2070 J. Opt. Soc. Am. A / Vol. 28, No. 10 / October 2011 Raveendran et al.

Pseudo-time particle filtering for diffuse

optical tomography

Tara Raveendran,1 Saurabh Gupta,1 Ram Mohan Vasu,1,* and Debasish Roy2
1
Department of Instrumentation and Applied Physics, Indian Institute of Science, Bangalore 560 012, India
2
Department of Civil Engineering, Indian Institute of Science, Bangalore 560 012, India
*Corresponding author: [email protected]

Received June 2, 2011; revised August 3, 2011; accepted August 4, 2011;

posted August 4, 2011 (Doc. ID 148570); published September 13, 2011
We recast the reconstruction problem of diffuse optical tomography (DOT) in a pseudo-dynamical framework and
develop a method to recover the optical parameters using particle filters, i.e., stochastic filters based on Monte
Carlo simulations. In particular, we have implemented two such filters, viz., the bootstrap (BS) filter and the
Gaussian-sum (GS) filter and employed them to recover optical absorption coefficient distribution from both
numerically simulated and experimentally generated photon fluence data. Using either indicator functions or
compactly supported continuous kernels to represent the unknown property distribution within the inhomoge-
neous inclusions, we have drastically reduced the number of parameters to be recovered and thus brought the
overall computation time to within reasonable limits. Even though the GS filter outperformed the BS filter in terms
of accuracy of reconstruction, both gave fairly accurate recovery of the height, radius, and location of the inclu-
sions. Since the present filtering algorithms do not use derivatives, we could demonstrate accurate contrast
recovery even in the middle of the object where the usual deterministic algorithms perform poorly owing to
the poor sensitivity of measurement of the parameters. Consistent with the fact that the DOT recovery, being
ill posed, admits multiple solutions, both the filters gave solutions that were verified to be admissible by the
closeness of the data computed through them to the data used in the filtering step (either numerically simulated
or experimentally generated). © 2011 Optical Society of America
OCIS codes: 100.3010, 110.6955, 170.4580, 170.6960.

1. INTRODUCTION use of measurement noise characteristics to choose the reg-

Noninvasive imaging and spectroscopy based on near- ularization operator, the L curve [8], and the generalized cross
infrared (NIR) radiation have been extensively researched validation [9], are commonly employed. They use the noise
in the past because of its usefulness in the diagnosis of cancer behavior in measurement to minimize the data-model misfit
in soft-tissue organs [1–3]. This stems from the fact that op- functional. But the data-model misfit is not only due to noise
tical absorption spectrum in the NIR region can be used to in measurements but also to the inadequacy of the photon
build quantitative images of hemoglobin concentration in its transport model to represent the true propagation of photons.
oxygenated and deoxygenated states, useful for extracting the Whereas regularization as applied in quasi-Newton iterations
functional signatures of cancer [3,4]. Light propagation in is unable to tackle the mismatch effectively, a stochastic
highly scattering media such as human tissue is modeled filtering approach can through an appropriate selection of
either by the radiative transfer equation (RTE) or its approx- the system process noise. Moreover, the Hilbert space setting
imation, the diffusion equation (DE) [5]. The optical property of deterministic reconstruction algorithms is unsuitable for a
distributions of the tissue, the absorption, and the reduced proper treatment of noise terms with bounded variations. In
scattering coefficients, which form part of the model of this work, we propose to recast the DOT problem within a
photon transport, are recovered from noisy measurements of pseudo-dynamical framework that enables the parameter re-
photon flux on the boundary of the object [5]. The associated covery through a particle filter, i.e., a stochastic filter based
inverse problem is usually solved by minimizing a data-model on a Monte Carlo (MC) simulation of the process equations.
misfit functional with the help of a variant of Newton iteration In particular, we presently employ a couple of variations of
[5]. A numerical treatment, as above, of the optical imaging the particle filter, e.g., the sampling importance resampling
inverse problem suffers from weaknesses such as severe (SIR) filter (also known as the bootstrap filter) and Gaussian-
ill-posedness (which necessitates regularization), slow con- sum filters to solve the inverse problem of DOT with statically
vergence, and the need for a strict stopping criterion [5]. collected data.
Tikhonov regularization or its variants are often employed A stochastic filtering approach has been made use of in the
to reduce ill-posedness, which, on the selection of the optimal past to solve the DOT problem, both in the case where the
regularization parameter, is known to yield improvement in optical parameters are time invariant or otherwise [10–12].
smoothness and quantitative accuracy of reconstruction [6]. This approach essentially treats these parameter fields (or,
However, arriving at the optimal choice of regularization rather, their spatially discretized vector forms) as additional
parameter in the conventional diffuse optical tomography states evolving as Wiener processes. The aim of such filtering
(DOT) image reconstruction problem is not easy. Selections is to compute the probability distribution of the parameters
based on Morozov’s discrepancy principle [7], which make conditioned on the observations (measurements) up to the

1084-7529/11/102070-12$15.00/0 © 2011 Optical Society of America

Raveendran et al. Vol. 28, No. 10 / October 2011 / J. Opt. Soc. Am. A 2071

current instant, with the estimate typically given by the mean should typically be associated with process equations, rather
of the distribution. While for linear inverse problems, the enters the drift term in the measurement stochastic differen-
Kalman filter (KF) provides the optimal, least-square estimate, tial equations (SDEs). This renders the measurement model
the nonlinear DOT problem has been tackled using the sub- nonlinear in the optical parameters to be recovered. Using
optimal extended KF (EKF) [12,13]. The EKF has also been both numerically simulated and experimental phantom data,
employed to solve the inverse (static) elasticity problem by we have shown the effectiveness of the PF approach in gen-
posing the static equilibrium equations as a pseudo-dynamical eral and the GSPF in particular in recovering the optical
system whose steady-state solution is equivalent to the static absorption coefficient field over the object domain, which
response sought for [14]. One of the advantages gained vis-à- is presently two dimensional. The rest of the paper is orga-
vis the gradient-based approach is a lesser dependence on nized as follows. In Section 2, we describe the forward pro-
(and, sometimes, avoidance of) explicit regularization. How- blem of frequency domain DOT followed by a brief account
ever, the pseudo-dynamical EKF (PD-EKF), which requires of the conventional Gauss–Newton approach and a more
repeated computations of the Frechet derivatives of the non- detailed account of the proposed PFs to solve the inverse pro-
linear system equations, is computationally expensive and blem at hand. Details of simulations and phantom experi-
also needs proper tuning of the process noise covariance for ments along with the reconstruction results are given in
the recursion to lead to acceptable estimates. Therefore, in Section 3. Conclusions of this study are put forth in Section 4.
Banerjee et al. [15], an ensemble KF (EnKF) is used to solve
the pseudo-dynamical equations in the context of a class of 2. METHODS: A THEORETICAL
static inverse problems. Compared to either the regularized FRAMEWORK
Gauss–Newton method or the PD-EKF, the EnKF is shown
A. Forward Problem of DOT
to be numerically more accurate and computationally expedi-
In DOT, the objective is to recover the parameter vector
ent. However, the essential limitations of any KF-based esti-
μ≔fκT μTa gT ∈ V μ ≔V κ ⊕V μa ⊂R2n obtained following the
mate (e.g., the restriction that the filtering distribution must
discretization of the associated real, positive, scalar param-
be Gaussian) remain.
eter fields κðxÞ and μa ðxÞ, which respectively denote the diffu-
Particle filtering represents a class of schemes that
sion and absorption coefficients with respect to the spatial
implement a recursive nonlinear Bayesian update by MC simu-
variable x ∈ Ω⊂Rq (where Ω denotes the object domain). The
lations. The required posterior (filtering) distribution is re-
discretization is usually through the finite element method
placed here by an empirical measure through an ensemble of
with μ representing the nodal vector. The object is illuminated
particles (random samples) with associated weights and is
by a source qðr0 ; tÞ ¼ qðr0 Þejωt , where r0 ∈ Rq is a point one
computed/propagated recursively based on these samples
transport-mean-free path inside the boundary of the object.
and weights. As the ensemble size increases indefinitely, the
The response to the illumination, the photon density distribu-
empirical measure should approach, in some sense, the true
tion φðx; ωÞ, is governed by the partial differential equation [5]
filtering distribution. The sequential importance sampling
(SIS) scheme forms the basis for many different versions of

jω
the particle filter (PF) developed and applied over a range −∇ · ðκðxÞ∇φðx; ωÞÞ þ μa ðxÞ þ φðx; ωÞ ¼ q0 ðr0 Þ;
of problems over the past few decades [16–20]. Most PF stra- c ð1aÞ
tegies involve at least two steps, viz. prediction and updating x ∈ Ω⊂Rq ;
(or filtering). The accuracy and efficacy of a PF scheme is
crucially dependent on the importance sampling function, ty- subject to the boundary condition
pically a Radon–Nikodym derivative of the posterior distribu-
tion with respect to the prediction distribution and is used to φðm; ωÞ þ 2AκðmÞð∂φðm; ωÞ=∂^
nÞ ¼ 0: ð1bÞ
weight the predicted particles. The SIR or bootstrap filter [19],
a variant of the SIS filter, uses the transitional distribution Here A is the Fresnel reflection coefficient and n ^ the unit out-
corresponding to the process equations (that represent the ward normal to ∂Ω at the point m. Let Φ ∈ V Φ ⊂Rd denote the
system dynamics) as the prediction distribution, which is thus discretized nodal vector corresponding to φðx; ωÞ. Assuming,
free of the latest measurement that is accounted for purely without a loss of generality, that the set of points fmi ; i ∈
through the importance sampling function. Here, a third step, ½1; mŠg at which measurements are made are at nodal loca-
called the resampling step, is used to draw (with replacement) tions, and denoting the vector ϕ ∈ V ϕ ⊂Rm to be an appropri-
uniformly weighted particles from a set of nonuniformly ate subset of Φ corresponding to the measurement nodes on
weighted ones. This last step tends to discard the particles ∂Ω, we denote the noisy measurement vector by ϕðmÞ ∈ Rm .
with small importance weights and replicate those with higher We have κðxÞ ¼ ð3ðμa ðxÞ þ μ0s ðxÞÞÞ−1 , μ0s ðxÞ being the reduced
weights and faces the undesirable prospect of filter degener- scattering coefficient.
acy and divergence [16]. The Gaussian-sum PF (GSPF) [21,22], The usual deterministic route to recover μ given ϕðmÞ in-
on the other hand, provides a relatively superior alternative volves minimizing the squared (ℓ2 ) norm of the data-model
wherein the densities (in the prediction and updating steps) misfit. The matrix in the associated normal equation has a
are approximated via finite Gaussian mixtures and sampling- nontrivial null space and hence cannot be inverted without
based methods used for sequential updates. The presently regularization. Variants of Tikhonov regularization, which pe-
adopted reconstruction setting uses a parsimonious represen- nalize the misfit functional, yield the optimization problem [5]
tation of the process states, which include only the discretized
optical parameter vectors to be estimated. The discretized 1 λ
min ‖ϕðmÞ − ϕðμÞ‖2 þ ‖LðμÞ‖2 ; ð2aÞ
forward model (i.e., the DE for photon propagation), which μ 2 2
2072 J. Opt. Soc. Am. A / Vol. 28, No. 10 / October 2011 Raveendran et al.

where LðμÞ is the regularization matrix and ϕðμÞ is obtained Direct integration over time step Δt results in
through
μðt þ ΔtÞ ¼ expð½−Mðμ ފðΔtÞÞμðtÞ
KðμÞΦ ¼ q: ð2bÞ Z tþΔt
þ expð½−Mðμ ފðt þ Δt − sÞÞf ds; ð7Þ
KðμÞ is the system matrix and q the force vector. The normal t
equation for the problem is
where f ¼ Mðμ Þμ − Vðμ Þ is a constant forcing term.
gðμÞ ¼ JðμÞT ðϕðmÞ − ϕðμÞÞ − λL0 ðμÞ ¼ 0; ð3Þ Selection of the linearization point t in ½t; t þ ΔtŠ at t ¼ t
or at t ¼ t þ Δt results in explicit phase space linearization
where JðμÞ ¼ ∂ϕðμÞ 0 ∂LðμÞ
∂μ and L ðμÞ ¼ ∂μ . The Newton method for (PSL) [26] or implicit locally transversal linearization [27], re-
solving Eq. (3) gives the iteration spectively. The explicit PSL map applied to Eq. (7) results in
an iteration that bypasses the computation of the inverse of
Δμiþ1 ¼ ðJT ðμi ÞJðμi Þ þ λL00 ðμi Þ JT J or its modifications. Many schemes to integrate Eq. (6),
described in [14], include (1) the forward Euler (explicit)
þ Fðμi ÞÞ−1 ðJT ðμi ÞðϕðmÞ
discretization leading to the damped Landweber iteration,
− ϕðμi ÞÞ − λL0 ðμi ÞÞ: ð4Þ μiþ1 ¼ μi − hVðμi Þ, where h ¼ tiþ1 − ti and (2) backward
Taylor discretization yielding μiþ1 − hVðμiþ1 Þ ¼ μi . It has
F, the discretized second derivative of the operator in Eq. (1), also been demonstrated that h plays a role similar to the reg-
is difficult to compute and hence is dropped in the approxi- ularization parameter of the LM algorithm [14].
mate scheme, such as the Gauss–Newton (GN) [15] or the
Levenberg–Marquardt (LM) [23–25] method. For instance, the C. Pseudo-Time PF for DOT
LM update equation is given by PFs, which are stochastic filters employing MC simulations,
offer an attractive alternative in view of their grounding
Δμiþ1 ¼ ðJT ðμi ÞJðμi Þ þ τiþ1 IÞ−1 ðJT ðμi ÞðϕðmÞ − ϕðμi ÞÞÞ: ð5Þ in the theory of stochastic processes, which not only enable
a proper stochastic modeling of the noise terms with un-
Here fτiþ1 > 0g is a monotonically decreasing set of regular- bounded variations but also account for possible model uncer-
ization parameters. Typically, one takes τ1 ¼ 10 and reduces tainties (e.g., those owing to the replacement of the RTE by
it by a factor of maxfdiagfJT ðμi ÞJðμi Þgg100:25 with every the DE). Moreover, unlike the EKF or the EnKF, PFs admit
iteration [6]. non-Gaussian features to be retained and reconstructed while
dispensing with the need to form Jacobians, a major source of
computational inexpediency and numerical error that could
B. Pseudo-Time Approach for DOT propagate quickly. To start with, we set about defining our
Choice of the penalization term and the regularization param- process and measurement equations while emphasizing a par-
eters, which also alter the character of the minimization simonious representation of the states to be reconstructed.
problem, greatly influences the closeness of solutions of the More specifically, we include in the process states only the
original and regularized problems. Only with a proper choice parameters to be reconstructed and allow them to evolve
of L [Eq. (2a)] can numerical difficulties arising out of a com- in pseudo-time as Wiener (Brownian motion) processes (more
pactly supported operator be tackled satisfactorily. An alter- broadly, as local martingales [28]). Just like the filtration
native, which also avoids an explicit inversion of the generated by the artificially added process noise, a similarly
linearized operator [as in Eq. (4)], can be had by introducing increasing artificial filtration must be generated for the mea-
an artificial dynamics to the parameters and seeking a solu- surement equation via an artificially added Wiener noise.
tion, which is a steady-state or an appropriately stopped re- In this way, the original measurement vector now becomes
sponse of the artificially dynamical system. Here “time” is a a stochastic process allowing for nontrivial updates of the
pseudo variable, which helps us recursively integrate the process parameters as recursions proceed. Moreover, the pro-
ordinary differential equations (ODEs). The evolution of the pagation of photon flux, as given by the inversion of the dis-
artificial dynamics may be interpreted to be on an invariant cretized DE Eq. (2b), is incorporated as the drift term of the
manifold (a Lie manifold), which could help in the develop- measurement equation so that the process and measurement
equations over t ∈ ðti ; tiþ1 Š are respectively given by
ment of robust numerical schemes. This approach has a par-
allel in the Landweber iteration that corresponds to an explicit
dμa ðtÞ ¼ dξðtÞ; ð8aÞ
Euler time discretization of the ODEs. With τ ¼ 0, the pseudo-
time ODE may be formed in ðt; t þ ΔtŠ using the LM map as dzðtÞ ¼ DK−1 ðμa Þqdt þ dϑðtÞ: ð8bÞ

μ_ þ Mðμ ÞðμðtÞ − μ Þ þ Vðμ Þ ¼ 0; ð6Þ Here z ∈ Rm is the set of measured photon flux
(assumed to be strictly nodal) and D ∈ Rm×n is a binary
where Mðμ Þ ¼ ½JT ðμ ÞJðμ ފ, Vðμ Þ ¼ ½JT ðμ ފðϕðμ Þ− coefficient matrix (i.e., with entries 0 or 1). Note that we have
ϕðmÞ Þ, J ∈ Rm×2n , and μ is the linearization point in V μ . We left out κ from the process equations as κðxÞ is often deemed
note that the measurement ϕðmÞ is, in general, complex. How- to be known to a level of confidence. ξðtÞ and ϑðtÞ are zero-
ever, as we have done in our studies, when restricted to the mean (low-intensity) vector Wiener processes with indepen-
special case of constant intensity illumination, i.e., ω ¼ 0, ϕðmÞ dently evolving scalar components. Equations (8a) and (8b)
is positive real. are SDEs expressed in incremental forms owing to the
Raveendran et al. Vol. 28, No. 10 / October 2011 / J. Opt. Soc. Am. A 2073

nondifferentiability of the Wiener noise processes. In this a reduced dimensional parameter vector vis-à-vis the nodal
setting, while the process SDEs are linear in μa , the measure- representation. Given the thrust of this study, which is to bring
ment SDEs [Eq. (8b)] are clearly nonlinear in μa and thus jus- out the essential advantages of a pseudo-time PF-based ap-
tify the use of a PF over the EKF or the EnKF to solve the proach over the regularization quasi-Newton method, we have
estimation problem. Note that the filtration F zt generated presently restricted our numerical explorations based on the
by zðtÞ includes F z0 , which is the σ field corresponding to representation given by Eq. (9a), leaving the other possibilities
the true (static) noise in the actual measurement ϕðmÞ . The to a future study.
noise vector ξðtÞ imparts μa ðtÞ the character of a (local) Unlike the PFs, the GN algorithm requires sensitivity
Wiener martingale (of nonzero mean). Since ϑðtÞ is fictitious, matrices for its implementation, which here is computed es-
its intensity (as measured by the variances of its scalar com- sentially using the recipe given in [5]. All the derivatives for a
ponents) should be small; however, an inadequately small in- source–detector pair are obtained from two Green’s func-
tensity may not enable a nonzero innovation in pseudo-time tions, one for the forward propagation equation (Gφ ) and
and hence a meaningful update of μa . the other for its adjoint (GΓ ). The modification that is required
Depending on the finite element discretization of the field when we use the new representation for the field μa ðxÞ (i.e.,
μa ðxÞ, n could be very large leading to a larger null space of the when a typical inhomogeneity is assumed constant within a
linearized operator (or an elevated multiplicity of possible so- circular region and represented by just four parameters) is
lutions, most of them being unphysical). Moreover, since PFs to compute the differential of the inhomogeneity when the
employ computation-intensive MC simulations, a larger pro- parameters are varied, which, when multiplied by Gφ · GΓ ,
cess state dimension implies substantially additional compu- provides the required derivatives. Further details on the deri-
tational overhead. In order to further reduce n, we propose to vative calculation for this new representation are part of
have a functional representation of the field μa ðxÞ such that another work currently being completed.
the undetermined coefficients of the representation (hence- The process SDEs (of dimension n ¼ 4P) are given
forth represented by the vector μp ) only serve as the process by Eq. (8a) with μp replacing μa . Now the aim of a PF is to
states, i.e., we have μa ðxÞ ¼ f ðμp ; xÞ. Keeping in mind that the obtain the conditional (projected) probability distribution
cross sections of the inhomogeneities are nearly circular, one Πiþ1j1:iþ1 ≔Πðμp;iþ1 jz1:iþ1 ≔fzT1 ; …; zTiþ1 gT Þ ¼ Πðμp;iþ1 jF ztiþ1 Þ
of the simplest representations of f will be through a set of (or the associated density piþ1j1:iþ1 , if it exists) of the states [in
indicator functions of the form this case, μp ðtÞ] conditioned on the measurements z1:iþ1 avail-
P able until the current time instant tiþ1 . The estimate of μp at
X
f ðμp ; xÞ ¼ μ
a þ H k I Ωk ðxÞ; tiþ1 is then given by π tiþ1 ðμp Þ ¼ μ ^ p;iþ1 ¼ E Πp;iþ1j1:iþ1 ðμp;iþ1 Þ,
k¼1 ð9aÞ where E denotes the expectation operator under the measure
H k ≥ 0 and Ωk ⊂Ω for each k: corresponding to piþ1j1:iþ1 . Thus, in continuous time, π t ðμp Þ ¼
^ p ðtÞ ¼ Eðμp ðtÞjzðsÞ; 0 < s ≤ tÞ may be considered as the con-
μ
Here I is the indicator function, ∂Ωk is a circle of radius ditional mean of a projected stochastic process that is F Zt
(spread) r k and centered at xk ¼ ðxk ; yk Þ for each k and
adapted (measurable) [where F Zt is the filtration generated
Ωk ∩Ωj ¼ ϕ for k ≠ j. Moreover, μ a denotes the background
by fzðsÞ; 0 < s ≤ tg]. For the DOT problem given by Eq. (1),
absorption coefficient. This μ a for normal tissue may either
the existence (and not the uniqueness) of μ ^ p;t may be proved
be assumed known or included as an additional scalar param-
via the concept of “dual predictable projection” [28]. Evolu-
eter in μp . Depending upon P, which is taken to be typically
tions of such estimates are governed by the nonlinear optimal
less than 4 in a practical malignancy detection experiment,
filtering SDEs [30]. For instance, we have the following SDE
we arrive at a still-reduced set of the parameter vector
for μ ^ p;t :
μp ≔fH k ; xk ; r k jk ∈ ½1; PŠg to be reconstructed. In case
f ðμp ; xÞ is required to be continuous in x, an alternative t;
dπ t ðμp Þ ¼ ½π t ðμp ðDK−1 ðμp ÞqÞT Þ − π t ðμp Þπ t ðDK−1 ðμp ÞqÞT Šdϑ
representation based on a set fφk ðxÞ; k ∈ ½1; PŠg of com-
pactly supported C ∞ functions (or mollifiers) φk ðxÞ [with ð10Þ
supp φk ðxÞ ¼ Ωk and r k being the radius of ∂Ωk ] is
where dϑ  t ¼ dϑt − π t ðDK−1 ðμp ÞqÞdt denotes the incremental
P
X innovation process. Unfortunately, the above equation in
f ðμp ; xÞ ¼ μ
a þ H k φk ðxÞ: ð9bÞ ^ p;t is not closed as the right-hand side (RHS) contains higher-
μ
k¼1
order conditional moments. In the PF, this closure problem is
The above representations of the inhomogeneity could be addressed through MC simulations of the process equations.
inadequate to model the practical case of a malignant tumor For an elementary review of various PF algorithms, we refer
with a boundary that might not be circular. In such cases, to [31].
a shape-based reconstruction strategy [29] would be pre-
ferred. Thus, when the support of the kth inhomogeneity is 1. Bootstrap Filter
a more general closed curve (in two dimensions) given in The bootstrap (BS) is a very basic form of PF, consisting of
polar coordinates Pby ∂Ωkjnθðr; θÞ with its Fourier series expan- three steps, viz. prediction, updating (weight calculation), and
sion ∂Ωk ðr; θÞ ≅ Qn¼−Q e ∂Ωnk , f∂Ωnk g is the additional set of resampling. Since measurements are available at discrete time
parameters to be included in μp . Again taking P to be typically instants, we can dispense with the time-continuous SDE forms
less than 4 in practical malignancy detection experiments, of the process and measurement equations and adopt their
we arrive at the following set of the parameter vector: following discrete counterparts instead (replacing μp by μ
ðnÞ
μp ≔fH k ; ∂Ωk jk ∈ ½1; PŠ; n ∈ ½0; QŠg. This would still provide for notational convenience):
2074 J. Opt. Soc. Am. A / Vol. 28, No. 10 / October 2011 Raveendran et al.

μiþ1 ¼ μ i þ σμ ζ i ; ð11aÞ Pðdμiþ1 =Z1:iþ1 Þ ≈ P ðNÞ ðdμiþ1 =Z 1:iþ1 Þ

N ð16Þ
Ziþ1 ¼ DKðμÞ−1 q þ ϑiþ1 ; ð11bÞ
X
¼ ð1=NÞ δμðuÞ ðdμiþ1 Þ:
iþ1
u¼1
where tiþ1 ¼ iΔt, σμ ∈ ℝn×n is a diagonal matrix of process
pffiffiffiffiffiffi
noise intensity and ζ i ∼ Nð0; ΔtÞ denotes the standard Wi- Here δμ ðdμÞ denotes the Dirac delta measure located at μ. The
ener increment over ½ti ; tiþ1 Š. Δt, the pseudo-time step, is pre- posterior distribution can then be expressed as
sently assumed to be uniform and n ¼ 1; 2; …mðt ∈ ð0; tm ÞÞ.
Note that the nonlinear measurement Eq. (11b), which is of
a convenient form, is not strictly the discretized counterpart
of the measurement SDE Eq. (8b), as the drift term in Eq. (11b)
has no explicit time dependence. The prediction stage in the
BS filter consists of generating a set of independent and iden-
ðuÞ
tically distributed (iid) particles fμi : u ∈ ½1; NŠg sampled
from the posterior distribution at t ¼ ti , i.e., pi=i ≔pðμi =Z1:i Þ
and the forward projection of these particles, over ðti ; tiþ1 Š
using a transition kernel χðμiþ1 =μi Þ derived from the process
equation. The subscript u refers to the uth particle [i.e., the
uth realization of the random variable μi ðωÞ]. The predicted
particles are then assigned a set of likelihoods or weights
ðuÞ
fwiþ1 g to account for the currently available measurement
Ziþ1 . The weights are in turn used in the subsequent resam-
ðuÞ
pling stage to obtain a new a set of iid particles fμiþ1 : u ∈
½1; NŠg that (following normalization) provides for an
empirical measure to approximate piþ1jiþ1 ≔pðμiþ1 =Z1:iþ1 Þ .
ðuÞ
Here the normalized weights wiþ1 are thus proportional to
the ratios of the true posterior and the prediction densities
(Radon–Nikodym derivatives of the associated changes of
measures).
Presently the prediction density (prior) at time tiþ1 is given
by [19]
Z
qðμiþ1 =Z1:i Þ ¼ pðμ i =Z1:i Þχðμiþ1 =μ i Þdμ i : ð12Þ

The posterior density pðμiþ1 =Z1:iþ1 Þ is then obtainable

by Bayes rule, making use of the likelihood kernel
gðZiþ1 =μiþ1 Þ of the observation Ziþ1 given μiþ1 , and the prior
qðμiþ1 =Z1:i Þ as

gðZiþ1 =μiþ1 Þqðμiþ1 =Z1:i Þ

pðμiþ1 =Z1:iþ1 Þ ¼ R ; ð13Þ
gðZiþ1 =μiþ1 Þqðμiþ1 =Z1:i Þdμiþ1

where the likelihood kernel function gðZiþ1 =μiþ1 Þ is obtained

from the observation equation as

gðZiþ1 =μiþ1 Þ ¼ pϑ;iþ1 ðZiþ1 − DKðμa Þ−1 qÞ: ð14Þ

pϑ;iþ1 is the zero-mean Gaussian density for Wiener measure-

ment noise ϑiþ1 . Accordingly, we have the following propor-
tionality between the unnormalized weights and likelihood
kernel:

~ ðuÞ
w
ðuÞ
iþ1 ∝ gðZiþ1 =μiþ1 Þ; ð15Þ
ðuÞ
where fμiþ1 g denotes the ensemble of predicted particles.
Fig. 1. (Color online) Parameters estimated through the BS and
The filtering distribution at t ¼ tiþ1 can be approximately re-
GS filters for an absorbing object with one inclusion. At the start
presented in the MC set up through the following empirical of recursion, we assume there are two inclusions. (a) h, (b) r, and
measure: (c) the Euclidean distance to the center of the recovered inclusion.
Raveendran et al. Vol. 28, No. 10 / October 2011 / J. Opt. Soc. Am. A 2075

Pðdμiþ1 =Z1:iþ1 Þ ≈ P ðNÞ ðdμiþ1 =Z1:iþ1 Þ where ℵðμi ; mil ; Σil Þ denotes the normal density correspond-
ing to the random vector μi with mean mil and covariance Σil ,
gðZiþ1 =μiþ1 ÞQðNÞ ðdμiþ1 =Z1:i Þ
¼R assumed to be nonsingular. Within the setup of the GSPF, the
GðZiþ1 =μiþ1 ÞQðNÞ ðdμiþ1 =Z1:i Þ
particles (which represent the empirical approximation to the
N
X ðuÞ above density) are split into N G groups, each containing
¼ wiþ1 δ ðuÞ dðμiþ1 Þ: ð17Þ
μiþ1 M≔N=N G particles drawn from each term (mixand) in the
u¼1
mixture. Now, by approximating the prediction density as a
ðuÞ Gaussian mixture at t ¼ tiþ1 , one has
The weights are normalized such that N
P
u¼1 wiþ1 ¼ 1. After
resampling [18] we get the unweighted (or equally weighted)
filtered particles corresponding to the empirical distribution
N
X
P N ðdμiþ1 =Z1:iþ1 Þ ¼ ð1=NÞ δμðuÞ ðdμiþ1 Þ: ð18aÞ
iþ1
u¼1

The estimate μ ^ iþ1 ¼ Eðμiþ1 =Z1:iþ1 Þ for the optical param-

eters of interest is presently given by

N
1X ðuÞ
^ iþ1 ¼
μ μ : ð18bÞ
N u¼1 iþ1

2. GSPF
The GSPF approximates the predictive and posterior distribu-
tions as Gaussian mixtures, i.e., weighted sums of Gaussian
densities to approximate the filtering density. This is moti-
vated by the fact that any non-Gaussian density can be ade-
quately approximated by a Gaussian mixture containing a
sufficiently large number of terms [21,22] and with the coeffi-
cients in the mixture properly adjusted. Thus, expressing the
posterior density at t ¼ ti approximately by a Gaussian mix-
ture, we have

NG
X
pðμi =Z1:i Þ ≈ wil ℵðμi ; mil ; Σil Þ; ð19Þ
l¼1

Fig. 3. (Color online) Parameters estimated through PF algorithm

Fig. 2. (Color online) Recovered μa images. (a) Gray-level images for a phantom with two inhomogeneities, (a) h, (b) r, and
and (b) cross-sectional plots through the center of the inclusion. (c) Euclidean distance.
2076 J. Opt. Soc. Am. A / Vol. 28, No. 10 / October 2011 Raveendran et al.

Table 1. Comparison of Integrated Photon Loss

From Simulated Data (Two Inhomogeneities)
Small Large Small Large Small Large
BS μa ¼ 0:0159 μa ¼ 0:0303 R ¼ 8:7863 R ¼ 7:2415 Photon loss ¼ 0:7562 Photon loss ¼ 0:6648
GSPF1 μa ¼ 0:0205 μa ¼ 0:0302 R ¼ 7:0849 R ¼ 7:2963 Photon loss ¼ 0:7479 Photon loss ¼ 0:6436
GSPF2 μa ¼ 0:0213 μa ¼ 0:0307 R ¼ 6:7002 R ¼ 7:2475 Photon loss ¼ 0:7517 Photon loss ¼ 0:6408
From Experimental Data
BS μa ¼ 0:0043 R ¼ 6:8656 Photon loss ¼ 0:9427
GSPF1 μa ¼ 0:0136 R ¼ 3:4887 Photon loss ¼ 0:9095
GSPF2 μa ¼ 0:0495 R ¼ 2:0667 Photon loss ¼ 0:8150

Z
NG
qðμiþ1 =Z1:i Þ ¼ χðμiþ1 =μi Þpðμi =Z1:i Þdμi
X
qðμiþ1 =Z1:i Þ ¼  iþ1;l ℵðμiþ1 ; m
w  iþ1;l Þ:
 iþ1;l ; Σ ð22Þ
l¼1
Z NG
X
¼ ℵðμiþ1 ; μi ; Σζiþ1 Þ wil On receiving the measurement Ziþ1 , the filtering (posterior)
l¼1
density pðμiþ1 =Z1:iþ1 Þ can similarly be approximated as a
× ℵðμi ; mil ; Σil Þdμi Gaussian mixture:
NG Z
gðZ1:iþ1 =μiþ1 Þqðμiþ1 =Z1:iþ1 Þ
X
¼ wil ℵðμiþ1 ; μi ; Σζiþ1 Þ pðμiþ1 =Z1:iþ1 Þ ¼ R
l¼1 gðZ1:iþ1 =μiþ1 Þqðμiþ1 =Z1:iþ1 Þdμiþ1
× ℵðμi ; mil ; Σil Þdμi ; ð20Þ ∝ gðZiþ1 =μiþ1 Þqðμiþ1 =Z1:i Þ
∝ gðZiþ1 =μiþ1 Þ
where, from the process Eq. (11a), we get χðμiþ1 =μi Þ to be NG
Gaussian with the mean vector μi and covariance matrix
X
×  iþ1;l ℵðμiþ1 ; m
w  iþ1;l Þ;
 iþ1;l ; Σ ð23Þ
Σζiþ1 ≔σTμ σμ Δt. As with the posterior density Eq. (19), the pre- l¼1

dicted particles are also split in N G groups. Moreover, the M where the likelihood kernel function gðZiþ1 =μiþ1 Þ is
predicted particles in any specific group are arrived at [via the obtained from Eq. (14). In practice, we obtain M samples
process Eq. (11a)] using those in the associated parent group ðuÞ M
μiþ1;l gu¼1 from each of the distributions ℵðμiþ1 ; m
f~  iþ1;l ;
(representing the last posterior) as the initial conditions. Thus,
the analytically intractable integrals in Eq. (20) are computed  iþ1;l Þ, l ¼ 1; …; N G and thus compute the weights fψ ðuÞ g≔
Σ iþ1;l
ðuÞ
within an MC setup; i.e., for each group we draw M samples ~ iþ1;l Þg. Since the lth mixand on the RHS of
fgðZiþ1 =μiþ1 ¼ μ
each from ℵðμi ; mil ; Σil Þ; l ∈ ½1; N G Š, leading to N predicted Eq. (23), given by ℵðμiþ1 ; m  iþ1;l ÞgðZiþ1 =μiþ1 Þ, can be
 iþ1;l ; Σ
ðuÞ M
particles ffμiþ1;l gu¼1 ; l ∈ ½1; N G Šg. Also, for l ¼ 1; …; N G , the approximated as Gaussian [22], the weighted samples
ðuÞ ðuÞ M
weights corresponding to each mixand (represented by one μiþ1;l ; ψ iþ1;l gu¼1 for the lth mixand remains approximately
f~
group M of particles in the GSPF) are updated as Gaussian. We can now estimate the corresponding mean
and covariance (l ¼ 1; …; N G ) as
w
 iþ1;l ¼ PN il :
w ð21Þ PM ðuÞ ðuÞ
l¼1 wil
G
u¼1ψ iþ1;l μ
~ iþ1;l
miþ1;l ¼ PM ðuÞ
; ð24Þ
u¼1 ψ iþ1;l
ðuÞ M
For the predicted groups of particles ffμiþ1;l gu¼1 ;
l ∈ ½1; N G Šg, one can readily compute for each group l the sam- PM ðuÞ ðuÞ ðuÞ T
u¼1 μiþ1;l − miþ1;l Þð~
ψ iþ1;l ð~ μiþ1;l − miþ1;l Þ
ple mean m  iþ1;l . For compu-
 iþ1;l and the sample covariance Σ Σiþ1;l ¼ : ð25Þ
PM ðuÞ
tational purposes, the prediction density can then be recast as u¼1 ψ iþ1;l

Fig. 4. (Color online) Reconstruction of absorption coefficient using (a) BS and (b) GSPF for a phantom with two inhomogeneities.
Raveendran et al. Vol. 28, No. 10 / October 2011 / J. Opt. Soc. Am. A 2077

Specifically, the estimate μ

^ iþ1 ¼ Eðμiþ1 =Z1:iþ1 Þ for the optical
parameters of interest is presently given by

NG
X
^ iþ1 ¼
μ wiþ1;l miþ1;l : ð29Þ
l¼1

Unlike the BS filter, the GSPF avoids the resampling stage,

thereby leading to reduced sample variance in the estimated
parameters. Moreover, the grouping of the set of particles,
with each group being drawn from each Gaussian term in
the mixture, also plays a critical role in faster convergence
and sample variance reduction.

3. VERIFICATION OF RECONSTRUCTION
METHODS
A. Through Numerical Simulations
Fig. 5. (Color online) Comparison of simulated reference data and
the data computed with the estimates by BS and GSPF. For all numerical simulations, we have considered two-
dimensional objects that are cross sections of a cylinder of
The mixand weights are then updated and normalized as diameter 86 mm. The background optical properties are taken
to be μba ¼ 0:01 mm−1 and μ0b −1
s ¼ 1:0 mm . In the first set of si-
PM ðuÞ mulations, the above object had a circular absorption inhomo-
ψ iþ1;l
w  il PN u¼1
~ iþ1;l ¼ w ðuÞ
; l ¼ 1; …; N G ; ð26Þ geneity of diameter 15 mm located at (21.5, 0 mm) [assuming
M
P
u¼1 ψ iþ1;l the center of the big circle at (0,0)] and of μa ¼ 0:02 mm−1 . For
G
l¼1
data gathering, we use one source and 16 detectors equiangu-
larly placed on the diametrically opposite side of the source.
w
~ iþ1;l
wiþ1;l ¼ : ð27Þ Rotating the source–detectors combination in unison, we
NG
X
~ iþ1;l have gathered 15 sets of 16 measurements each by solving
w
l¼1
the forward equation [Eq. (1)] after finite element (FE) discre-
tization that used 850 three-noded triangular elements (invol-
The filtering density at t ¼ tiþ1 may now be represented as ving 451 nodes). The data are in fact the photon flow through
the detector which, because of the boundary condition of
NG
X Eq. (1b), is proportional to ϕ. The measurement data [ϕðmÞ ]
pðμiþ1 =Z1:iþ1 Þ ¼ wiþ1;l ℵðμiþ1 ; miþ1;l ; Σiþ1;l Þ: ð28Þ are obtained by adding a suitable percentage of noise to
l¼1 the computed ϕ. For the above object the parameter vector

Fig. 6. (Color online) Recovered gray-level images from noisy fluence data using (a) BS filter, (b) GN method. The percentage noise in data is 3.

Fig. 7. (Color online) Recovered gray-level images from noisy fluence data using (a) BS filter (b) GN method. The percentage noise in data is 5.
2078 J. Opt. Soc. Am. A / Vol. 28, No. 10 / October 2011 Raveendran et al.

and close to the background value for the false. Figures 1(b)
and 1(c) show the estimation of r and the Euclidian distance
to the center, corresponding to the true inhomogeneity from
which we gather that, whereas the BS-filter estimate of r is
closer to the true value, for the Euclidean distance, the GS-
filter estimate is more accurate. Figure 2 gives gray-level
images of the original and the recovered objects and cross-
sectional plots [Fig. 2(b)] through the center of the inclusions.
Figure 2(b) has also the reconstruction from the GN algo-
rithm. The initial noise intensities for the process noise are
chosen to be σ 2h ¼ 10−6 (unit in mm−2 ), σ 2r ¼ 3, σ 2x ¼ 7, and
σ 2y ¼ 7 (units in mm2 ). The standard deviation of the
observation noise is taken in all simulations to be 1% of the
maximum absolute recorded measurement.
In the second numerical experiment, we have used the
Fig. 8. (Color online) Cross section of the phantom with a central same absorbing object (and similar data set) as the one used
inhomogeneity estimated via BS and GSPF along with GN reconstruc- above except for the modification that it has two absorbing
tion and reference. inhomogeneities. They are both circular of diameter 15 mm
located at (21.5, 0 mm) and (−21.5, 0 mm) and of value 0.02
μp to be estimated consists of the height (h), radius (r), and and 0:03 ðmm−1 Þ, respectively. Figure 3 shows recovered
the coordinates ðx; yÞ of the center of the inhomogeneity. The parameters (h, r, and the Euclidean distance to the center)
initial noise intensities for the process noise is obtained as of the two inhomogeneities obtained via the BS and GS filters.
Σζ1 ≔diag½σ 2h ; σ 2r ; σ 2x ; σ 2y Š. These noise intensities are made to The initial intensities for the process noise have been chosen
exponentially decrease with each pseudo-time step Δt ¼ as σ 2h ¼ 10−5 , σ 2r ¼ 5, σ 2x ¼ 7, and σ 2y ¼ 7. Figure 3(a) shows the
tnþ1 − tn ¼ 0:01, (assumed uniform throughout the recursion). reconstructed h for the two inclusions. It is seen that, whereas
for the inclusion 2, both BS and GS filters give reconstructions
We reconstruct the parameters from noisy ϕðmÞ initially as-
close to the actual 0:03 mm−1 , for the first the GS filter gave a
suming that there are two inhomogeneous inclusions and
more accurate recovery with the BS filter converging to a
therefore eight unknowns to be estimated. Figure 1(a) shows value less than the true value of 0:02 mm−1 . The results for the
the reconstructed h for the true and false inclusions from radii are shown in Fig. 3(b) where for inhomogeneity 2, both
which it is clear that both the BS and Gaussian-sum (GS) fil- the filters converge to a value very close to the actual 7:5 mm.
ters estimated h close to its true value for the true inclusion For the other inhomogeneity the GS filter converged to a value

Fig. 9. (Color online) Reconstruction of μa using (b) GN, (c) GSPF, and (d) BS schemes for an object with a central inclusion; (a) is the reference.
Raveendran et al. Vol. 28, No. 10 / October 2011 / J. Opt. Soc. Am. A 2079

close to the actual one and the BS filter to a slightly higher [Eq. (1)] used to propagate photons, as also verified by the
(∼8:5 mm). It is interesting to note that the reconstruction almost exact match of the computed ϕ on the boundary using
of the first inhomogeneity by the BS filter, which is larger the recovered parameters with the measured (or numerically
in size but smaller in height, i.e., the μa value, results in the generated) data we had started the algorithm with (Fig. 5).
same overall light absorption [calculated as expð−μa DÞ, where Figure 4 gives gray-level maps of the recovered μa from the
μa and D are returned absorption coefficient and diameter of two filters.
the inclusion, respectively] as that from the GS filter, which We have also applied the BS filter to estimate parameters
has D smaller and μa larger. Table 1 gives details of this com- from noisy data. To the simulated ϕ corresponding to the
object with two inhomogeneities, we have added Gaussian
parison. These two results are admissible solutions of the DE
noise with the 3% and 5% strength. The gray-level images of
the reconstructions are given in Figs. 6 and 7, respectively,
wherein for comparison, reconstructions using the GN
scheme are also shown. The results from the BS filter are un-
affected by data noise of the strength considered here. The
superior performance (and lesser spread of the reconstructed
inhomogeneities) is attributable to a more accurate stochastic
projection scheme on the noise filtration.
In our next numerical example, we have considered an ob-
ject with an inhomogeneous μa inclusion in the center, a case
for which contrast recovery with deterministic algorithms is
the most inadequate owing to the relatively poor sensitivity of
measurement to the unknown parameters. We demonstrate
here the superior performance of the PF-based estimation
(which does not require any derivatives) for a proper recovery
of contrast in this case. The size of background cylinder and
its optical properties are the same as those in the earlier
example. There is a circular inhomogeneity at the center of
diameter 15:00 mm and μa ¼ 0:02 mm−1 . We have chosen
the initial 23 intensities for the process noise as σ 2h ¼ 10−5 ,
σ 2r ¼ 3, σ 2x ¼ 7, and σ 2y ¼ 7. The cross-sectional plot and the
reconstructed inclusion using BS, GS, and GN algorithms
are shown in Fig. 8. Figure 9 shows the corresponding
gray-level plots. It is seen that both of the PF-based methods
give an accurate recovery of contrast when compared to the
GN scheme.

B. Through Reconstruction from Experimental Data

Finally, the working of the PFs is tested by estimating
unknown μa from experimentally gathered fluence data. The
experimental data are the same as that used in an earlier pub-
lication from us [32] wherein a video-rate NIR tomography
system was used to gather data pertaining to an object with
an inhomogeneous inclusion of intralipid solution whose μa is
varied dynamically. The object itself is cylindrical of diameter
27 mm, and intralipid anomaly of ∼4 mm diameter is at

Fig. 10. (Color online) Recovered (a) h, (b) r, and (c) the Euclidean
distance, estimated through BS, GSPF1, and GSPF2 filters from the Fig. 11. (Color online) Reconstructed gray-level image from experi-
experimental data. mental data using the BS.
2080 J. Opt. Soc. Am. A / Vol. 28, No. 10 / October 2011 Raveendran et al.

position (10, 0 mm). There are eight fibers around the object
serving as sources and detectors; when one delivers light,
all the others receive the exiting light. A CCD array is used
to image and quantify the intensity of light from the fibers.
Altogether, 64 measurements of intensity are taken.
We have employed both BS and GSPF to estimate the dy-
namic μa corresponding to the data collected at time t ¼ 100 s.
The results are presented in Figs. 10–13. Figure 10 shows the
estimated h, r, and the Euclidean distance from the origin to

Fig. 15. (Color online) Comparison of experimental data with

computed data using the estimates obtained by the GSPF filter for
which the number of nodes in the FE discretization is 800.

the center of the inclusion using the BS filter and two different
GS filters, namely GSPF1 and GSPF2, which are different
because σ 2h is chosen to be 10−6 and 10−4 , respectively (all
the other noise intensities are identical, which are σ 2r ¼ 1,
Fig. 12. (Color online) Reconstructed gray-level image from
σ 2x ¼ 7, and σ 2y ¼ 7). The noise intensities for the BS filter
experimental data using the GSPF1. are kept the same as for GSPF1. We see that the heights re-
turned (i.e., the value of μa ) differ, the smallest for BS, then
for GSPF1, and the largest for GSPF2. The radii returned
[Fig. 10(b)] show a reversal of this pattern, the largest for
BS, the smallest for GSPF2, and the GSPF1 coming in be-
tween. As we have remarked in Subsection 3.A dealing with
numerically simulated data, here again photon loss as esti-
mated as expð−μa DÞ remains almost the same for the three
reconstructions. (See also Table 1.) To prove that the recon-
structions from all the PFs are valid solutions, we solved the
forward equation (discretized with number of nodes and ele-
ments taken to be 451 and 850) with the recovered optical
parameters as inputs. The results are in Fig. 14, which shows
an almost perfect match of the computed data with the experi-
mental one. Figure 15 shows a similar comparison just for
GSPF-generated data to show that, when the total number
Fig. 13. (Color online) Reconstructed gray-level image from of nodes is 800, the forward equation has converged without
experimental data using the GSPF2. any mesh-density-dependent variation in the solution.

4. CONCLUSIONS
We have developed two pseudo-time PF-based reconstruction
strategies for solving the inverse problem of DOT. Of the two,
the BS filter is the more basic one, which is improved upon
in the GS filter, wherein Gaussian mixture representations
of the prediction and filter densities provide for a more uni-
form population of the particles within the sample space of the
process variables (i.e., the optical parameters to be recon-
structed) and hence a reduced possibility of filter divergence.
The cardinal advantage of the implemented PF-based recon-
struction algorithm is that it does not use a sensitivity matrix
en route to parameter recovery, and therefore the contrast
recovery away from the source and detector (which is usually
poor in an algorithm using a deterministic error minimiza-
tion approach owing to the poor sensitivity of data on param-
Fig. 14. (Color online) Comparison of experimental data with
eters) is quite good. To render the MC-based implementation
computed data using the estimates obtained by BS, GSPF1, and of the PF efficient, we have drastically reduced the number
GSPF2 filters. of unknowns by representing the unknown inclusions by
Raveendran et al. Vol. 28, No. 10 / October 2011 / J. Opt. Soc. Am. A 2081

appropriate indicator functions while riding on a uniform from sparse and noisy data sets: near-infrared fluorescence
tomography,” Proc. Natl. Acad. Sci. USA 99, 9619–9624 (2002).
background property. Considering the fact that, in diagnostic
13. M. J. Eppstein, D. E. Dougherty, D. J. Hawrysz, and E. M. Sevick-
imaging of malignancy, the tumor is a localized perturbation Muraca, “Three-dimensional Bayesian optical image reconstruc-
in the average normal tissue property, limited to one or two tion with domain decomposition,” IEEE Trans. Med. Imag. 20,
in number, our representation here is often suitable and 147–163 (2001).
adequate. The advantage we reap out of this representation 14. B. Banerjee, D. Roy, and R. M. Vasu, “A pseudo-dynamic sub-
optimal filter for elastography under static loading and measure-
and the recovery only of the needed parameters, is that the ments,” Phys. Med. Biol. 54, 285–305 (2009).
diffusion-induced spread is eliminated. The other advantages 15. B. Banerjee, D. Roy, and R. M. Vasu, “A pseudo-dynamical sys-
are in a more appropriate formal treatment of the measure- tems approach to a class of inverse problems in engineering,”
ment noise as well as the ability to reconstruct from relatively Proc. R. Soc. A 465, 1561–1579 (2009).
poorer signal-to-noise ratio data. Moreover, unlike the var- 16. A. Doucet, S. Godsill, and C. Andrieu, “On sequential Monte
Carlo sampling methods for Bayesian filtering,” Stat. Comput.
iants of the pseudo-time KF-based approaches [14,15,33], the 10, 197–208 (2000).
presently adopted formulations are strictly derivative-free and 17. A. Doucet, N. de Freitas, and N. Gordon, Sequential Monte Carlo
not restricted by the questionable assumption of Gaussianity Methods in Practice (Academic, 2001).
of the unknowns. 18. S. Arulampalam, N. Maskell, N. Gordon, and T. Clapp,
“A tutorial on particle filters for online nonlinear/non-Gaussian
Bayesian tracking,” IEEE Trans. Signal Process. 50, 174–188
ACKNOWLEDGMENTS (2002).
The experimental data on the DOT, which has also been used 19. N. J. Gordon, D. J. Salmond, and A. F. M. Smith, “Novel approach
to nonlinear/non-Gaussian Bayesian state estimation,” in IEE
in one of our earlier papers [32], has been kindly provided Proceedings F Radar and Signal Processing (IEEE, 1993),
by D. Piao. Vol. 140, pp. 107–113.
20. K. Murphy and S. Russell, “Rao-Blackwellised particle filtering
REFERENCES for dynamic Bayesian networks,” in Sequential Monte Carlo
Methods in Practice, A. Doucet, N. de Freitas, and N. Gordon,
1. S. Srinivasan, B. W. Pogue, S. Jiang, H. Dehghani, C. Kogel,
eds. (Academic 2001), pp. 499–515.
S. Soho, J. J. Gibson, T. D. Tosteson, S. P. Poplack, and K. D. 21. J. H. Kotecha and P. M. Djuric, “Gaussian sum particle filtering,”
Paulsen, “Interpreting hemoglobin and water concentration, IEEE Trans. Signal Process. 51, 2602–2612 (2003).
oxygen saturation and scattering measured in vivo by near-
22. J. H. Kotecha and P. M. Djuric, “Gaussian sum particle filtering
infrared breast tomography,” Proc. Natl. Acad. Sci. USA 100,
for dynamic state space models,” in Proceedings of IEEE Inter-
12349–12354 (2003). national Conference on Acoustics, Speech, and Signal Proces-
2. D. A. Boas, D. H. Brooks, E. L. Miller, C. A. DiMarzio, M. Kilmer, sing (IEEE, 2001), pp. 3465–3468.
R. J. Gaudette, and Q. Zhang, “Imaging the body with diffuse
23. M. Schweiger, S. R. Arridge, and I. Nissila, “Gauss–Newton
optical tomography,” IEEE Signal Process. Mag. 18(6), 57–75
method for image reconstruction in diffuse optical tomography,”
(2001). Phys. Med. Biol. 50, 2365–2386 (2005).
3. B. W. Pogue, S. P. Poplack, T. O. McBride, W. A. Wells, O. K. S., 24. K. Levenberg, “A method for the solution of certain non-
U. L. Osterberg, and K. D. Paulsen, “Quantitative hemoglobin
linear problems in least squares,” Quart. Appl. Math. 2, 164–168
tomography with diffuse near-infrared spectroscopy: pilot re-
(1944).
sults in the breast,” Radiology 218, 261–266 (2001). 25. D. W. Marquardt, “An algorithm for least squares estimation
4. B. J. Tromberg, N. Shah, R. Lanning, A. Cerussi, J. Espinoza, of nonlinear parameters,” J. Soc. Ind. Appl. Math. 11, 431–441
T. Pham, L. Svaasand, and J. Butler, “Non-invasive in vivo
(1963).
characterization of breast tumors using photon migration
26. D. Roy, “Explorations of the phase space linearization method
spectroscopy,” Neoplasia 2, 26–40 (2000). for deterministic and stochastic non-linear dynamical systems,”
5. S. R. Arridge, “Optical tomography in medical imaging,” Inverse Nonlinear Dyn. 23, 225–258 (2000).
Probl. 15, R41–R93 (1999).
27. D. Roy, “A new numeric-analytical principle for nonlinear deter-
6. P. K. Yalavarthy, B. W. Pogue, H. Dehghani, and K. D. Paulsen,
ministic and stochastic dynamical systems,” Proc. R. Soc.
“Weight-matrix structured regularization provides optimal gen- London Ser. A 457, 539–566 (2001).
eralized least-squares estimate in diffuse optical tomography,” 28. G. Kallianpur, Stochastic Filtering Theory (Academic, 1980).
Med. Phys. 34, 2085–2098 (2007).
29. A. D. Zacharopoulos, M. Schweiger, V. Kolehmainen, and
7. C. R. Vogel, Computational Methods for Inverse Problems
S. Arridge, “3D shape based reconstruction of experimental data
(Academic, 2002). in diffuse optical tomography,” Opt. Express 17, 18940–18956
8. P. C. Hansen, “Analysis of discrete ill-posed problems by means (2009).
of the L-curve,” SIAM Rev. 34, 561–580 (1992).
30. R. Lipster and A. Shiryaev, Statistics of Random Processes
9. H. W. Engl, M. Hanke, and A. Neubauer, Regularization of the
(Academic, 2001).
Inverse Problem (Academic, 1996). 31. P. Fernhead, “Sequential Monte Carlo methods in filter theory,”
10. V. Kolehmainen, S. Prince, S. R. Arridge, and J. P. Kaipo, “State- Ph.D. thesis (University of Oxford, 1998).
estimation approach to the nonstationary optical tomography
32. S. Gupta, P. K. Yalavarthy, D. Roy, D. Piao, and R. M. Vasu,
problem,” J. Opt. Soc. Am. A 20, 876–889 (2003).
“Singular value decomposition based computationally efficient
11. M. J. Eppstein, D. E. Dougherty, T. L. Troy, and E. M. Sevic- algorithm for rapid dynamic near-infrared diffuse optical tomo-
Muraca, “Biomedical optical tomography using dynamic param- graphy,” Med. Phys. 36, 5559–5567 (2009).
etrization and Bayesian conditioning on photon migration
33. B. Banerjee, D. Roy, and R. M. Vasu, “Efficient implementations
measurements,” Appl. Opt. 38, 2138–2150 (1999).
of a pseudo-dynamical stochastic filtering strategy for static
12. M. J. Eppstein, D. J. Hawrysz, A. Godavarty, and E. M. Sevick- elastography,” Med. Phys. 36, 3470–3476 (2009).
Muraca, “Three-dimensional, Baysian image reconstruction