Sunteți pe pagina 1din 14

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO.

5, MAY 2011

1129

Compressive Diffuse Optical Tomography:


Noniterative Exact Reconstruction Using
Joint Sparsity
Okkyun Lee, Student Member, IEEE, Jong Min Kim, Yoram Bresler, Fellow, IEEE, and
Jong Chul Ye*, Member, IEEE

AbstractDiffuse optical tomography (DOT) is a sensitive and


relatively low cost imaging modality that reconstructs optical properties of a highly scattering medium. However, due to the diffusive
nature of light propagation, the problem is severely ill-conditioned
and highly nonlinear. Even though nonlinear iterative methods
have been commonly used, they are computationally expensive
especially for three dimensional imaging geometry. Recently, compressed sensing theory has provided a systematic understanding of
high resolution reconstruction of sparse objects in many imaging
problems; hence, the goal of this paper is to extend the theory to
the diffuse optical tomography problem. The main contributions of
this paper are to formulate the imaging problem as a joint sparse
recovery problem in a compressive sensing framework and to
propose a novel noniterative and exact inversion algorithm that
achieves the 0 optimality as the rank of measurement increases to
the unknown sparsity level. The algorithm is based on the recently
discovered generalized MUSIC criterion, which exploits the advantages of both compressive sensing and array signal processing.
A theoretical criterion for optimizing the imaging geometry is
provided, and simulation results confirm that the new algorithm
outperforms the existing algorithms and reliably reconstructs
the optical inhomogeneities when we assume that the optical
background is known to a reasonable accuracy.
Index TermsDiffuse optical tomography (DOT), generalized
MUSIC criterion, joint sparsity, multiple measurement vector
(MMV), p-thresholding, simultaneous orthogonal matching pursuit (S-OMP).

I. INTRODUCTION
IFFUSE optical tomography (DOT) is a sensitive and
relatively low-cost imaging modality that has recently
been studied quite extensively [1][4]. The objective of diffuse
optical tomography is to reconstruct the optical properties of

Manuscript received February 10, 2011; accepted February 27, 2011. Date of
publication March 10, 2011; date of current version May 04, 2011. This work
was supported by the Korea Science and Engineering Foundation (KOSEF)
grant funded by the Korea government (MEST) under Grant 2010-0000855.
The work of Y. Bresler was supported by the U.S. National Science Foundation
under Grant CCF-06-35234. Asterisk indicates corresponding author.
O. Lee and J. M. Kim are with the Department of Bio and Brain Engineering,
Korea Advanced Institute of Science and Technology, Daejon 305-701, Korea
(e-mail: okkyun2@kaist.ac.kr; franzkim@gmail.com).
Y. Bresler is with the Coordinated Science Lab, University of Illinois at Urbana-Champaign, Urbana, IL 61801 USA (e-mail: ybresler@uiuc.edu).
*J. C. Ye is with the Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejon 305-701, Korea (e-mail:
jong.ye@kaist.ac.kr).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TMI.2011.2125983

a cross-section of a highly scattering medium such as tissue,


based on measurements of scattered and attenuated optical
flux. The optical properties of interest include absorption and
scattering parameter variations, and fluorescent yields and lifetime. This technique presents significantly lower health risks
than X-ray imaging techniques [1][4]. Furthermore, due to
the extensive development of bioluminescence and fluorescent
molecular probes, diffuse optical tomography has emerged as
an important molecular imaging tool [4], [5].
A major difficulty in optical imaging, however, is that the
signal within tissue experiences significant scatter. In this situation, the photons should be treated as particles that elastically
scatter through a random medium. The imaging problem is
then to estimate the unknown optical coefficient variation from
the scattered field measurements at the detector. However,
the problem is highly nonlinear due to the nonlinear coupling
between the unknown absorption perturbation and the photon
flux. Furthermore, the problem is severely ill-posed due to the
diffusive nature of light propagation. Conventionally, there
have been two types of inversion approaches to deal with the
nonlinearity: iterative methods and linearization methods. The
iterative methods require recalculating the unknown flux during
each iteration. For example, the LevenbergMarquardt method
(or the so-called distorted Born iterative method) [6][8] or
Bayesian approaches using the ICD/Born method [9] belong to
the iterative reconstruction methods. In order to further reduce
the computational burden, a multigrid algorithm was proposed
[10]. However, even with these high performance algorithms,
the computational overhead of the nonlinear iterative algorithms is still prohibitive, especially for 3-D imaging, since
Greens function needs to be recalculated at each iteration.
Instead of using nonlinear iterative approaches, several groups
have employed linearization approaches. The normalized Born
approach [11], and the linearized operator approach [12][14]
are among them. Usually the computation times for these types
of algorithms are within a few minutes even for a large 3-D
phantom. However, the linearization approaches fail when
the contrast between the homogeneous background and the
unknown target is beyond the Born approximation limit.
The main contribution of this paper is a novel noniterative and
exact reconstruction algorithm that addresses the diffuse optical
tomography problem. The algorithm has three distinguishing
features. First, similarly to linearized approaches, the noniterative nature of the algorithm does not require numerical diffusion
equation solvers during iterations. This significantly reduces the

0278-0062/$26.00 2011 IEEE

1130

computational time. Second, the exactness of the algorithm overcomes the Born approximation errors that arise in linearization
methods in the presence of high contrast optical inhomogenities. Third, as will be shown in the simulation, compared to
the conventional approaches, the sparsity constraint makes the
reconstruction problem less ill-posed. As far as we know, our
algorithm is the first of its kind in diffuse optical tomography to
have all three features at the same time. At this point, it is important to emphasize that unlike the existing linearized solver based
on Born-type approximation, such linearization is avoided in the
proposed method, since it is not assumed that the perturbations
from the known background are small in magnitude. (They are
instead assumed to have small supportwhich does not linearize
the forward model). In contrast, linearized solvers assume small
perturbations from the background. Therefore, the advantages of
the proposed algorithm are that the algorithm is exact, and not
subject to the limitations of linearization methods.
The main technical breakthrough of the algorithm comes from
the recent theory of joint sparse recovery in compressed sensing
[15][20]. More specifically, in most diffuse optical tomography
problems, the optical parameter changes locally against the
background, for example, due to the angiogenesis of cancers or
the brain hemodynamic response, occupying only a small portion
of the whole field of view (FOV). Furthermore, the unknown
support of the absorption change is usually invariant during a set
of distinct illuminations. Therefore, we can formulate the diffuse
optical tomography as a joint sparse recovery problem, where we
estimate a set of unknown signals with a common sparse support
[15][20]. The advantage of this formulation is that as soon
as the common sparse support is estimated, the unknown total
optical flux and the optical parameters can be easily calculated
using a FoldyLax type linear integral equation [21][24].
From a historical perspective, our algorithm is an extension
of the MUSIC algorithm in the wave inverse scattering theory
[25][27] and diffuse optical tomography [21], [28]. Even
though the existing MUSIC approaches can be regarded as noniterative and exact in some sense [29], the maximum sparsity level
that these approaches can recover is limited when the number of
sources or detectors is small; hence, the algorithm cannot take
advantage of an increasing number of sources (or receivers) as
long as the number of its counterpart is fixed. This asymmetric
imaging geometry is quite often encountered in practice, where
massive detector arrays are used to measure the scattered optical flux from a limited number of illuminations. On the other
hand, joint sparse recovery algorithms, such as simultaneous
orthogonal matching pursuit (S-OMP) [16] and -thresholding
[17] provide performance superior to MUSIC when the number
of illuminations is small [21]; however, they become inferior
to MUSIC as the number of available snapshots increases. This
apparent dichotomy between sensor array signal processing and
compressive sensing algorithms for the joint sparse recovery
problem has spurred extensive investigations, resulting in
diverse approaches and theories [18], [19], [30][33].
A recent breakthrough in this area has created a new class
of algorithms that are based on a generalized MUSIC criterion,
which has been developed independently by the authors [18],
[19]. More specifically, when the number of targets is , and
independent illuminations are available, the new algorithm finds

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 5, MAY 2011

targets using a compressive sensing algorithm such as


S-OMP or -thresholding, and the remaining targets are recovered using a generalized MUSIC criterion [18], [19]. This
hybrid approach significantly improves the performance of estimating jointly sparse signals and achieves the sparse recovery
bound using a finite number of illuminations. The resulting
bound suggests that the maximum number of recoverable targets
is up to the average number of sources and detectors, which is
a significant improvement over the existing methods. Furthermore, for extended targets whose sparsity level is not well-defined, the generalized MUSIC criterion still provides nonparametric super-resolution imaging spectrum that can be used to estimate the target shapes. Hence, the algorithm is very practical.
The rest of the paper is organized as follows. Mathematical
preliminaries are summarized in Section II and problem formulation is given in Section III. Conventional MMV algorithms
and their limitations are described in Section IV. Section V describes the generalized MUSIC criterion, and imaging geometry
optimization is discussed in Section VI. Numerical simulation
is provided in VII, and the RIP condition and piecewise constant or linear approximation of optical parameter distribution
are discussed in Section VIII, which is followed by the conclusion in Section IX.
II. MATHEMATICAL PRELIMINARIES
and
correspond to the th row
Throughout the paper,
and the th column of matrix , respectively. When is an
index set,
corresponds to a submatrix collecting corresponding rows of and columns of , respectively; and
denotes the index set for nonzero rows. The following definitions are often used throughout the paper.
Definition 1: [34] The rows in
(or columns) are in general
position if any collection of them are linearly independent.
Definition 2 (Mutual Coherence): [35] For a sensing matrix
, the mutual coherence
is given
by

where the superscript denotes the Hermitian transpose.


Definition 3 [Restricted Isometry Property (RIP)]: [36] A
sensing matrix
is said to have a -restricted isometry property (RIP) if there are left and right RIP constants
such that

for all

such that
. A single RIP constant
is often called the RIP constant.
III. PROBLEM FORMULATION

In this section, we provide imaging physics for the continuous


wave diffuse optical imaging geometry since it admits a convenient detection scheme using massive detector arrays such as a
CCD (charge coupled device) camera. Extension of the theory for
frequency-domain or time-domain DOT is also straightforward.

LEE et al.: COMPRESSIVE DIFFUSE OPTICAL TOMOGRAPHY: NONITERATIVE EXACT RECONSTRUCTION USING JOINT SPARSITY

A. Diffusion Equation Formulation for the Forward Problem


In a highly scattering medium with low absorption, the
photon flux density from a continuous wave source
for a
certain illumination pattern is modeled by the continuous-domain diffusion equation [13]
(2)
and
are norHere,
malized absorption and the diffusion coefficients, where
is the absorption coefficient,
is the reduced scattering coefficient, and is the speed of light, respectively. Additionally,
the photon flux must satisfy boundary conditions. In general,
the following form of the extrapolated boundary condition is
employed
(3)
,
where the extrapolation length is given by
denotes a coefficient that depends on the relative refractive
denotes the background diffusion constant,
index [13], [39],
and denotes the normal vector of the measurement surfaces.
In this paper, we are interested in imaging the absorption
contrast. Since angiogenesis in cancer or vasodilatation in
neural activation causes a change in mainly absorption properties of hemoglobins, absorption contrast is very important
in optical bio-imaging [40], [41]. Moreover, tissues usually
have scattering coefficient whose spatial variation is slow, and
[42],
which dominates the absorption coefficient,
so the diffusion coefficients can be regarded as approximately
. Accordingly, we can decomconstant such that
,
pose the absorption coefficients into
is the unknown absorption perturbation. Unlike
where
linearization methods, we do not assume that the perturbation
is small relative to the background absorption coefficient .
Then, the intensity of scattered diffuse light at for the th
snapshot can be represented by [13]

(4)
where

[13], the diffuse wave number


; the homogeneous Greens function
is

where the illumination dependent induced current


given by

1131

is
(8)

is the unknown total flux that depends on the


where
illumination pattern, and denotes the FOV.
In this paper, optical inhomogeneities are assumed to be
sparsely distributed, i.e.,
(9)
denotes a Dirac delta function and the 3-D lowhere
of the -targets are unknown. For simplicity,
cations
are regwe assume that the possible target locations
ularly spaced on a fixed resolution grid size, which implies that
are selected from -possible locations
. Then,
the scattered optical flux from the th illumination is given by

(10)
where
denotes the unknown total incident flux that satisfies the FoldyLax equation [21][24]

(11)
where all the multiple scattering effects are included except the
self-field to avoid singularity [24].
B. Multiple Measurement Vector Formulation
This section shows that the diffuse optical imaging problem
using fixed detector geometry can be converted into a multiple
measurement vector (MMV) problem. The fixed detector geometry is one of the most commonly employed measurement
geometries in diffuse optical tomography for applications such
as breast imaging [43], fluorescence mediated tomography [44],
or brain imaging [45]. We consider multiple snapshot imaging
using distinct illumination patterns. More specifically, using the
sparse source model in (9), the th snapshot of the induced current is given by
(12)

(5)
and the homogenous flux

where
th illumination is given by

is given by
(6)

. Then the measurement for the

(13)

Now, we arrive at the following measurement equation:


By collecting the measurement of (10) at detector positions
, the imaging problem can be represented in a matrix
form
(7)

(14)

1132

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 5, MAY 2011

where the scattered flux


, and the corresponding
are given by
noise matrix
..
.

..

..
.

..
.

..

..
.

Therefore,

can be calculated using least square fitting [21]

(15)

with the sensing matrix of


..
.

..

..
.

(16)

(22)

constructed as follows:

and the induced current matrix

IV. MMV ALGORITHMS


..
.

..
.

..

..
.

(17)

. Now, define
whose rows are only nonzero when
that counts the number of rows
the row-diversity measure
in
with nonzero elements:
. We
further define an active index set of target

), Feng
Under the noiseless measurement scenario (i.e.,
and Bresler [46] and Chen and Huo [15] provided the suffi. Recently, Davies
cient condition for the uniqueness of
and Eldar [47] showed that the sufficient condition is also indeed necessary.
Theorem 1: (Feng and Bresler [46], Chen and Huo [15],
denote the rank of matrix
Davies and Eldar [47]): Let
. Then,
with
has a unique solution if and only if

(18)
(23)
, where
denotes
Note that
within the FOV. Since we are interested in
a photon flux at
imaging targets within an area reachable by photon propagation,
the photon distribution can be safely assumed to be a positive
if and only
value within the entire FOV. Hence,
. Hence, the row-diversity
is equal to the
if
. Then, the
number of nonzero targets, i.e.,
diffuse optical tomography problem can be stated as follows:
(19)

C. Exact Inverse Scattering


Assuming that we obtain the estimate of active set , the coris given by
responding least-squares solution
(20)
where the superscript denotes the pseudo-inverse. However,
our imaging goal is to find the absorption coefficient distribution
rather than the induced current distribution
. Recall that for a given estimate
, the
estimate of the total flux
is a readily available quantity due to the FoldyLax equation (11). Hence, we have the
following vector equation:

Theorem 1 provides insight into the ultimate number of targets


formulation of DOT can recover. Since we have
that a
and
, the maximum number
of targets that the
formulation can recover is given by
(24)
which corresponds to the average of the numbers of the source
and detector elements. Next, we discuss whether conventional
algorithms can achieve the bound.
A. Simultaneous Orthogonal Matching Pursuit (S-OMP) [15],
[16]
The S-OMP algorithm is a greedy algorithm that is performed
by the following procedure:
and
;
at the first iteration, set
iterations,
and
after
, where
is the orthogonal projection onto
;
such that
and
select
set
.
Worst-case analysis of S-OMP [17] shows that a sufficient condition for the success of S-OMP is
(25)

..
.

..
.
(21)

which is exactly the same as Tropps exact recovery conditions


for the single measurement vector problem [48], suggesting that

LEE et al.: COMPRESSIVE DIFFUSE OPTICAL TOMOGRAPHY: NONITERATIVE EXACT RECONSTRUCTION USING JOINT SPARSITY

the maximum sparsity level may not be improved with an increasing number of snapshots, even in a noiseless case. Numerical simulation also shows the performance saturation of S-OMP
beyond a certain number of snapshots [17].
B. MUSIC Algorithm [46], [49]
The MUSIC algorithm [49] was originally developed to estimate continuous parameters such as the bearing angle or DOA.
However, as shown by Feng and Bresler [46], [50], the MUSIC
criterion can none the less be modified to identify the support set
from a finite index set as follows. Refering to
with
,
let the columns of
be an orthonormal basis
for the orthogonal complement of the range space of , that
. The subspace
is often called noise
is,
subspace.
in
Theorem 2: [46], [49] (MUSIC Criterion) Let
and assume that
, and
satisfy
,
, where
and
. Assume further
and
are in general
that the columns of a sensing matrix
position, that is, any collection of columns of are linearly
if and only if
independent. Then, for any
(26)
The quantity
is known as the MUSIC spectrum
and when plotted against its peaks reveal the support set .
if
Note that the MUSIC criterion (26) holds for all
the columns of are in general position. Using the compressive
sensing terminology, this implies that the recoverable sparsity
level by MUSIC for the noiseless measurement case is given by
(27)
where the last equality comes from the definition of the spark
[35]. Therefore, the bound (23) can be achieved by MUSIC
. However, for any
, the
when
MUSIC condition (26) does not hold, and even for the noiseless case perfect recovery cannot be guaranteed. This is a major
drawback of MUSIC compared to compressive sensing algorithms.
C. Two-Thresholding [17]
In two-thresholding, we select a set

with

such that
(28)

Gribonval et al. [17] demonstrated that the performance of twothresholding is often not as good as that of S-OMP, which suggests that two-thresholding is not likely to achieve the -bound
(23) with a finite number of snapshots, even if the measurements
are noiseless.
Another variation [46], [50] of the thresholding method is
to find the correlation with the signal or noise subspace of
rather than data itself. This is closely related to the MUSIC
algorithm. More specifically, we select a set with
such
that it is least correlated with noise subspace
(29)

1133

Note that this is basically the same as thresholding the MUSIC


spectrum. Recently, Fannjiang [29] derived an explicit threshold
in terms of RIP constants such that MUSIC can find the correct support of input signals in the presence of noise. He also
provided a comparison between basis pursuit and MUSIC, and
showed that MUSIC with an optimal threshold is better than CS
methods when the measurement matrix has a full rank. However, he did not provide a method to overcome the limitations
of MUSIC when the number of snapshots is not sufficient or the
signal matrix does not have full rank.
V. GENERALIZED MUSIC CRITERION FOR DOT PROBLEM
A. Theory
As discussed before, MUSIC can achieve the optimality
has full rank, which requires that
when the signal matrix
the number of snapshots be sufficiently large. However, in
some DOT problems, only a limited number of illuminations is allowed for many practical reasons. For example, in
neuro-imaging applications, due to the time-varying hemodynamics, we cannot increase the number of illuminations
sufficiently without sacrificing temporal resolution. Hence,
MUSIC can not address the requirements of DOT. On the
other hand, compressive sensing approaches can estimate
a sparse support with high probability from few snapshots.
Moreover, average case analysis [51] shows that the probability
of exact recovery of sparse support increases with increasing
number of snapshots. However, worst-case analysis for MMV
algorithms such as two-thresholding and S-OMP shows that
even if the measurements are noiseless, the bounds for exact
support recovery are not improved with an increasing number
of snapshots, nor are the bounds achieved [17], [32], [51]. As
a consequence, for the intermediate range of number of illumi), neither MUSIC nor the compressive
nations (i.e.,
sensing approaches are satisfactory. To address this, the authors
have independently proposed a new class of algorithms called
Compressive MUSIC [18] or SA-MUSIC [19]. The main idea
of these algorithms is to generalize the MUSIC criterion to the
rank-deficient case.
Theorem 3 (Generalized MUSIC Criterion): [18], [19] Con, where
,
sider the MMV problem
,
and
. Assume that
. Suppose that satisfies the RIP condition
and the nonzero rows of
are in
with
with
general position. Suppose, furthermore,
. Then, for any
(30)
.
if and only if
Note that, unlike the classical MUSIC criterion, we have a
in Theorem 3. This
particular RIP condition
condition has several interesting relationships with (23). First, it
if and only if
is easy to see that
. But, by Theorem 1, because
is assumed, the latter inequality is a necessary and sufficient condito have a unique solution with
.
tion for

1134

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 5, MAY 2011

Hence, the condition


in Theorem 3 is equivalent to the requirement that the -sparse solution of
be unique. In other words, this RIP condition is the minimal requirement that is required for the success of any recovery algorithm, not just the generalized MUSIC criterion.
is based
Note that the generalized MUSIC criterion for
on the rank of a matrix, which is prone to error when measurement is corrupted by additive noise. Hence, rather than (30),
the following equivalent criterion is more practical.
Corollary 1: [18], [19] Assume that
,
,
,
, and
are the same as in
Theorem 3. Let
,
and
is
the orthogonal projection on
. Then
(31)
.
if and only if
Essentially, Corollary 1 implies that the modified MUSIC algorithm is guaranteed to succeed, if the conditions of Theorem
elements used as initialization are correct.
3 hold, and the
indices of with CS-based alTherefore, we determine
gorithms such as S-OMP or two-thresholding, for which a probabilistic performance guarantee can be given for achieving exact
reconstruction. After that process, we recover remaining indices of with the generalized MUSIC criterion given in Corollary 1, and this reconstruction process is deterministic in the
noiseless case. This hybridization makes the new MUSIC applicable for all ranges of , outperforming all existing methods.
So far, we have assumed that the measurement is noisefree. For the case of a noisy measurement , the corresponding
noise subspace estimate is not exact. However, the following
components of the support used
theorem shows that if the
to initialize the generalized MUSIC criterion are correct, that is,
, then the generalized MUSIC estimate is consisthe
tent and achieves the correct estimation of the remaining -indices for a sufficiently large SNR.
Theorem 4: [18] Consider the noisy MMV problem

where
,
, and
noise. Suppose, furthermore,
a finite threshold

is an additive
. Then, there exist
such that if
(32)

then for any

and

, we have

Furthermore,
is an increasing function of
, and
.
An explicit expression for can be given [18], but is not essential to this paper. From the above discussion, we observe that
the recovery performance of the generalized MUSIC algorithm
is closely related to: 1) the condition number of the measure, 2) the left RIP constant
, and 3) the
ment matrix
geometric properties of the two subspaces
and

determined by the angles between the two subspaces. These observations will be used in optimizing the sensing geometry of
diffuse optical tomography.
B. Generalized MUSIC Spectrum for Extended Targets
,
, and Condition (31) is the
Note that when
same as the MUSIC criterion (26). Since MUSIC was originally
developed for continuous parameter estimation, the similarity
of the generalized MUSIC criterion to the original MUSIC due
to Corollary 1 suggests a continuous form of the generalized
MUSIC algorithm. This is very important for practical DOT of
extended targets where the sparsity level is not well-defined.
A related consideration is the choice of discretization grid interval, which is subject to conflicting requirements. On the one
hand, if we choose a large grid interval, the initial estimate of the
support is prone to discretization error. On the other hand,
reducing the discretization interval has two deleterious effects.
First, it increases the sparsity level for a spatially distributed
target of a given size without increasing , thus increasing the
of error-prone steps of determining the partial supnumber
port. Second, as will be discussed in Section VIII, reducing the
grid interval typically increases the mutual coherence between
the atoms of the resulting sensing matrix . The RIP constant
of usually increases as well, making recovery more difficult.
In [52], it has been shown that the best achievable resolution for
DOT is around 1 2 mm. Therefore, there is little if any advantage in choosing a much smaller grid interval, and we chose
0.5 mm as the grid interval for finding the partial support. After
the partial support is found, the remaining support can be found
using a continuous form generalized MUSIC criterion that does
not depends on a particular grid interval.
In the DOT problem, the columns of the sensing matrix
are not usually normalized. Therefore, we define the normalized
version of the generalized MUSIC spectrum as
(35)
. This normalization plays an
where
especially important role in the DOT problem, because the
value of the Greens function, which determines the norm of
the columns of , is exponentially decaying with increasing
distance between a detector and a target. Hence, the normalization imposes uniform weighting for the spectrum regardless of
the depth of inhomogenities [28]. This enables the application
of a constant threshold to the generalized MUSIC spectrum
to estimate the support of the inhomogeneities. The detailed
algorithm to calculate the generalized MUSIC spectrum is
given in Table I.
VI. IMPLEMENTATION ISSUES
A. Source-Detector Geometry Optimization
Imaging geometry optimization has been an important issue
in DOT [53][56]. For example, the singular value spectrum
method has been used for both source and detector optimization [55]. For a given fixed detector geometry, optimized illumination patterns can be calculated using the CramrRao bound

LEE et al.: COMPRESSIVE DIFFUSE OPTICAL TOMOGRAPHY: NONITERATIVE EXACT RECONSTRUCTION USING JOINT SPARSITY

TABLE I
GENERALIZED MUSIC SPECTRUM

1135

supports is challenging in the


However, even finding the
is close to 1. This is because the
DOT problem, since
forward mapping of the DOT problem is a compact mapping
[57] that has many clustered singular values around zero; hence,
the column vectors (so-called atoms) of the sensing matrix are
highly correlated. Since the performance of S-OMP or the
-thresholding algorithm depends on the RIP of the dictionary
[17], [35], [51], high correlation between atoms significantly
reduces the recovery performance. To overcome this difficulty,
we used dictionary preconditioning [58]. More specifically,
applying a preconditioning weighting matrix
produces the following sensing equation:
(36)

[56]. However, this approach depends on a particular target configurations.


This paper provides a novel solution where the source and
detector geometry optimization problem becomes a purely algebraic one and can be addressed very effectively using the
theory we have developed so far. Note that in our recovery formulation, the upper bound expression (23) decouples the role of
source and detector. More specifically, in order to maximize the
and
should
number of recoverable targets,
is a geometric property of the
be maximized. Since
sensing matrix , it can be maximized by optimizing the detector geometry without a priori knowledge of the targets. This
optimization can be performed completely independently not
only from the source configuration but also from the unknown
targets. This is one of the main differences with the CramrRao
bound approach [56]. After the detector geometry is optimized
and fixed, the source illumination pattern can be optimized by
. At this stage, the optimization problem
maximizing
becomes target-dependent.
This geometry optimization procedure can be further elaborated as follows, even for noisy measurement. More specifically, the required SNR bound for the perfect recovery of the
should be
remaining targets is given in (32). Hence,
reduced and the condition number for noiseless measurements
should be improved. Since

where
denotes the minimum singular values of subma, the left RIP constant can be minimized by increasing the
trix
minimum singular value, which can be optimized by detector
can be done by
geometry. Then, the minimization of
varies depending
optimizing the source geometry since
on the source illumination. The aforementioned geometry optimization problem can be performed sequentially.
B. Dictionary Preconditioning
Recall that the generalized MUSIC criterion guarantees
support components
to find the remaining targets if
are correctly found using conventional compressed sensing
components is easier than
algorithms. Note that finding
finding the components as theoretically demonstrated in [18].

It is well known that the RIP constant can be affected by precan be


conditioning. This indicates that the new dictionary
made to have a better RIP constant. However, optimizing to
have the best RIP is a combinatorial optimization problem. Instead, it can be shown that this optimization problem can be approximated as the minimization of Frobenius norm resulting in
the optimal preconditioner [58]

Consequently, the correlation of the signal after the optimal preconditioning becomes
(37)
which corresponds to the pseudo-inverse. Considering that the
pseudo-inverse reconstruction is one of the popular approaches
for the DOT inverse problem, the correlation can be easily combined with the conventional fast algorithms.
However, due to the ill-posedness of matrix , the optimal
preconditioner of (36) may severely amplify the noise. To mitigate this, we regularize the preconditioner as follows:

for a small regularization parameter


, which determines
the smallest singular value. We chose to make the resulting
dictionary
have a fixed condition
number for all simulation we have conducted. Even with this
regularization, for a sufficiently small , the noise components
of can be significantly amplified by the preconditioner; and
comwith such noise amplification, it is more likely that
ponents of the support will be correctly determined than all
components. Hence, the generalized MUSIC algorithm is also
advantageous when incorporated with the preconditioner.
VII. NUMERICAL EXPERIMENTS
A. 2-D Simulation With Point Targets
In order to demonstrate the performance of the proposed
method, we reconstructed regularly placed targets with a
various number of illuminations, and then calculated the perfect reconstruction ratio under a noisy scenario. The perfect
recovery ratio is defined as the percentage of correct identification of all supports. First, the geometry of the 2-D simulation
is illustrated in Fig. 1(a). The FOV of the simulation is set

1136

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 5, MAY 2011

Fig. 1. Imaging geometry for 2-D simulation. (a) FOV of the simulation and
(b) source configuration.

Fig. 2. Perfect reconstruction fractions in 2-D simulation for regularly spaced


, 3, and 5 when SNR 40
targets with different number of illuminations r
dB, and the target spacing  is (a) 7 mm, (b) 5 mm.

=1

Fig. 3. Perfect reconstruction fractions in 2-D simulation for regularly spaced


and the target spacing  is
targets with SNR of 30, 40, and 50 dB when r
(a) 7 mm, (b) 5 mm.

=3

to 5 cm 5 cm. A total of 17 17 detectors with pitch of 5 mm


are located at the detector plane as shown in Fig. 1(a). Targets
are regularly placed at the center of the FOV at intervals of
and 5 mm as illustrated in Fig. 1(a). We sequentially increased the number of targets from the center of the FOV while
maintaining the same pitch of . The point source configuration
for each number of illuminations is described in Fig. 1(b).
The optical parameters for the homogeneous background are
mm ,
mm , whereas the absorption
change is
mm . We added Gaussian noise with
various SNRs to the scattered flux to verify the noise robustness
of our algorithm. The results are obtained by averaging 300
runs with independent noise realizations for each simulated
scenario.
We performed simulations for regularly spaced targets with
various number of illuminations, SNRs, and target pitches
. In the proposed algorithm,
support components are
found using S-OMP with preconditioning, and the remaining

Fig. 4. Perfect reconstruction fractions in 2-D simulation for detector optimization using the proposed algorithm with r
, 3, and 5 illuminations, SNR
40 dB, and target separation  of 7 mm. The source configuration is fixed to
that illustrated in Fig. 1.

=1

Fig. 5. Perfect reconstruction fractions in 2-D simulation for source optimization using the proposed algorithm with r
and five illuminations, SNR 40 dB,
and target interval  of (a) 7 mm, (b) 5 mm. Detector pitch is fixed to 5 mm.

=3

components are found using the generalized MUSIC criterion.


In Fig. 2(a) and (b), when noise is present, the performance of
the S-OMP algorithm saturates even with increasing number of
illumination patterns . However, as shown in Fig. 2(a) and (b),
increases.
our algorithm shows gradual improvement as
The noise sensitivity of the proposed method is evaluated in
and 5 mm with
illuminations.
Fig. 3(a) and (b) for
The proposed algorithm is seen to be much more robust to
noise than S-OMP. In Figs. 2 and 3, we also observe that the
performance of our algorithm degrades gracefully as target
separation decreases.
B. 2-D Simulation With Source-Detector Geometry
Optimization
Using the 2-D imaging geometry described above, this section demonstrates by simulation the importance of source-detector geometry optimization. For the detector optimization, we
fixed the number of detectors as 17 17 and the source configuration as illustrated in Fig. 1, with different detector pitches.
In Fig. 4, regardless of the detector pitch, the enhancement of
the performance by using MMVs with multiple snapshots are
obvious. However, the detector pitch of 5 mm shows superior
performance to the pitch of 3 mm. For the same number of detector elements, smaller pitch yields a larger left RIP constant,
compared to larger detector pitch. Therefore, a better detector
configuration is the one with a larger detector pitch.
Fig. 5(a) and (b) illustrates perfect reconstruction fractions
for different source configurations with detector pitch of 5 mm.
The source configuration corresponding to each curve in the plot

LEE et al.: COMPRESSIVE DIFFUSE OPTICAL TOMOGRAPHY: NONITERATIVE EXACT RECONSTRUCTION USING JOINT SPARSITY

1137

Fig. 7. Perfect reconstruction fractions in 3-D simulation for regularly spaced


, 10, and 17 illuminations, SNR 40 dB, and target interval 
targets with r
of (a) 7 mm, (b) 5 mm.

=3

Fig. 6. Imaging geometry for 3-D simulation. (a) FOV of the simulation, (b) detector and source configuration, and (c) the targets.

is illustrated by an inset. We can easily confirm that the performance of the configuration with uniform distance sources is inferior to that with nonuniform source distribution for both cases
mm and 5 mm. Furthermore, we observe nonmonoof
tonic behavior of perfect reconstruction fractions in a nonoptimized source geometry. This is because a configuration with
regularly placed sources turns out to generate symmetric incident fields, which degrade the conditioning of the measurement
matrix. This symmetry may produce some unobservable shadow
depending on neighbor configurations, which may result in the
nonmonotonic behavior of the perfect reconstruction fraction.
Therefore, we implemented the source optimization by locating
the sources at positions as irregular as possible.
C. 3-D Simulation With Point Targets
The geometry of the 3-D simulation is illustrated in Fig. 6(a).
The FOV of the simulation is set to cm
cm
cm.
A total of 3 3 detectors with a pitch of 10 mm are located
along six surfaces of a cube as shown in Fig. 6(b). A line
source is used in this simulation. We increased the number
of sources sequentially. Targets are regularly placed at the
and 5 mm as illuscenter the FOV with a distance of
trated in Fig. 6(c). We sequentially increased the number of
targets from the center while maintaining their distance of .
The optical parameters for the homogeneous background are
mm ,
mm , whereas the absorption
mm . We added Gaussian noise of
change is
various SNRs to verify the noise robustness of our algorithm.
The results were obtained by averaging 300 runs with independent noise realizations.
Fig. 7(a) and (b) illustrates that even though the number of
illuminations increases, S-OMP can recover only one target reliably. The inferior performance is especially noticeable under
noise. However, similar to 2-D cases, even when S-OMP fails to
recover all the targets, the overall reconstruction performance of
the proposed method was significantly better. Note that for both
mm and
mm, the proposed algorithm has signifmm, the number of
icant advantages over S-OMP. For
resolved targets is reduced. This is because the restricted isometry constant increases (degrades) and the compressive sensing
support components is not efficient.
stage of reconstructing
However, the MUSIC step still works well and we have a significant gain over S-OMP. Fig. 8(a) and (b) demonstrates that

Fig. 8. Perfect reconstruction fractions in 3-D simulation for regularly spaced


illuminations and target
targets with SNR of 30, 40, and 50 dB, and r
intervals  of (a) 7 mm, (b) 5 mm.

= 10

in 3-D too our algorithm is quite robust for noise whereas the
performance of the S-OMP algorithm for various noise levels is
inferior. These results confirm that the proposed algorithm also
works well in the 3-D geometry.
D. Extended Target Simulation for Molecular Imaging
Applications
To demonstrate the performance of the proposed method in
more realistic problems, we performed a simulation study for a
molecular imaging applications using the Digimouse computational mouse phantom [59]. The goal of the study is to localize
tumors within the mouse. Unlike conventional studies that use
fluorescent molecular probes to target the cancer volume, this
study is to image the tumors based on the intrinsic absorption
changes induced by angeogenesis. We assume that the background optical parameters are available by segmenting the organs using an independent X-ray CT scan or a standard template, whereas the location and the absorption parameters of tumors are unknown.
The source and detector geometry is illustrated in Fig. 9(a).
We use 12 point sources and 738 detectors placed on the surface of a cylinder, which is filled with matching fluid having the
same optical parameter as the mouse skin. The radius and the
height of the cylinder are 16 mm and 32 mm, respectively. The
mouse phantom is illustrated in Fig. 9(b) in which four tumors
are indicated with red arrows. The various optical parameters
of the mouse organs are summarized in Table II for the incident
wavelength of 800 nm, whereas the absorption change for the tumm . For accurate simulation, we solved
mors is
the diffusion equation using the GPU accelerated MonteCarlo
was added to
simulation [60]. Gaussian noise of
the scattered flux.

1138

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 5, MAY 2011

Fig. 9. Imaging geometry for mouse phantom simulation. (a) Source and detector geometry and (b) mouse phantom (the red arrows indicate tumors).

TABLE II
OPTICAL PROPERTIES FOR THE MOUSE [61], [62]

The sparsity level cannot be set a priori, since it depends


on the grid size. Hence, as described in Table I, we attempted to
reconstruct inhomogenities using the generalized MUSIC spectrum. We chose a reconstruction grid size of 0.5 mm, and chose
516 support components using -thresholding and assumed they
support with the grid size of 0.5 mm.
corresponded to the
Then, the generalized MUSIC spectrum in (33) was calculated.
Fig. 10(a) and (b) illustrates the location of the thresholded
spectra of the conventional MUSIC and the generalized MUSIC
criterion, respectively, embedded in the original phantom. The
rendered results show that reconstructed tumors in the liver and
the nearby kidney are connected together into a big tumor for
the case of the conventional MUSIC algorithm, whereas our algorithm clearly shows each tumor in the liver and kidney, separately. Fig. 10(c) and (d) shows linearized reconstruction results based on the Born approximation using Tikhonov regularization and -penalty regularization, respectively. Even though
the -penalty provides better results than Tikhonov regularization due to the sparsity imposing penalty, the performance
is still worse compared to the proposed method. For example,
the tumor in the left lung area was not detected in Fig. 10(d).
Table III summarizes the mean square error (MSE) values of the
reconstructed absorption coefficients and Table IV summarizes
the Hausdorff distance [63] between true supports and estimated
supports for each tumor. Both of these quantitative measures
clearly show the advantages of the proposed method.
In order to test the robustness of the algorithm when the background estimation is not accurate, we performed the simulation
study with Digimouse in Fig. 11(a) and (b). Fig. 11(a) shows the
reconstruction results using the proposed method when the absorption parameter estimation for the left kidney is 1% deviated
from the true value, whereas Fig. 11(b) shows the results from
10% deviation from the truth. As the estimation of the background Greens function became less accurate, the resolution of
the reconstruction became worse. However, the degradation is
graceful, which confirms the practicality of the algorithm.

Fig. 10. 3-D reconstruction of tumors embedded in the mouse phantom with
(a) conventional MUSIC, (b) generalized MUSIC criterion, and linearized reconstruction approaches using (c) Tikhonov regularization, and (d) l1-penalty regularization. The scattered flux is corrupted by additive Gaussian noise of SNR
40 dB.

VIII. DISCUSSION
A. RIP Condition
Recall that the original DOT problem is governed by a continuous integral equation. Due to the diffusive nature of photon
propagation, the inverse problem is severely ill-posed. One may
therefore wonder whether the RIP condition for the generalized
MUSIC criterion is satisfied. In general, the RIP constant is difficult to calculate since it requires combinatorial optimization.
However, in this section we consider a special DOT geometry
for which we can explicitly calculate the mutual coherence of
the sensing matrix. It is known that the RIP constant and the
mutual coherence are related as
, so calculating
mutual coherence can bound the RIP constant, which appears in
our theoretical results. As shown in the following theorem, for
a special detector geometry the mutual coherence is a function
of the grid interval and the optical parameters, which indicates
that the reconstruction grid interval can be controlled to satisfy
the RIP condition for the generalized MUSIC criterion. The geometry considered is a popular planar imaging geometry with a
massive detector array such as CCD.
Theorem 5: Consider a planar imaging geometry, and assume
all inhomogeneities are confined within the region
.
In this case, the mutual coherence is given by
(38)
where

denotes the grid interval, and

LEE et al.: COMPRESSIVE DIFFUSE OPTICAL TOMOGRAPHY: NONITERATIVE EXACT RECONSTRUCTION USING JOINT SPARSITY

1139

TABLE III
MSE FOR THE VARIOUS RECONSTRUCTION METHODS

TABLE IV
HAUSDORFF DISTANCE FOR THE VARIOUS RECONSTRUCTION METHODS

can be absorbed into the


The role of basis function
sensing matrix such that the resulting Greens function becomes
the integration of the original Greens function and the basis
(41)
The effect of the basis then is to change the RIP constant of
the sensing matrix, and hence the overall reconstruction performance. In future research, we will investigate the optimal representation of the optical parameter variation that strikes a balance
between the discretization error and RIP condition.
Fig. 11. 3-D reconstruction of tumors embedded in the mouse phantom when
the absorption parameter estimation of the left kidney are (a) 1% and (b) 10%
deviated from the true value.

Proof: See Appendix A.


Theorem 5 clearly shows the dependence on the grid interval.
Importantly, we observe that
is monotone decreasing, it
follows that by choosing a large grid interval, we can improve
the mutual coherence and the RIP constant, accordingly. A
caveat is that increasing the grid interval increases the discretization error in the measurement model. The trade-off
between the RIP and the discretization error merits systematic
study, since this is a common fundamental problem for all inverse problems that originate from an integral equation forward
mapping. The issue is, however, beyond the scope of the paper.
B. Piecewise Constant/Linear Approximation of Optical
Parameter Distribution
In this paper, we assume that the absorption parameter distribution is described by the sum of weighted and shifted delta
functions. However, a more practical description is piecewise
constant or spline approximation of the absorption parameter
distribution as
(39)
where
denotes the basis function. Then, the theory we
have developed in this paper is still applicable, since the induced
current can be approximated by

(40)

IX. CONCLUSION
This paper described a novel noniterative exact reconstruction
formula for diffuse optical tomography using a fixed detector
geometry with multiple source configurations. By exploiting
the sparsity of the optical parameter perturbations under various illuminations with fixed detector geometry, the reconstruction problem can be formulated as a joint sparse approximation problem. The recently proposed generalized MUSIC algorithm can therefore be used for the resulting MMV problem. A
source/detector optimization problem, and the design of a preconditioner were discussed. Furthermore, extended target estimation problems were addressed systematically using the new
algorithm. Extensive numerical simulations showed that the algorithm is significantly superior to previous algorithms and robust against noise, and misestimation of the background absorption coefficient. Owing to its simplicity and effectiveness, we
believe that the proposed algorithm opens a new direction in
DOT research.
APPENDIX A
PROOF OF THEOREM 5
Subject to the assumptions of Theorem 5, the mutual coherence can be readily approximated since the aforementioned real
space formulation can be equivalently converted into a Fourier
space formulation, often called the FourierLaplace formulation
[12]. We will use this formulation to derive the mutual coherence of the sensing matrix corresponding to the forward mapping of DOT problem.
In this geometry, the unperturbed Greens function is given
by [12]
(A1)

1140

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 5, MAY 2011

where
,
, and denotes the spatial angular
denotes the kernel
frequency with respect to , and
function. For the case of a single plane detector located at
and free boundary condition by [13], the analytic function of
Greens function in the spatial domain is

where the third equation is derived by substituting the Greens


with (A1), integrating with respect to , and
function
exploiting the property of a delta function; and the last equality
.
comes from (A4) in Lemma A1 by setting
and
is given by
Next, the correlation between atoms

(A2)
and the corresponding the kernel function is given by [13]
(A3)
. The following lemma is very
where
useful in calculating the mutual coherence.
Lemma A.1: Let
,
, and denote the spatial angular frequency with respect to . Suppose, furthermore,
. Then, we have the following equality:
(A9)

(A4)
Proof: Using (A1), (A2), and (A3), we have the following
result:

where the third equation is again derived by substituting the


and
with (A1), integrating
Greens function
with respect to , and exploiting the property of a delta function;
and the last equality comes from (A4) in Lemma A1. Therefore,
the normalized correlation is given by

(A10)
(A5)
Integrating both sides with respect to from to yields (A4).
This concludes the proof.
Now, to calculate the mutual coherence, we need to calculate
the normalized correlation between two atoms, say
and

where and denote the column indices with the corresponding


and
,
coordinates of the voxels as
respectively. First, we calculate the normalization factor. Suppose that the detector pitch is sufficiently small to satisfy the
Nyquist condition. Then, we have

where

. Without loss of generality, we assume


. Since the numerator is a monotonic decreasing func, we know that the maximum of the correlation
tion with
should occur at
, i.e.,
. Therefore, the mutual
coherence should be calculated by

(A11)
Now, our goal is to specify the condition
maximum normalized correlation. Let
and
and

to provide the
,
. Then, we have

(A6)
(A7)

(A8)

(A12)
Our goal is to demonstrate that (A12) is a monotone decreasing
function with respect to , and monotone increasing with respect
to . For this, we need the following lemma for the function
.
,
has the following properties.
Lemma A.2: For
1)
is monotone increasing.
is concave.
2)

LEE et al.: COMPRESSIVE DIFFUSE OPTICAL TOMOGRAPHY: NONITERATIVE EXACT RECONSTRUCTION USING JOINT SPARSITY

3)

is convex.
Proof: Note that

Hence, for
are given by

REFERENCES

the first and the second derivatives of

(A13)

(A14)
Hence,
is monotone increasing and concave. Finally, since
is concave and
is nonincreasing and convex,
is convex. This concludes the proof.
Using Lemma A2, we have
(A15)
This implies that (A12) is a monotone decreasing function with
respect to . Now, the first derivative of (A12) with respect to
is given by

(A16)
where the last inequality comes from convexity of
and
Jensens inequality. Hence, (A12) is monotone increasing with
respect to .
Therefore, for a grid interval and the FOV of the slab ge, the minimum
and
ometry from
, so the mutual coherence is given
the maximum
by

(A17)
This concludes the proof.

1141

[1] C. H. Contag, S. D. Spilman, P. R. Contag, M. Oshiro, and B. Dennery,


Visualizing gene expression in living mammals using a bioluminescent reporter, Photochem. Photobiol., no. 66, pp. 523531, 1997.
[2] C. L. Hutchinson, T. L. Troy, and E. M. Sevick-Muraca, Fluorescence
lifetime determination in tissues and other random media from measurement of excitation and emission kinetics, Appl. Opt., no. 35, pp.
23252332, 1996.
[3] B. W. Pogue, M. S. Patterson, H. Jiang, and K. D. Paulsen, Initial
assessment of a simple system for frequency domain diffuse optical
tomography, Phys. Med. Biol., vol. 40, pp. 17091729, 1995.
[4] R. Weissleder, Scaling down imaging: Molecular mapping of cancer
in mice, Nature Rev. Cancer, vol. 2, pp. 18, 2002.
[5] R. Weissleder and V. Ntziachristos, Shedding light onto live molecular targets, Nature Med., vol. 9, no. 1, pp. 123128, 2003.
[6] H. Jiang, K. D. Paulsen, U. L. Osterberg, B. W. Pogue, and M. S. Patterson, Optical image reconstruction using frequency-domain data:
Simulation and experiment, J. Opt. Soc. Am. A, vol. 13, no. 2, pp.
253266, Feb. 1996.
[7] Y. Yao, Y. Wang, Y. Pei, W. Zhu, and R. L. Barbour, Frequency domain optical imaging of absorption and scattering distributions by a
Born iterative method, J. Opt. Soc. Am. A, vol. 14, no. 1, pp. 325342,
Jan. 1997.
[8] V. A. Markel, J. A. OSullivan, and J. C. Schotland, Inverse problem
in optical diffusion tomography. IV. Nonlinear inversion formulas, J.
Opt. Soc. Am. A, vol. 20, no. 5, pp. 903912, May 2003.
[9] J. C. Ye, K. J. Webb, C. A. Bouman, and R. P. Millane, Optical
diffusion tomography using iterative coordinate descent optimization
in a Bayesian framework, J. Opt. Soc. Am. A, vol. 16, no. 10, pp.
24002412, Oct. 1999.
[10] J. C. Ye, C. A. Bouman, K. J. Webb, and R. P. Millane, Nonlinear
multigrid algorithms for Bayesian optical diffusion tomography, IEEE
Trans. Image Process., vol. 10, no. 6, pp. 909922, Jun. 2001.
[11] V. Ntziachristos and R. Weissleder, Experimental three-dimensional
fluorescence reconstruction of diffuse media by use of a normalized
Born approximation, Opt. Lett., vol. 26, no. 12, pp. 893895, Jun.
2001.
[12] V. A. Markel and J. C. Schotland, Inverse problem in optical diffusion
tomography. I. Fourier-Laplace inversion formulas, J. Opt. Soc. Am.
A, vol. 18, no. 6, pp. 13361347, Jun. 2001.
[13] V. A. Markel and J. C. Schotland, Inverse problem in optical diffusion
tomography. II. Role of boundary conditions, J. Opt. Soc. Am. A, vol.
19, no. 3, pp. 558566, 2002.
[14] V. A. Markel, V. Mital, and J. C. Schotland, Inverse problem in optical diffusion tomography. III. inversion formulas and singular-value
decomposition, J. Opt. Soc. Am. A, vol. 20, no. 5, pp. 890902, May
2003.
[15] J. Chen and X. Huo, Theoretical results on sparse representations of
multiple measurement vectors, IEEE Trans. Signal Process., vol. 54,
no. 12, pp. 46344643, Dec. 2006.
[16] J. A. Tropp, A. C. Gilbert, and M. J. Strauss, Algorithms for simultaneous sparse approximation. Part I: Greedy pursuit, Signal Process.,
vol. 86, no. 3, pp. 572588, 2006.
[17] R. Gribonval, H. Rauhut, K. Schnass, and P. Vandergheynst, Atoms
of all channels, unite! Average case analysis of multi-channel sparse
recovery using greedy algorithms, J. Fourier Anal. Appl., vol. 14, no.
5, pp. 655687, 2008.
[18] J. M. Kim, O. K. Lee, and J. C. Ye, Compressive MUSIC: A missing
link between compressive sensing and array signal processing, 2010.
[19] K. Lee and Y. Bresler, Subspace-Augmented MUSIC for Joint Sparse
Recovery, .
[20] M. F. Duarte, S. Sarvotham, D. Baron, M. B. Wakin, and R. G. Baraniuk, Distributed compressed sensing of jointly sparse signals, in
Asilomar Conf. Signals, Sys. Comput., 2005, pp. 15371541.
[21] J. C. Ye, S. Y. Lee, and Y. Bresler, Exact reconstruction formula for
diffuse optical tomography using simultaneous sparse representation,
in Proc. 2008 IEEE Int. Symp. Biomed. Imag.: From Nano to Macro,
Paris, France, May 2008, pp. 16211624.
[22] L. Tsang, J. A. Kong, K.-H. Ding, and C. O. Ao, Scattering of Electromagnetic Waves: Numerical Simulations. New York: Wiley, 2000.
[23] M. I. Mishchenko, L. D. Travis, and A. A. Lacis, Multiple Scattering of
Light by Particles: Radiative Transfer and Coherent Backscattering.
Cambridge, U.K.: Cambridge Univ. Press, 2006.
[24] A. C. Fannjiang, Compressive inverse scattering: I. High-frequency
SIMO/MISO and MIMO measurements, Inverse Problems, vol. 26,
pp. 129, 2010.

1142

[25] A. Kirsch, Characterization of the shape of a scattering obstacle using


the spectral data of the far field operator, Inverse Problems, vol. 14,
pp. 14891512, 1998.
[26] M. Cheney, The linear sampling method and the MUSIC algorithm,
Inverse Problems, vol. 17, no. 4, pp. 591596, 2001.
[27] E. A. Marengo and F. K. Gruber, Subspace-based localization and inverse scattering of multiply scattering point targets, EURASIP J. Adv.
Signal Process., vol. 17342, p. 16, 2007.
[28] M. Pfister and B. Scholz, Localization of fluorescence spots with
space-space MUSIC for mammographylike measurement systems, J.
Biomed. Opt., vol. 9, p. 481, 2004.
[29] A. C. Fannjiang, The MUSIC algorithm for sparse objects: A compressed sensing analysis, Inverse Problems, vol. 27, p. 035013, 2011.
[30] D. P. Wipf, Bayesian methods for finding sparse representations,
Ph.D. dissertation, Univ. California, San Diego, 2006.
[31] J. A. Tropp, Algorithms for simultaneous sparse approximation. Part
II: Convex relaxation, Signal Process., vol. 86, no. 3, pp. 589602,
2006.
[32] M. Mishali and Y. C. Eldar, Reduce and boost: Recovering arbitrary
sets of jointly sparse vectors, IEEE Trans. Signal Process., vol. 56, no.
10, pp. 46924702, Oct. 2010.
[33] Y. C. Eldar, P. Kuppinger, and H. Bolcskei, Block-sparse signals: Uncertainty relations and efficient recovery, IEEE Trans. Signal Process.,
vol. 58, pp. 30423054, 2010.
[34] D. L. Donoho, Neighborly polytopes and sparse solution of underdetermined linear equations Dept. Stat., Stanford Univ., Tech. Rep.,
2005.
[35] D. L. Donoho and M. Elad, Optimally sparse representation in general
(non-orthogonal) dictionaries via l minimization, Proc. Nat. Acad.
Sci. USA, vol. 100, no. 5, pp. 21972202, 2003.
[36] E. Candes, J. Romberg, and T. Tao, Robust uncertainty principles:
Exact signal reconstruction from highly incomplete frequency information, IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489509, Feb.
2006.
[37] A. Aldroubi, Oblique projections in atomic spaces, Proc. Amer.
Math. Soc., vol. 124, pp. 20512060, 1996.
[38] M. Unser and A. Aldroubi, A general sampling theory for non-ideal
acquisition devices, IEEE Trans. Signal Process., vol. 42, pp.
29152925, 1994.
[39] D. Contini, F. Martelli, and G. Zaccanti, Photon migration through a
turbid slab described by a model based on diffusion approximation. I.
Theory, Appl. Opt., vol. 36, pp. 45874599, 1997.
[40] V. Ntziachristos and B. Chance, Probing physiology and molecular
function using optical imaging: Applications to breast cancer, Breast
Cancer Res., vol. 3, pp. 4146, 2001.
[41] D. A. Boas, D. H. Brooks, E. L. Miller, C. A. DiMarzio, M. Kilmer,
R. J. Gaudette, and Q. Zhang, Imaging the body with diffuse optical
tomography, IEEE Signal Process. Mag., vol. 18, no. 6, pp. 5775,
Nov. 2001.
[42] M. Pfister and B. Scholz, Localization of fluorescence spots with
space-space music for mammographylike measurement systems, J.
Biomed. Opt., vol. 9, pp. 481487, May/Jun. 2004.
[43] S. D. Konecky, G. Y. Panasyuk, K. Lee, V. Markel, A. G. Yodh, and
J. C. Schotland, Imaging complex structures with diffuse light, Opt.
Express, vol. 16, no. 7, pp. 50485060, 2008.
[44] V. Ntziachristos, C. Tung, C. Bremer, and R. Weissleder, Fluorescence-mediated tomography resolves protease activity in vivo, Nat.
Med., vol. 8, pp. 757760, 2002.

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 5, MAY 2011

[45] M. A. Franceschini and D. A. Boas, Noninvasive measurement of neuronal activity with near-infrared optical imaging, NeuroImage, vol. 21,
no. 1, pp. 372386, 2004.
[46] P. Feng, Universal minimum-rate sampling and spectrum-blind reconstruction for multiband signals, Ph.D. dissertation, Univ. Illinois, Urbana-Champaign, 1997.
[47] M. E. Davies and Y. C. Eldar, Rank awareness in joint sparse recovery, 2010.
[48] J. A. Tropp, Greed is good: Algorithmic results for sparse approximation, IEEE Trans. Inf. Theory, vol. 50, no. 10, pp. 22312242, Oct.
2004.
[49] R. Schmidt, Multiple emitter location and signal parameter estimation, IEEE Trans. Antennas Propag., vol. 34, no. 3, pp. 276280, Mar.
1986.
[50] P. Feng and Y. Bresler, Spectrum-blind minimum-rate sampling and
reconstruction of multiband signals, in Proc. ICASSP, Atlanta, GA,
May 1996, vol. 3, pp. 16881691.
[51] Y. C. Eldar and H. Rauhut, Average case anlysis of multichannel
sparse recovery using convex relaxation, IEEE Trans. Information
Theory, vol. 56, pp. 505519, 2010.
[52] V. Ntziachristos, C. Bremer, and R. Weissleder, Fluorescence imaging
with near-infrared light: New technological advances that enable in
vivo molecular imaging, Eur. Radiol., vol. 13, no. 1, pp. 195208,
Jan. 2003.
[53] A. Joshi, W. Bangerth, and E. M. Sevick-Muraca, Non-contact fluorescence optical tomography with scanning patterned illumination,
Opt. Express, vol. 19, no. 14, pp. 65166534, Jul. 2006.
[54] T. Lasser and V. Ntziachristos, Optimization of 360 projection
fluorescence molecular tomography, Med. Image Anal., vol. 11, pp.
389399, Apr. 2007.
[55] J. P. Culver, V. Ntziachristos, M. J. Holbrooke, and A. G. Yodh, Optimization of optode arrangements for diffuse optical tomography: A
singular-value analysis, Opt. Lett., vol. 26, no. 10, pp. 701703, May
15, 2001.
[56] J. Dutta, S. Ahn, A. A. Joshi, and R. M. Leahy, Illumination pattern optimization for fluorescence tomography: Theory and simulation
studies, Phys. Med. Biol., vol. 55, no. 10, pp. 29612982, Apr. 2010.
[57] M. Guven, B. Yazici, K. Kwon, E. Giladi, and X. Intes, Effect of discretization error and adaptive mesh generation in diffuse optical absorbing imaging: I, Inverse Problems, vol. 23, pp. 11151133, 2007.
[58] K. Schnass and P. Vandergheynst, Dictionary preconditioning for
greedy algorithms, IEEE Trans. Signal Process., vol. 56, no. 5, pp.
19942002, May 2008.
[59] B. Dogdas, D. Stout, A. F. Chatziioannou, and R. M. Leahy, Digimouse: A 3-D whole body mouse atlas from CT and cryosection data,
Phys. Med. Biol., vol. 52, pp. 577587, 2007.
[60] Q. Fang and D. A. Boas, Monte Carlo simulation of photon migration
in 3-D turbid media accelerated by graphics processing units, Opt.
Express, vol. 17, no. 22, pp. 2017820190, Oct. 2009.
[61] G. Alexandrakis, F. R. Rannou, and A. F. Chatziioannou, Tomographic bioluminescence imaging by use of a combined optical-PET
(OPET) system: A computer simulation feasibility study, Phys. Med.
Biol., vol. 50, pp. 42254241, 2005.
[62] S. A. Prahl, Optical properties spectra 2001, Oregon Medical Laser
Clinic [Online]. Available: http://omlc.ogi.edu/spectra/index.html
[63] D. P. Huttenlocher, G. A. Klanderman, and W. J. Rucklidge, Comparing images using the Hausdorff distance, IEEE Trans. Pattern Anal.
Mach. Intell., vol. 15, no. 9, pp. 850863, Sep. 1993.

S-ar putea să vă placă și