This
process allows monitoring of the specimen’s orthotopic growth and metastases over time, and
provides a means of understanding and predicting clinical behavior of HCC (Fu et al 1991).
To simulate HCC in the mouse, we first artificially created a primary implantation site in the
liver of the Digimouse based on published data (Hoffman 1999). The longest diameter of this
implanted lesion was approximately 8 mm. Since metastatic lesions on follow-up can also be
found within the liver itself, we created one small metastatic lesion (approximately 1.2 mm in
diameter) on the liver boundary at about 5 mm from the primary lesion. We assume that Alexa
700 will accumulate at the two sites due to increased blood flow from tumor neovascularization.
A representative horizontal section through the CT of the Digimouse showing the simulated
spatial distribution of Alexa 700 is shown in figure 5(a). This distribution in reality is in
3D though only 2D sections are shown for visual clarity. For imaging, simulated data were
generated using the optical setup in figure 1 and using the method described in section 2.3.
Images were reconstructed using the regularized pseudo-inverse method. The SNRs were set
to 20 dB (MWI-MD), 27.8 dB (SWI-MD and MWI-AD) and 35.5 dB (SWI-AD) as described
in section 2.5. The regularization parameter in each case was chosen from the resolution-
variance plot to correspond to equal noise variance. This allowed the noise levels to be matched
for the four types of data acquisition.
Excitation spectroscopy in multispectral fluorescence tomography 4697

0
10
SWI−AD
SWI−MD
−1 MWI−AD
10 MWI−MD

Relative variance −2
10

−3
10

−4
10

−5
10
1.6 2 2.4 2.8 3.2 3.6 4
FWHM (mm)

Figure 4. Average spatial resolution versus relative estimator variance as a function of

regularization parameter α. Relative variance is the ratio of the computed variance to the variance
of the SWI-AD model at α = 10−13 .

3. Results

3.1. The SVD study

Our results from the SVD of the four forward models are shown in figure 3. The SWI-
AD model has better conditioning compared to other models for higher singular values;
however, the drop-off of the spectrum is steep. Therefore, it will tend to be less robust to
noise compared to other models. The SWI-MD model has better conditioning than the SWI-
AD model implying that there certainly is a reduction in ill-posedness because multispectral
detection of emission data was used. The observation that multispectral detection was indeed
useful is further strengthened by the comparison between MWI-AD and MWI-MD where the
same excitation wavelength bins were used. In this case, the MWI-MD model clearly has
better conditioning than MWI-AD.
A comparison between the SWI-AD, SWI-MD and MWI-AD models indicates that
although there certainly is an advantage when multispectral detection is used, the improvement
in conditioning achieved is less than that achieved through illumination at multiple
wavelengths. The MWI-MD model has the best conditioning among all four models indicating
that information in both illumination and emission domains need to be exploited for optimal
outcomes.

3.2. Resolution-variance comparisons

The results of our resolution-variance analysis are shown in figure 4, where we have plotted
the average spatial resolution, measured by the full width at half-maximum of a point spread
function, over the 100 representative sources against relative estimator variance as a function
of the regularization parameter α. The relative variance in each case is defined as the ratio of
the computed variance to the variance of the SWI-AD model at α = 10−13 . For a constant
spatial resolution value and over all values of α, the MWI-MD model has the least variance.
This, in turn, implies that for any fixed value of spatial resolution, the images reconstructed
4698 A J Chaudhari et al

Figure 5. Reconstruction from noisy data in a mouse model of human hepatocellular carcinoma.
(a) A representative horizontal section from the CT of the Digimouse (gray scale) with the simulated
distribution of the Alexa dye overlaid (in color), reconstructed images obtained from using the
SWI-AD (b), SWI-MD (c), MWI-AD (d) and MWI-MD (e) models overlaid on the CT section. The
regularization parameter in each case was chosen from the resolution-variance plots to correspond
to equal noise variance. The reconstructed optical images were linearly interpolated to match the
CT image resolution. Reconstructed values greater than or equal to 10% of the maximum value
are displayed. Reconstructed images without interpolation and thresholding are shown in figure 6.

using the MWI-MD model will deviate least from their mean values compared to other models.
Thus, this model provides a more stable solution compared to other models.
The SWI-AD model has the worst resolution-variance characteristics, while the MWI-
AD model performs better than the SWI-MD model overall. We observe a relatively large
SNR gain (values between 2.5 and 20, with an average of 8) for the MWI-MD-based results
compared to those based on MWI-AD. This SNR gain for MWI-MD on an average was
close to 20 compared to the SWI-MD results and in the thousands for SWI-AD. Over all
variance values, we consistently observe a 0.2–0.3 mm and 0.5–0.6 mm improvement in
spatial resolution for the MWI-MD-based results compared to those obtained for the MWI-
AD and SWI-MD models respectively at matched noise levels. With respect to SWI-AD, the
MWI-MD-based results show an average spatial resolution improvement of close to 1.5 mm.
These results validate our findings from the SVD study that multispectral detection is indeed
useful, illumination at multiple wavelengths is better than illumination at a single wavelength
and information from both illumination and emission domains must be used synergistically to
obtain the best result.

3.3. Image reconstruction study in a realistic mouse phantom

Results from our reconstruction study are shown in figure 5. In figure 5(b), we show an
image reconstructed using the SWI-AD approach. The reconstructed distribution shows only
broad similarity to the true distribution from figure 5(a), with the deeper structure and the
smaller lesion not visualized. Even in figure 5(c), where a reconstruction result using the
SWI-MD approach is shown, superficial parts of the primary site are reconstructed better
than the deeper parts. This effect can be attributed to decreased SNR for deep sources and
Excitation spectroscopy in multispectral fluorescence tomography 4699

Figure 6. Reconstructed optical images from figure 5 without any thresholding. (a) The simulated
distribution of Alexa 700. (b)–(e) Reconstructed images using the SWI-AD, SWI-MD, MWI-AD
and MWI-MD models, respectively. The color scale used for these images is the inverse of the
conventional gray scale.

hence decreased spatial resolution that consequently blurs out the signal. Mathematically,
noise is able to significantly affect the small singular values which typically carry depth
information (Chaudhari 2006). The reconstruction result using the MWI-AD model is shown in
figure 5(d). In this case, deeper sections of the primary lesion appear to be reconstructed better
than those in reconstructed images generated from the SWI-MD approach. The best result is
obtained using the MWI-MD approach and the corresponding reconstructed source distribution
is shown in figure 5(e). The reconstructed image in this case shows the metastatic lesion in
addition to the primary lesion. This metastatic lesion is deep in tissue and was not visible in the
reconstructed images from the other three models. Thus, MWI-MD approaches provide the
best quantitative ability and depth sensitivity (both by-products of improved spatial resolution
and robustness to noise), while the SWI-AD model provides the worst. Reconstructed images
without thresholding and without interpolation (required to match to the CT image resolution)
are shown in figure 6.

4. Discussion

The imaging performance using forward models that incorporated excitation at multiple
wavelengths was compared to models that used excitation at a single wavelength and
achromatic or multispectral detection. Multispectral detection techniques led to better
conditioned forward models compared to those that used achromatic detection. Excitation
at multiple wavelengths added another favorable dimension to the otherwise ill-posed OFT
inverse problem. An imaging study conducted using a realistic mouse model showed the
potential of our excitation spectroscopy approach in multispectral OFT for imaging of small
lesions deep in tissue. This study also confirmed our findings from SVD and resolution-
variance analyses. Because the focus of this paper was spectral-domain analysis, we chose
to have the animal surface uniformly illuminated. The conditioning of all our models will
improve even further if optimal spatial distributions for the illumination (Joshi et al 2006b,
Deliolanis et al 2008) are coupled with our spectral approaches.
4700 A J Chaudhari et al

The proposed method capitalizes on the fact that higher variability in tissue optical
properties is observed for illumination photons compared to emission photons. This variability
then leads to some amount of independent information in the models and yields better
performance for image reconstruction. The results shown in this paper were limited to
one representative fluorophore. We anticipate that the benefits of excitation spectroscopy
would reduce compared to those shown in this work for fluorescent probes that have their
absorption spectrum red-shifted compared to that of Alexa 700. However, even for these
probes, there will be higher variability of optical properties for excitation photons compared
to emission photons. Thus, MWI-MD approaches will still be valuable. Better results than
those shown here may be achieved for probes whose absorption spectra blue-shift with respect
to that of Alexa 700. However, there will be limited penetration of excitation light in tissue at
these wavelengths. Thus, careful determination of the benefits of the MWI-MD approach is
necessary for different probes. We recommend the SVD and resolution-variance analyses as
quantitative methods to facilitate this determination.
Imaging systems that are able to measure multispectral emission data have become
available commercially. The proposed method additionally requires the capability of
illuminating the subject at multiple wavelengths. One possible way of achieving this is
by using a broadband light source with a tunable filter (Gao et al 2004, Leavesley et al
2008). The tunable filter in this case should ideally have a narrow passband, sharp cutoff, high
attenuation in the stop-band and fast switching times. The broadband source should possess
sufficient power so that light levels after filtering are suitable for imaging. A second technique
of achieving excitation at multiple wavelengths may involve separate lasers for illumination,
but the cost of this setup may be significantly higher.
One limitation of the presented study is that tissue autofluorescence was not taken into
account explicitly. Tissue autofluorescence is a function of wavelength (Anderson and Parrish
1981) and thus the excitation wavelengths to be used need to be chosen carefully. The influence
of autofluorescence on the optical reconstruction result may be reduced by machine learning
approaches (Mansfield et al 2005) or by explicitly accounting for it in the formulation of the
forward model. The latter technique needs further investigation.
In this paper, a large number of reconstructed images were needed for our evaluation
and thus we used the direct regularized pseudo-inverse-based solution. The assembly of all
the models took close to 5 h on a Dual Core AMD OpteronTM 2.33 GHz processor. The
computation of the SVD for each model took close to 25 min. Additionally, no non-negativity
constraint was enforced on the solution. Fast iterative approaches for on-the-fly computation
of the forward models and reconstructed images that incorporate non-negativity constraints
have been investigated (Ahn et al 2008a, 2008b). Our proposed method will greatly benefit
from these developments.

5. Conclusions

We have proposed a method for OFT that utilizes excitation spectroscopic information
in addition to the detection of optical multispectral emission data. We compared this
approach with conventional approaches that use only illumination at a single wavelength
with either achromatic or multispectral data using one representative fluorescent dye (Alexa
700 R
). Studies involving singular value analyses and spatial resolution versus variance curves
unanimously show benefits of using our approach of excitation spectroscopy. Our results in a
simulated realistic mouse model of human hepatocellular carcinoma show that the proposed
approach allows for exploiting both excitation and emission domain information to reduce the
ill-posedness of the OFT inverse problem. Our results also indicate that for the Alexa 700 R
Excitation spectroscopy in multispectral fluorescence tomography 4701

dye, excitation spectroscopy was better suited for extracting depth information compared to
emission spectroscopy (measurement of multispectral emission data). This should be true for
NIR probes in general, since the excitation band is blue-shifted compared to the emission band
and thus more depth-dependent spectral information will be available in the excitation band.
There are three critical limitations of this study. First, tissue autofluorescence has not been
taken into consideration. Second, the diffusion approximation to the RTE used for forward
model generation may not be accurate in certain anatomical regions of the animal. Third,
we have demonstrated our method for only one representative dye and thus our quantitative
results have limited scope. Careful determination of the benefits of the proposed approach is
necessary for different probes.

Acknowledgments

The authors would like to thank Dr Felix Darvas from the University of Washington-Seattle
and Dr Gregory S Mitchell from the University of California-Davis for useful discussions.
This work was funded in part by the National Institutes of Health Grant R01CA121783 and
R44CA13824, by the American Cancer Society award IRG-95-125-07 and by the Susan G
Komen Foundation award BCTR0707455. This work was also made possible by grant no
UL1 RR024146 from the National Center for Research Resources (NCRR), a component of
the National Institutes of Health (NIH) and NIH Roadmap for Medical Research.

References

Adams K, Ke S, Kwon S, Liang F, Fan Z, Lu Y, Hirschi K, Mawad M, Barry M and Sevick-Muraca E 2007
Comparison of visible and near-infrared wavelength-excitable fluorescent dyes for molecular imaging of cancer
J. Biomed. Opt. 12 024017
Ahn S, Chaudhari A J, Darvas F, Bouman C A and Leahy R M 2008a Fast iterative image reconstruction methods for
fully 3D multispectral bioluminescence tomography Phys. Med. Biol. 53 3921–42
Ahn S, Dutta J, Chaudhari A J and Leahy R M 2008b Computationally efficient image reconstruction methods for
multispectral optical fluorescence tomography using FEM-based forward models Abstract Book, Joint Molecular
Imaging Conf. pp 329–30
Alessio A and Kinahan P 2006 PET Image Reconstruction, Nuclear Medicine 2nd edn (Amsterdam: Elsevier)
Alexandrakis G, Rannou F R and Chatziioannou A F 2005 Tomographic bioluminescence imaging by use of a
combined optical-PET (OPET) system: a computer simulation feasibility study Phys. Med. Biol. 50 4225–41
Anderson R and Parrish J 1981 The optics of human skin. J. Invest. Dermatol. 77 13–9
Arridge S 1999 Optical tomography in medical imaging Inverse Problems 15 41–93
Arridge S R and Hebden J C 1997 Optical imaging in medicine: II. Modelling and reconstruction Phys. Med. Biol.
42 841–53
Arridge S R, Schweiger M, Hiroaka M and Delpy D T 1993 A finite element approach for modeling photon transport
in tissue Phys. Med. Biol. 20 299–309
Axelsson J, Svensson J and Andersson-Engels S 2007 Spatially varying regularization based on spectrally resolved
fluorescence emission in fluorescence molecular tomography Opt. Express 15 13574–84
Bargo P, Prahl S A, Goodell T T, Sleven R A, Koval G, Blair G and Jacques S L 2005 In vivo determination of optical
properties of normal and tumor tissue with white light reflectance and an empirical light transport model during
endoscopy J. Biomed. Opt. 10 034018
Boas D, Brooks D H, Miller E L, DiMarzio C A, Kilmer M, Gaudette R J and Zhang Q 2001 Imaging the body with
diffuse optical tomography IEEE Signal Process. Mag. 18 57–75
Chance B 1991 Optical method Annu. Rev. Biophys. Biophys. Chem. 20 1–28
Chang J, Graber H L and Barbour R L 1997 Imaging of fluorescence in highly scattering media IEEE Trans. Biomed.
Eng. 44 810–22
Chaudhari A J 2006 Hyperspectral and multispectral optical bioluminescence and fluorescence tomography in small
animal imaging PhD Thesis University of Southern California
Chaudhari A J, Darvas F, Bading J R, Moats R A, Conti P S, Smith D J, Cherry S R and Leahy R M 2005 Hyperspectral
and multispectral bioluminescence optical tomography for small animal imaging Phys. Med. Biol. 50 5421–41
4702 A J Chaudhari et al

Chaudhari A J, Joshi A A, Darvas F and Leahy R M 2007 A method for atlas-based volumetric registration with
surface constraints for optical bioluminescence tomography in small animal imaging Proc. SPIE 6510 651024
Cheong W F, Prahl S A and Welch A J 1990 A review of the optical properties of biological tissues IEEE J. Quantum
Electron. 26 2166–85
Cherry S R 2004 In vivo molecular and genomic imaging: new challenges for imaging physics Phys. Med. Biol.
49 R13–48
Cong A and Wang G 2005 A finite-element-based reconstruction method for 3D fluorescence tomography Opt.
Express 13 9847–57
Cong A X and Wang G 2006 Multispectral bioluminescence tomography: methodology and simulation Int. J. Biomed.
Imaging 2006 57614
Dahlquist G and Bjorck A 2008 Numerical Methods in Scientific Computing vol 1 (Philadelphia, PA: Society for
Industrial and Applied Mathematics)
Dehghani H, Arridge S, Schweiger M and Delpy D 2000 Optical tomography in the presence of void regions J. Opt.
Soc. Am. A 17 1659–70
Dehghani H, Davis S C, Jiang S, Pogue B W, Paulsen K D and Patterson M S 2006 Spectrally resolved bioluminescence
optical tomography Opt. Lett. 3 365–7
Deliolanis N, Kasmieh R, Wurdinger T, Tannous B, Shah K and Ntziachristos V 2008 Performance of the red-shifted
fluorescent proteins in deep-tissue molecular imaging applications J. Biomed. Opt. 13 044008
Dogdas B, Stout D, Chatziioannou A F and Leahy R M 2007 Digimouse: a 3D whole body mouse atlas from CT and
cryosection data Phys. Med. Biol. 52 577–87
Fu X, Besterman J, Monosov A and Hoffman R 1991 Models of human metastatic colon cancer in nude mice
orthotopically constructed by using histologically intact patient specimens Proc. Natl Acad. Sci. 88 9345–9
Gao X, Cui Y, Levenson R, Chung L and Nie S 2004 In vivo cancer targeting and imaging with semiconductor
quantum dots Nature Biotechnol. 22 969–76
Giepmans B N G, Adams S R, Ellisman M H and Tsien R Y 2006 The fluorescent toolbox for assessing protein
location and function Science 312 217–24
Godavarty A, Zheng C, Eppstein M J and Sevick-Muraca E M 2004 Fluorescence enhanced optical imaging of large
phantoms using single and simultaneous dual point illumination geometries Med. Phys. 31 183–90
Guven M, Yazici B and Ntziachristos V 2007 Fluorescence optical tomography with a priori information Proc.
SPIE 6431 643107
Haskell R C, Svaasand L O, Tsay T T, Feng T C, McAdams M S and Tromberg B J 1994 Boundary conditions for
the diffusion equation in radiative transfer J. Opt. Soc. Am. A 11 2727–40
Hassan M, Riley J, Chernomordik V, Smith P, Pursley R, Lee S B, Capala J and Gandjbakhche A H 2007 Fluorescence
lifetime imaging system for in vivo studies Mol. Imaging 6 229–36
Hielscher A 2005 Optical tomographic imaging of small animals Curr. Opin. Biotechnol. 16 79–88
Hoffman R 1999 Orthotopic metastatic mouse models for anticancer drug discovery and evaluation: a bridge to the
clinic Invest. New Drugs 17 343–60
Joshi A, Bangerth W and Sevick-Muraca E M 2004 Adaptive finite element based tomography for fluorescence optical
imaging in tissue Opt. Express 12 5402–17
Joshi A, Bangerth W, Hwang K, Rasmussen J C and Sevick-Muraca E M 2006a Fully adaptive FEM based
fluorescence optical tomography from time-dependent measurements with area illumination and detection
Med. Phys. 33 1299
Joshi A, Bangerth W and Sevick-Muraca E 2006b Non-contact fluorescence optical tomography with scanning
patterned illumination Opt. Express 14 6516–34
Joshi A, Rasmussen J, Sevick-Muraca E, Wareing T and McGhee J 2008 Radiative transport-based frequency-domain
fluorescence tomography Phys. Med. Biol. 53 2069–88
Klose A, Ntziachristos V and Hielscher A 2005 The inverse source problem based on the radiative transfer equation
in optical molecular imaging J. Comput. Phys. 202 323–45
Klose A D and Hielscher A H 2003 Fluorescence tomography with simulated data based on the equation of radiative
transfer Opt. Lett. 28 1019–21
Koenig A, Herve L, Josserand V, Berger M, Boutet J, da Silva A, Dinten J, Peltie P, Coll J and Rizo P 2008 In vivo
mice lung tumor follow-up with fluorescence diffuse optical tomography J. Biomed. Opt. 13 011008
Kumar A T N, Raymond S B, Dunn A K, Bacskai B J and Boas D A 2008 A time domain fluorescence tomography
system for small animal imaging IEEE Trans. Med. Imaging 27 1152–63
Kuo C, Coquoz O, Troy T L, Xu H and Rice B W 2007 Three-dimensional reconstruction of in vivo bioluminescent
sources based on multispectral imaging J. Biomed. Opt. 12 024007
Lasser T, Soubret A, Ripoll J and Ntziachristos V 2008 Surface reconstruction for free-space 360o fluorescence
molecular tomography and the effects of animal motion IEEE Trans. Med. Imaging 27 188–94
Excitation spectroscopy in multispectral fluorescence tomography 4703

Leavesley S, Jiang Y, Patsekin V, Rajwa B and Robinson J 2008 An excitation wavelength-scanning spectral imaging
system for preclinical imaging Rev. Sci. Instrum. 79 023707
Luker G and Luker K 2008 Optical imaging: current applications and future directions J. Nucl. Med. 49 1–4
Lv Y, Tian J, Cong W, Wang G and Kumar D 2006 MicroCT-guided bioluminescence tomography based on the
adaptive finite element tomographic algorithm 28th Ann. Meeting IEEE Engineering in Medicine and Biol
Society pp 381–4
Mansfield J, Gossage K, Hoyt C and Levenson R 2005 Autofluorescence removal, multiplexing, and automated
analysis methods for in-vivo fluorescence imaging J. Biomed. Opt. 10 041207
Massoud T F and Gambhir S S 2003 Molecular imaging in living subjects: seeing fundamental biological processes
in a new light Genes Dev. 17 545–80
Meyer H, Garofalakis A, Zacharakis G, Psycharakis S, Mamalaki C, Kioussis D, Economou E, Ntziachristos V and
Ripoll J 2007 Noncontact optical imaging in mice with full angular coverage and automatic surface extraction
Appl. Opt. 46 3617–27
Milstein A B, Oh S, Webb K J, Bouman C A, Zhang Q, Boas D A and Millane R P 2003 Fluorescence optical diffusion
tomography Appl. Opt. 42 3081–94
Neefjes J and Dantuma N P 2004 Fluorescent probes for proteolysis: tools for drug discovery Nature Rev. Drug
Discov. 3 58–69
Ntziachristos V 2006 Fluorescence molecular tomography Annu. Rev. Biomed. Eng. 8 1–33
Ntziachristos V, Ripoll J, Wang L V and Weissleder R 2005 Looking and listening to light: the evolution of whole-body
photonic imaging Nature Biotechnol. 23 313–20
Ntziachristos V, Ripoll J and Weissleder R 2002 Would near-infrared fluorescence signals propagate through large
human organs for clinical studies? Opt. Lett. 27 333–5
Ntziachristos V, Schellenberger E A, Ripoll J, Yessayan D, Graves E, Bogdanov A Jr, Josephson L and
Weissleder R 2004 Visualization of antitumor treatment by means of fluorescence molecular tomography
with an annexin V-Cy5.5 conjugate Proc. Natl Acad. Sci. 101 12294–9
Ntziachristos V and Weissleder R 2001 Experimental three-dimensional fluorescence reconstruction of diffuse media
by use of a normalized born approximation Opt. Lett. 26 893–5
Ntziachristos V and Weissleder R 2002 Charge-coupled-device based scanner for tomography of fluorescent near-
infrared probes in turbid media Med. Phys. 29 803–9
Paithankar D Y, Chen A U, Pogue B W, Patterson M S and Sevick-Muraca E M 1997 Imaging of fluorescent yield
and lifetime from multiply scattered light reemitted from random media Appl. Opt. 36 2260–72
Patterson M S and Pogue B W 1994 Mathematical model for time-resolved and frequency-domain fluorescence
spectroscopy in biological tissues Appl. Opt. 33 1963–74
Patterson M S, Wilson B C and Wyman D R 1992 The propagation of optical radiation in tissue: I. Models of radiation
transport and their application Lasers Med. Sci. 6 155–68
Patwardhan S, Bloch S, Achilefu S and Culver J 2005 Time-dependent whole-body fluorescence tomography of probe
bio-distributions in mice Opt. Express 13 2564–77
Paulsen K D and Jiang H 1995 Spatially-varying optical property reconstructions using a finite element diffusion
equation approximation Med. Phys. 22 691–701
Pinaud F, Michalet X, Bentolila L A, Tsay J M, Doose S, Li J J, Iyer G and Weiss S 2006 Advances in fluorescence
imaging with quantum dot bio-probes Biomaterials 27 1679–87
Pogue B W, Gibbs S L, Chen B and Savellano M 2004 Fluorescence imaging in vivo: raster scanned point-source
imaging provides more accurate quantification than broad beam geometries Technol. Cancer Res. Treat. 3 15–21
Ren K, Bal G and Hielscher A 2007 Transport-and diffusion-based optical tomography in small domains: a
comparative study Appl. Opt. 46 6669–79
Reynolds J S, Thompson C A, Webb K J, LaPlant F P and Ben-Amotz D 1997 Frequency domain modeling of
reradiation in highly scattering media Appl. Opt. 36 2252–9
Rice B W, Cable M D and Nelson M B 2001 In vivo imaging of light-emitting probes J. Biomed. Opt. 6 432–40
Rice B W, Xu H and Kuo C 2006 Surface construction using combined photographic and structured light information
Patent Application 0268153
Ripoll J and Ntziachristos V 2004 Imaging scattering media from a distance: theory and applications of non-contact
optical tomography Mod. Phys. Lett. B 18 1403–31
Ripoll J, Schulz R B and Ntziachristos V 2003 Free-space propagation of diffuse light: theory and experiments Phys.
Rev. Lett. 91 103901
Roy R, Godavarty A and Sevick-Muraca E 2003 Fluorescence-enhanced optical tomography using referenced
measurements of heterogeneous media. IEEE Trans. Med. Imaging 22 824–36
Roy R, Godavarty A and Sevick-Muraca E M 2006 Fluorescence-enhanced optical tomography of a large tissue
phantom using point illumination geometries J. Biomed. Opt. 11 044007
4704 A J Chaudhari et al

Schulz R, Ripoll J and Ntziachristos V 2004 Experimental fluorescence tomography of tissues with noncontact
measurements IEEE Trans. Med. Imaging 23 492–500
Soubret A, Ripoll J and Ntziachristos V 2005 Accuracy of fluorescent tomography in the presence of heterogeneities:
study of the normalized Born ratio IEEE Trans. Med. Imaging 24 1377–86
Sun F, Tang Z, Liu K, Xue O, Gao D, Yu Y, Zhou X and Ma Z 1996 Metastatic models of human liver cancer
in nude mice orthotopically constructed by using histologically intact patient specimens J. Cancer Res. Clin.
Oncol. 122 397–402
Swartling J, Svensson J, Bengtsson D, Terike K and Andersson-Engels S 2005 Fluorescence spectra provide
information on the depth of fluorescent lesions in tissue Appl. Opt. 44 1934–41
Tikhonov A and Arsenin V 1977 Solution of Ill-Posed Problems (New York: Wiley)
Wang G, Cong W, Durairaj K, Qian X, Shen H, Sinn P, Hoffman E, McLennan G and Henry M 2006 In vivo mouse
studies with bioluminescence tomography Opt. Express 14 7801–9
Weber W et al 2008 Technology Insight: novel imaging of molecular targets is an emerging area crucial to the
development of targeted drugs Nature Clin. Pract. Oncol. 5 44
Weissleder R and Ntziachristos V 2003 Shedding light onto live molecular targets Nature Med. 9 123–8
Weissleder R, Tung C H, Mahmood U and Bogdanov A 1999 In vivo imaging of tumors with protease-activated
near-infrared fluorescent probes Nature Biotechnol. 17 375–8
Zacharakis G, Kambara H, Shih H, Ripoll J, Grimm J, Saeki Y, Weissleder R and Ntziachristos V 2005 Volumetric
tomography of fluorescent proteins through small animals in vivo Proc. Natl Acad. Sci. 102 18252–7
Zavattini G, Vecchi S, Mitchell G, Weisser U, Leahy R M, Pichler B J, Smith D J and Cherry S R 2006 A hyperspectral
fluorescence system for 3D in vivo optical imaging Phys. Med. Biol. 51 2029–43