You are on page 1of 22

Applied Mathematics and Computation 161 (2005) 149–170

www.elsevier.com/locate/amc

1D inverse problem in diffusion based


optical tomography using
iteratively regularized
Gauss–Newton algorithm q
a,* b
T. Khan , A. Smirnova
a
Department of Mathematical Sciences, O-201 Martin Hall, P.O. Box 340975,
Clemson, SC 29634-0975, USA
b
Department of Mathematics and Statistics, Georgia State University, Atlanta, GA 30303, USA

Abstract
In this paper, we investigate an one-dimensional inverse problem in diffusion based
optical tomography using iteratively regularized Gauss–Newton (IRGN) algorithm for
ill-posed nonlinear problems. We devise a stable reconstruction algorithm for the in-
verse problem using iterative regularization with Armijo–Goldstein–Wolf (AGW) type
line search strategy. We demonstrate the efficacy of the IRGN combined with AGW by
reconstructing the scattering parameter relevant to the inverse problem in optical
tomography.
Ó 2003 Elsevier Inc. All rights reserved.

Keywords: Inverse problems; Nonlinear ill-posed; Iteratively regularized Gauss–Newton; Biomed-


ical imaging; Reconstruction algorithms

1. Introduction

In the last decade research in the area of biomedical optics has flourished. In
particular there has been considerable new development in biomedical imaging

q
This work is supported by NSF grant (award number DMS-0207050).
*
Corresponding author.
E-mail addresses: khan@clemson.edu (T. Khan), matabs@suez.cs.gsu.edu (A. Smirnova).

0096-3003/$ - see front matter Ó 2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2003.12.019
150 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

using optical tomography [2,8,10,18]. Optical tomography is a way to probe


highly scattering media using low-energy visible or near infra-red light (NIR)
and then to reconstruct images of these media. Light in the near-infrared range
(wavelength from 700 to 1200 nm) penetrates tissue and interacts with it. The
predominant effects are absorption and scattering [6,11,14]. We assume: given a
set of measurements of transmitted light between pairs of points on the surface
of an object, there exists a unique distribution of internal scatters and
absorbers which would yield that set. The formation of an image for the optical
properties of the tissue from a series of boundary measurements is the inverse
problem of optical absorption and scattering tomography (OAST) or optical
tomography (OT) for short. The widely accepted photon transport model is the
radiative transfer equation (RTE). The transport equation is an integro-dif-
ferential equation for the radiance and has spatially dependent diffusion and
absorption parameters as coefficients which are a priori unknown. Hence the
problem is to infer from the measurements of some function of the photon
density on the boundary, the coefficients of absorption and diffusion in the
tissue.
A low order diffusion approximation to the RTE has been derived and
studied in the last several years. This is an approximation to the RTE by a
parabolic differential equation in the time domain and by an elliptic differential
equation in the frequency domain [2]. The diffusion approximation to the
transport equation has been widely used to calculate photon migration in
biological tissues [15]. The existing computational methods for the inverse
problem for photon migration in biological tissues are almost exclusively based
on the diffusion approximation [7].
It is well known that the diffusion based inverse problem in optical
tomography is exponentially ill-posed or unstable [2,23]. In order to under-
stand the effect of iteratively regularized Gauss–Newton (IRGN) algorithm on
the stability of the inverse problem, we investigated the one-dimensional in-
verse problem in diffusion based optical tomography. The outline of the paper
is as follows. In Section 2, we discuss the transport model and its diffusion
approximation. In Section 3, we describe system approximation using finite
element method (FEM). In Section 4, we describe the iterative Gauss–Newton
algorithm for the inverse problem. In Section 5, we discuss our results and
work in progress.

2. Photon transport model

In optical imaging, low-energy visible light is used to illuminate the bio-


logical tissue. The illumination of the tissue can be modelled as a photon
transport phenomenon. The process is described by the most widely applied
equation in optical imaging, the radiative transfer or transport equation (RTE)
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 151

[9,17]. Let X  Rn , n P 2 be bounded with C 2 ––smooth boundary oX, mðxÞ


denote the outward unit normal to oX at the point x 2 oX, and C is defined as
 
C :¼ ðx; s; tÞ 2 oX  S n1  ½0; T
;  mðxÞ s > 0 :

  S n1  ½0; T
Þ and f 2 L1 ðX
Let a, b 2 L1 ðXÞ with a > b > 0, u 2 H 1 ðX 
n1
S  ½0; T
Þ. Then the RTE is given by

1 ou
ðx; s; tÞ þ s ruðx; s; tÞ þ aðxÞuðx; s; tÞ
c ot Z
 bðxÞ Hðs s0 Þuðx; s0 ; tÞ ds0 ¼ f ðx; s; tÞ ð1Þ
S n1

together with the initial and boundary conditions

uðx; s; 0Þ ¼ 0 in X  S n1 ; ð2Þ

uðx; s; tÞ ¼ 0 on C ; ð3Þ

where uðx; s; tÞ describes the density function of particles (photons) which


travel in X at time t though the point x in the direction s 2 S n1 , unit sphere
in Rn . The parameter a is the total cross section or attenuation, and b is
the scattering cross section. The difference l :¼ a  b has the physical meaning
of an absorption cross section. The parameters a, b and l are assumed to
be real, nonnegative and bounded functions of the position. The parameters a
and b are the sought-for tissue parameters and c is the velocity of light.
The function H is the scattering phase function characterizing the intensity of
a wave incident in direction s0 scattered in the direction s. It is assumed to be
a real, nonnegative function and is normalized to one. The inverse problem
is to recover a and b from measurements of some given functionals of the
outgoing density uj jCþ at the boundary oC for ms different set of source dis-
tribution fj , j ¼ 1; . . . ; ms . An example of a measured quantity in optical
tomography is
Z
1
gðx; tÞ ¼ mðxÞ suðx; s; tÞ ds; x 2 oX; t P 0; ð4Þ
4p S 2

where g is known for many incoming fluxes fj .


Simpler deterministic models can be derived from RTE for a constant
refractive index by expanding the density u and source f in spherical harmonics
and retaining a limited number of terms [2,3,5,20]. Due to the prevalence of
scattering, the flux is essentially isotropic a small distance away from the
sources; i.e., it depends only linearly on s. Thus we may describe the process
adequately by the first two moments
152 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170
Z
1
u0 ðx; tÞ ¼ uðx; s; tÞ ds;
4p S 2
Z
1
u1 ðx; tÞ ¼ suðx; s; tÞ ds
4p S 2
of u. We assume that u depends only linearly on s, i.e.,
uðx; s; tÞ ¼ c0 u0 ðx; tÞ þ c1 s u1 ðx; tÞ; ð5Þ
where the constants c0 ¼ 1 and c1 ¼ 3 are easily obtained by computing the
moments of order 0 and 1. If we let H  is the mean scattering cosine
Z
 ¼ 1
H s0 sHðs s0 Þ ds0
4p S 2
which does not depend on s and if we introduce the reduced scattering cross
 and
section b0 ¼ ð1  HÞ
1

3ðl þ b0 Þ
which is the diffusion coefficient. We get diffusion approximation P0 (DA):
1 ou0
ðx; tÞ  r DðxÞru0 ðx; tÞ þ la ðxÞu0 ðx; tÞ ¼ f0 ðx; tÞ ð6Þ
c ot
together with initial condition
u0 ðx; 0Þ ¼ 0 in X; ð7Þ
and the boundary condition
ou0
u0 ðx; tÞ þ 2DðxÞ ðx; tÞ ¼ 0; x 2 oX: ð8Þ
om
Similarly the equation for the measurement reads
ou0
gðx; tÞ ¼ mðxÞ u1 ðx; tÞ ¼ DðxÞmðxÞ ru0 ðx; tÞ ¼ DðxÞ : ð9Þ
om
Frequency domain diffusion approximation can easily be obtained by
Fourier transforming the time-domain equation. The frequency domain ana-
logue of Eq. (6) is given by
 
ix
r DðxÞru0 ðx; xÞ þ la ðxÞ þ u0 ðx; xÞ ¼ f0 ðx; xÞ; ð10Þ
c
where we made use of the relation
o
 ix:
ot
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 153

The frequency domain DE is elliptic where the time-domain DE is para-


bolic, an important distinction for numerical solutions. Furthermore, the dif-
fusion approximation to the radiative transfer model can be written in the time
independent (dc) case as in [2]
r DðxÞru0 ðxÞ þ la ðxÞu0 ðxÞ ¼ f0 ðxÞ: ð11Þ
The associated boundary condition is
ou0
u0 ðxÞ þ 2DðxÞ ðxÞ ¼ 0; x 2 oX: ð12Þ
om
If we let X be the domain under consideration with surface oX, the weak
forward problem corresponding to Eq. (11) is to find uðxÞ 2 H 1 ðXÞ such that
for all vðxÞ 2 H 1 ðXÞ, the following variational equation is satisfied
Z Z Z Z
1
rv Dru dX þ vla u dX þ vu ds ¼ vf dX: ð13Þ
X X oX 2 X

Now, we can define the forward problem as: given sources fj in X and q in Q,
a vector of model parameters, for example the coefficient of diffusion D and the
T
coefficient of absorption la (i.e. q ¼ ðD; la Þ ) that belongs to a parameter set Q,
find the data u on oX and the inverse problem as: given data z on oX find q. We
can recast the forward problem in an abstract setting as the following
parameter dependent equation:
r DðxÞruðx; qÞ þ la ðxÞuðx; qÞ ¼ f ðxÞ; ð14Þ
where x is in X, uðqÞ is in an appropriate abstract space H , and f represents a
source or a forcing distribution. In general, measurement of uðqÞ may not be
possible, only some observable part CuðqÞ of the actual state uðqÞ may be
measured. In this abstract setting, the objective of the inverse or parameter
estimation problem is to choose a parameter q in Q, that minimizes an error
criterion or cost functional J ðuðqÞ; CuðqÞ; qÞ over all possible q in Q e Q
subject to uðqÞ satisfying the diffusion approximation. A typical observation
operator is
 m  m
ou 1
Cf uðqÞ ¼  D ðxi ; q; f Þ ¼ uðxi ; q; f Þ ; ð15Þ
om i¼1 2 i¼1

where xi is in oX, m is the number of measurements, and the second equality


comes from the boundary condition (12). For example, in one dimension with
X ¼ ½0; L
, we have m 2 f1; 2g and x1 ¼ 0, x2 ¼ L. A typical cost functional Jk is
given as
m

2
1X ms X

fj f
2
Jk ðqÞ ¼
Ci uðqÞ  zi j
þ kkq  ^qk ; ð16Þ
2 j¼1 i¼1
154 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170
f
where zi j is the measured data at the boundary for a given source fj , k is the
regularization parameter, and ^q is an a priori choice, for example could be just
the nominal values for the parameter. Now composing uðqÞ and CuðqÞ we
obtain the parameter-to-output mapping: T ½q
¼ Cu. This is the nonlinear
mapping of diffusion based optical tomography in abstract setting.

2.1. A simple example

In what follows, we illustrate the theoretical concepts with a simple example


in one dimension. Let X ¼ ½0; L
, the diffusion approximation with constant
background is the Strum–Louiville equation:
f
r2 u þ q2 u ¼ ; ð17Þ
D
where q2 ¼ la =D is constant, with the Rubin boundary condition:
uð0Þ  2Druð0Þ ¼ 0;
ð18Þ
uðLÞ þ 2DruðLÞ ¼ 0:
The inverse problem is to estimate the scalar q from the data z measured at
x ¼ 0 or x ¼ L.
For example, the GreenÕs function (solution for a delta distribution source
f ðxÞ ¼ dðx  gÞ where xs ¼ g is the location of the source) for the problem (17)
and (18) can be solved analytically and the solution for x < g is:
ðeqg  ceqg Þðeqx  beqx Þ
uðx; g; qÞ ¼ ; ð19Þ
2qDðb  cÞ
where
e2Lq 1  2Dq
c¼ ; b¼ :
b 1 þ 2Dq
If we measure Dru m at x ¼ 0, then the inverse problem is to estimate q
from the parameter to output map which is given by
ou ðeqg  ceqg Þð1 þ bÞ
T ½q
¼ Cuðx; g; qÞ ¼ D ð0; g; qÞ ¼ ; ð20Þ
om 2ðb  cÞ
which is a nonlinear function of the parameter q as discussed. In Fig. 1, we plot
Jk ðqÞ for the 1D diffusion approximation with Rubin boundary conditions for
a homogeneous background medium with la ¼ 0:012 mm1 and D ¼ 0:33 mm.
This is the simulation offfi the 1D diffusion approximation on the interval
pffiffiffiffiffiffiffiffiffiffi
ð0; 43:0Þ with q ¼ la =D. We computed cost functional Jk ðqÞ ¼ jTq  z1 j2 þ
kkqk2 for q0 ¼ 0:1907 mm1 (corresponding to z1 ¼ Tq0 , x1 ¼ 0 la ¼ 0:012
mm1 , and D ¼ 0:33 mm) over a range of q starting from 0:14 to 0:4. The solid
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 155

–6
x 10
7

cost functional Jλ(q) 6

0
0.15 0.2 0.25 0.3 0.35 0.4
parameter value q

Fig. 1. Parameter estimation and regularization.

curve represents J without regularization (k ¼ 0) and the broken curve rep-


resents Jk with regularization parameter k ¼ 105 . From Fig. 1, it is clear that
without regularization
pffiffiffiffiffiffiffiffiffiffi
ffi ðk ¼ 0Þ the function J is rather insensitive to parameter
q ¼ la =D (i.e. the numerical method starting with an overestimate of the true
parameter is bound to fail). But with regularization ðk ¼ 105 Þ, Jk is more
convex. We note here that the regularization has changed the problem so that
we are solving for a minimum qk that is no longer the same as the problem
without regularization, mainly q0 . This simple example for a constant back-
ground illustrates the complexity of a nonlinear ill-posed inverse problem. In
the rest of the paper we investigate further the one-dimensional inverse prob-
lem in general with spatially varying parameters D and la .

3. System approximation

The general analytic method for solving ODEs or PDEs containing a delta
distribution source is the GreenÕs function method. However for complex
geometries, the analytic solution is intractable. Therefore one requires
numerical solutions. The FEM is more versatile than other methods including
the finite difference method because of its ease in complex geometries and
modelling boundary effects. The FEM is a variational method used to
approximate the solution by a family of finite dimensional basis functions.
Then the forward problem is reduced to one of linear algebra and one com-
putes the approximate solution using FEM codes.
156 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

If we let X ¼ ½0; L
be the 1D domain under consideration with surface
oX ¼ f0; Lg, the weak formulation of the problem corresponding to Eq. (11) is
to find uðxÞ 2 H 1 ðXÞ such that for all vðxÞ 2 H 1 ðXÞ, the following variational
equation is satisfied
Z Z
1 1
rv Dru dX þ la vu dX þ vð0Þuð0Þ þ vðLÞuðLÞ
X X 2 2
Z
¼ vf dX: ð21Þ
X

The FEM is derived by projecting the weak form of (21) onto a finite
dimensional function space. The finite dimensional function space is the set of
continuous and twice differentiable, piecewise cubic polynomials on subinter-
vals of the domain X. We approximate u with the functions
X
N p þ1

uNp ðxÞ ¼ cfi Bi ðxÞ; ð22Þ


i¼1

where
x x 
i
Bi ðxÞ ¼ B
h
with partitions
x1 < x0 ¼ 0 < x1 < < xNp 1 < xNp ¼ L < xNp þ1 ;
and stepsize
h ¼ xi  xi1 for 0 6 i < Np þ 1;
and BðxÞ is the cubic B-spline. By using the test functions v ¼ Bj on the weak
form of the PDE in (21), we obtain the Galerkin approximation for all
1 6 j 6 Np þ 1
Z Z
1 1
rBj DruNp dX þ la Bj uNp dX þ Bj ð0ÞuNp ð0Þ þ Bj ðLÞuNp ðLÞ
X X 2 2
Z
¼ Bj f dX: ð23Þ
X

From the above we obtain a system of equations for cfi where the coefficients
of the system equations are given by
 ij þ Mij þ Pij ;
Aij ¼ D ð24Þ
where the diffusion inner product
Z
 ij ¼
D rBi DrBj dX; ð25Þ
X
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 157

the absorption inner product


Z
Mij ¼ Bi la Bj dX; ð26Þ
X

and the pointwise inner product is


1
Pij ¼ ðBi ð0ÞBj ð0Þ þ Bi ðLÞBj ðLÞÞ: ð27Þ
2
We computed the integrals for D  and M using the Gaussian quadrature.
Therefore we obtain the system
ACf ¼ F; ð28Þ

where the forcing vector is


Z
Fi ¼ Bi f dX;
X

and Cf is the ðNp þ 3Þ-vector containing the unknown coefficients cfi :

Cf ¼ ðcf1 ; cf0 ; . . . ; cfNp þ1 ÞT : ð29Þ

This linear system then can be solved using any one of the number of
techniques.

3.1. Forward simulation

To test our forward solver, we first computed the approximate solutions in


two cases: (1) constant values of D and la , using f ¼ ex (see Fig. 2, and
compare to solution in Fig. 3 where f is shown in Fig. 4 and absolute error is
shown in Fig. 5)

3
computed solution
2.5

1.5

0.5

0
0 10 20 30 40 50

Fig. 2. Constant values of D and la , f ¼ expðxÞ.


158 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

3
exact solution
2.5

1.5

0.5

0
0 10 20 30 40 50

Fig. 3. Constant values of D and la , f ¼ expðxÞ.

0.8 forcing distribution

0.6

0.4

0.2

0
0 10 20 30 40 50

Fig. 4. Constant values of D and la , f ¼ expðxÞ.

–3
x 10
1

0.5
absolute error

–0.5

1
0 10 20 30 40 50

Fig. 5. Constant values of D and la , f ¼ expðxÞ.


T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 159

Computed solution for approx delta distrib source


3

2.5

1.5

0.5

0
0 10 20 30 40 50
jxxs j2
Fig. 6. Constant values of D and la , f ¼ pffiffiffiffiffiffi
A ffi
2p2
e 22 .

A jxx j2
 s
f ¼ pffiffiffiffiffiffiffiffiffi e 22 ; ð30Þ
2p2

with A ¼ 1 and  ¼ 0:1 where xs ¼ 42:5 is the source location (see Fig. 8). Here
this particular f is an approximation to the dðx  xs Þ source for which the
GreenÕs function solution is available (see Fig. 6 for approximate solution and
compare it to analytic solution in Fig. 7. The absolute error is given in Fig. 9)
Fig. 13 illustrates numerical simulation with nonconstant D and la (see Figs. 10
and 11), and f as in (30) with A ¼ 1,  ¼ 0:1, and xs ¼ 42:5 (Fig. 12).

4. Inversion using iterative regularized Gauss–Newton algorithm

In the last section, we outlined the numerical method for the forward
problem. In this section, we will describe the computational method for the

Exact solution for delta distributed source


3

2.5

1.5

0.5

0
0 10 20 30 40 50
jxxs j2
Fig. 7. Constant values of D and la , f ¼ pffiffiffiffiffiffi
A ffi
2p2
e 22 .
160 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

Approximated delta distributed source


3.5

2.5

1.5

0.5

0
0 10 20 30 40 50
jxxs j2
Fig. 8. Constant values of D and la , f ¼ pffiffiffiffiffiffi
A ffi
e 22 .
2p2

Absolute error
0.05

0.05

0.1

0.15
0 10 20 30 40 50
jxxs j2
Fig. 9. Constant values of D and la , f ¼ pffiffiffiffiffiffi
A ffi
2p2
e 22 .

The coefficient of absorption mua


–3
x 10
12

11

10

6
0 10 20 30 40 50
jxxs j2
A ffi 
Fig. 10. Nonconstant values of D and la , f ¼ pffiffiffiffiffiffi
2p2
e 22 .
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 161

The coefficient of diffusion D


0.6

0.5

0.4

0.3

0 10 20 30 40 50
jxxs j2
A ffi 
Fig. 11. Nonconstant values of D and la , f ¼ pffiffiffiffiffiffi
2p2
e 22 .

Approximated delta distributed source f


3.5

2.5

1.5

0.5

0
0 10 20 30 40 50
jxxs j2
A ffi 
Fig. 12. Nonconstant values of D and la , f ¼ pffiffiffiffiffiffi
2p2
e 22 .

Computed solution for variable D and mu a


2.5

1.5

0.5

0
0 10 20 30 40 50
jxxs j2
A ffi 
Fig. 13. Nonconstant values of D and la , f ¼ pffiffiffiffiffiffi
2p2
e 22 .
162 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

inverse problem. We approximate the infinite dimensional parameters D and la


with the functions
X
N
DðxÞ ¼ ^ k ðxÞ;
dk B ð31Þ
k¼1

X
N
la ðxÞ ¼ ^ k ðxÞ;
ek B ð32Þ
k¼1

where B^ k are the basis functions may be on a different set of grid points than
the grid points for the basis functions Bk . The model parameter vector q is
approximated by
T
qN ¼ ðd1 ; . . . ; dN ; e1 ; . . . ; eN Þ ; ð33Þ
where qN is to be identified from the minimization of the cost functional Jk
m

2
  1X ms X

fj Np  N  f
2
Jk qN ¼
Ci u q  zi j
þ kkqN  ^qN k ; ð34Þ
2 j¼1 i¼1

where T ½qN
¼ CuNp ðqN Þ and k is the regularization parameter.

4.1. Frechet derivative

Any numerical method for solving a nonlinear inverse problem requires


computation of the Frechet derivative and it is adjoint via linearization. T ½q

associates with q ¼ ðD; la ÞT say the detector response of the form


ou
T ½q
¼ CuðqÞ ¼ D ; ð35Þ
om
where u is the solution to Eq. (11) with the forcing distribution f . The inverse
problem is to determine D; la from a set of
f f
Ci j uðqÞ ¼ zi j ð36Þ
detector outputs for a set of sources fj , j ¼ 1; . . . ; ms .
Therefore we derive the linearization of T ½q
. Let us assume that approxi-
mations D0 and l0 for D and la is available such that D ¼ D0 þ h1 , la ¼ l0 þ h2
with h1 and h2 small, and for simplicity, D ¼ D0 near oX. Let v0 be the solution
of
r D0 ðxÞrv0 ðxÞ þ l0 ðxÞv0 ðxÞ ¼ f0 ðxÞ ð37Þ
with boundary condition
ov0
v0 ðxÞ þ 2D0 ðxÞ ðxÞ ¼ 0; x 2 oX: ð38Þ
om
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 163

Let u ¼ v0 þ v where u solves Eqs. (11) and (12) and subtracting Eqs. (37)
and (38) from Eqs. (11) and (12), we obtain ignoring second-order terms in h1 ,
h2 , and v

r D0 ðxÞrvðxÞ þ l0 ðxÞvðxÞ ¼ r h1 ðxÞrv0 ðxÞ  h2 ðxÞv0 ðxÞ ð39Þ

with boundary condition


ov
vðxÞ þ 2DðxÞ ðxÞ ¼ 0; x 2 oX: ð40Þ
om
The linearization T 0 ½q
is given by
 
h ov
T ½q
1 ¼ D ;
0
ð41Þ
h2 om

where v is the solution to Eq. (39). The adjoint of this linearized equation (39)
can be derived using the GreenÕs formula
Z 
ð  r Drv þ la vÞw  vðr Drw þ la w Þ dX ð42Þ
X
Z !
ov  
ow
¼ D  v
w ds ð43Þ
oX om om

and the adjoint of T 0 ½q


is
    
h rv0 rw
T 0 ½q
1 r ¼ ; ð44Þ
h2 v0 w
where w is a solution of the adjoint equation
r DðxÞrw ðxÞ þ la ðxÞw ðxÞ ¼ 0 ð45Þ
with the boundary condition
ow
w ðxÞ þ 2DðxÞ ðxÞ ¼ r ; x 2 oX: ð46Þ
om

4.2. Discrete sensitivity relations

Now we will derive the discrete sensitivity relations which will be used for
the Jacobian calculation for our numerical scheme below. If we plug in the
expression for DðxÞ and lðxÞ Eqs. (31) and (32) into the equations for the
system matrices D and M Eqs. (25) and (26) respectively, we get
X
N
 ij ¼
D  k;ij ;
dk D ð47Þ
k¼1
164 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

where
Z
 k;ij ¼
D ^ k rBi rBj dX;
B ð48Þ
X

and similarly
X
N
Mij ¼ ek Mk;ij ; ð49Þ
k¼1

where
Z
Mk;ij ¼ ^ k Bi Bj dX:
B ð50Þ
X

Now if we differentiate Eq. (28) with respect to the parameters qk and if we


assume F is independent of qk we get
oCf oA
A þ Cf ¼ 0; ð51Þ
oqk oqk
and if we solve for the partial derivatives with respect to the parameters qk
oCf  þ M þ P Þ1 Dk Cf ;
¼ ðD 1 6 k 6 N; ð52Þ
oqk
oCf  þ M þ P Þ1 Mk Cf ;
¼ ðD N þ 1 6 k 6 2N ; ð53Þ
oqk
where A ¼ D  þ M þ P, D  k and Mk are the ðNp þ 3Þ  ðNp þ 3Þ matrices with
 k;ij and Mk;ij respectively.
elements D

4.3. Numerical scheme

Now we describe our numerical scheme for the inverse problem. The inverse
problem in Eq. (34) can be casted as a finite dimensional nonlinear minimi-
zation problem mainly
m

2 1  
  1X ms X

fj Np  N  f
2 1Xmtot
 
min J qN ¼
Ci u q  zi j
¼ kR qN k ¼ ri2 qN
N
q 2R 2N 2 j¼1 i¼1 2 2 i¼1
T
with zero or nonzero residual, where, RðqN Þ ¼ ðr1 ðqN Þ; r2 ðqN Þ; . . . ; rmtot ðqN ÞÞ
f f
with mtot ¼ m  ms , rn ðqN Þ ¼ Ci j uNp ðqN Þ  zi j with n ¼ i þ ðj  1Þm, and the
N
norm kRk is called the residual at q . Let H  ðqN Þ ¼ K  ðqN ÞK ðqN Þ, where K ðqN Þ
N
is the Jacobian matrix of Rðq Þ,
 
  orn ðqN Þ
K qN ¼ ; n ¼ 1; . . . ; mtot ; k ¼ 1; . . . ; N : ð54Þ
oqk
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 165

The gradient is given by


     
rJ qN ¼ K  qN R qN ; ð55Þ
and the Hessian is given by
     
r2 J qN ¼ H  qN þ A qN ; ð56Þ
P m
where AðqN Þ ¼ i¼1 tot
ri r2 ri ðqN Þ. Thus, if the Jacobian matrix K ðqN Þ is avail-
able, we know the first part H  of the Hessian r2 J ðqN Þ without requiring any
second-order information.
For example, if we use an observation operator as in Eq. (15) we get
 Np 
orn ðqN Þ o ou N 1 ouNp
¼ D ðxi ; q ; fj Þ ¼ ðxi ; qN ; fj Þ: ð57Þ
oqk oqk om 2 oqk
If we now plug the expression for uNp from Eq. (22) into Eq. (57) we get
Np þ1
orn ðqN Þ 1 X
f
ocl j  T oCfj ;
¼ Bl ðxi Þ ¼ B ð58Þ
oqk 2 l¼1 oqk oqk

where
 
 ¼ 1 B1 ðxi Þ; B0 ðxi Þ; . . . ; BN ðxi Þ; BN þ1 ðxi Þ T ;
B ð59Þ
2 p p

and Cfj is given by Eq. (29). Therefore using the sensitivity equations (52) and
(53) for the expression oCfj =oqk in Eq. (58) we can compute the Jacobian
matrix KqN for our inverse calculation. One can also use the adjoint formu-
lation to derive this Jacobian matrix which can be beneficial in two and three
dimensions. However in one dimension speed of the forward sensitivity rela-
tions is adequate to compute the Jacobian of the inverse problem.
Given the Jacobian of the inverse problem, we can define the IRGN method
as follows:
        
qNkþ1 ¼ qNk  ck ½K  qNk K qNk þ kk I
1 K  qNk R qNk þ kk ðqNk  ^qN Þ ;

where kk is a regularizing sequence that satisfies the conditions:


kk
kk > 0; 16 6 r; lim kk ¼ 0:
kkþ1 k!1

The reader may consult [1,4,16] for the convergence analysis of IRGN under
different types of conditions on the initial guess q0 . The best convergence rate
was achieved when we used the iteratively regularized analog of a line search
procedure suggested in [22] for the well-posed case. Namely, a backtracking
strategy with ck ¼ 1; 1=2; 1=4, until the following two requirements were
simultaneously fulfilled:
Jkk ðqkþ1 Þ 6 Jkk ðqk Þ þ dchrJkk ðqk Þ; sk i ð60Þ
166 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

which is the Armijo–Goldstein strategy and


krJkk ðqkþ1 Þk 6 qkrJkk ðqk Þk ð61Þ

which is the Wolfe type strategy. Here sk is the solution to ðK  ðqk ÞKðqk Þ þ
kk IÞsk ¼ ðK  ðqk ÞRðqk Þ þ kk qk Þ. Our code uses the parameters d ¼ 0:0001 and
q ¼ 0:99 as in [13].

4.4. Simulation results

The optimization of the functional Jk was performed using the IRGN iter-
ative method described above. After 20 iterations using equally spaced grid
points with Np ¼ 40 and N ¼ 20 with k0 ¼ 1:0E  6 with the choice of cubic B-
splines satisfying Rubin boundary condition for the solution and Neumann
boundary condition for the parameter, we reconstructed the D profile for a set
of 10 sources. We modelled the sources as in Eq. (30) with xs ¼ 0:1, 0.45, 0.8,
1.15, 1.5, 41.0, 41.3750, 41.75, 41.1250, 42.5. The reconstructed profile is de-
picted in Fig. 14. The line corresponds to true D, the circles correspond to the
optimal profile for the space of N -dimensional cubic B-spline basis functions
for D, and the stars correspond to the constructed D using IRGN.
Since the problem is exponentially ill-posed the proper choice of the initial
value of kk was extremely important for the numerical simulations. In Figs. 15–
18 the dependence of the accuracy of the reconstructed solutions on the choice
of k0 is illustrated. For a posteriori stopping of IRGN iterations the generalized
discrepancy principle (see [4]) was implemented, i.e. the iterations were stopped
at the first index ^k, for which the residual kRðq^Nk Þk is less than or equal to sr,
s > 1:

1.2
1.1
1
0.9
0.8
0.7
0.6
0.5 exaact D
0.4 ****** computed D
oooo optimal D
initial D
0.3
0.2
0 10 20 30 40 50

Fig. 14. lk ¼ 106 /k.


T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 167

1.2

0.8

D
0.6

exact D
0.4 computed D
initial D
0.2
0 10 20 30 40 50

Fig. 15. lk ¼ 103 /k.

1.2

0.8
D

0.6

exact D
0.4 computed D
initial D
0.2
0 10 20 30 40 50

Fig. 16. lk ¼ 104 /k.

1.2

0.8
D

0.6

exact D
0.4 computed D
initial D
0.2
0 10 20 30 40 50

Fig. 17. lk ¼ 105 /k.

kRðq^Nk Þk2 6 sr < kRðqNk Þk2 ; 0 6 k < ^k; ð62Þ

where r is the error of the simulated data in L2 -norm.


168 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

1.2

0.8

D
0.6

exact D
0.4 computed D
initial D
0.2
0 10 20 30 40 50

Fig. 18. lk ¼ 106 /k.

Table 1
Effect of regularization parameter
k0 IRGN iteration Inner iterations Relative error Residual
1.0E)1 10 56 0.12677 7.5060E)4
1.0E)2 10 56 0.11652 4.2413E)4
1.0E)3 10 56 0.09710 2.3323E)4
1.0E)4 10 36 0.08179 1.10069E)4
1.0E)5 4 11 0.07483 3.5921E)5
1.0E)6 6 28 0.07024 4.5446E)5
1.0E)7 1 1 div div

In Table 1 one can see the summary of measured accuracy and work done
for values of k0 in the range from 1.0E)4 to 1.0E)7. For every k0 IRGN
approximations were calculated until either (62) was satisfied with sr ¼ 1:0E)4
or the number of iterations k became greater than 10. One can notice that the
required accuracy of the residual was attained for two values of k0 :1.0E)5 and
1.0E)6. For k0 ¼ 1:0E)4 the residual is very close to 1.0E)4 but still greater.
For other regularizing sequences kk , for example kk ¼ k0 =k n , 0:5 6 n, and
kk ¼ k0 = expðkÞ the optimal values of the parameters can also be found, how-
ever when kk ¼ k0 =k the accuracy of reconstruction is higher.

5. Conclusion

In this paper, we investigated an one-dimensional inverse problem in dif-


fusion based optical tomography. We devised a stable reconstruction algorithm
for the inverse problem using IRGN with Armijo–Goldstein–Wolf (AGW)
type line search. The algorithm worked very well for the ill-posed inverse
problem, for example in comparison to the MATLAB Levenberg–Marquardt
trust region nonlinear least square routine LSQNONLIN [12,19,21] with
T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170 169

1.2 1.2
1.1 1.1
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 exact D 0.5 exact D
0.4 ****** computed D 0.4 ****** computed D
oooo optimal D oooo optimal D
0.3 initial D 0.3 initial D
0.2 0.2
(a) 0 10 20 30 40 50 (b) 0 10 20 30 40 50

Fig. 19. D reconstructed using (a) IRGN with AGW algorithm, (b) Levenberg–Marquardt trust
region algorithm.

Np ¼ 40 and N ¼ 20, the reconstruction D profile is depicted in Fig. 19(b). The


line corresponds to true D, circles correspond to the optimal D, and stars
correspond to the constructed D. The comparison of Fig. 19(a) and (b) dem-
onstrates that the Levenberg–Marquardt trust region did not converge to the
optimal solution where IRGN/AGW did converge accurately. Therefore, we
have demonstrated the efficacy of our algorithm with accurate reconstruction
of the scattering parameter relevant to optical tomography. We are currently
investigating theoretical convergence and implementation of the proposed
IRGN–AGW algorithm in 2D and 3D cases with clinical data for early
detection of breast cancer at the Biomedical Optics Laboratory at Clemson
University.

References

[1] A.B. Bakushinsky, Iterative methods for nonlinear operator equations without regularity, new
approach, Dokl. Russian Acad. Sci. 330 (1993) 282–284.
[2] S.R. Arridge, Optical tomography in medical imaging: topical review, Inverse Problems 15
(1999) R41–R93.
[3] S.R. Arridge, J.C. Hebden, Optical imaging in medicine: 2. Modelling and reconstruction,
Phys. Med. Biol. 42 (1997) 841–853.
[4] B. Blaschke, A. Neubauer, O. Scherzer, On convergence rates for the iteratively regularized
Gauss–Newton method, IMA J. Num. Anal. 17 (1997) 421–436.
[5] H. Bremmer, Random volume scattering, Radiat. Sci. J. Res. 680 (1964) 967–981.
[6] B. Chance, R.R. Alfano (Eds.), Photon migration and imaging in random media, and tissues,
vol. 1888, SPIE, 1993.
[7] B. Chance, R.R. Alfano (Eds.), Optical Tomography, Photon migration, and spectroscopy of
tissue and model media: theory, human studies, and instrumentations, Part 1 and 2, vol. 2389,
SPIE, 1995.
[8] B. Chance, C.E. Cooper, D.T. Delpy, E.O.R. Reynolds, Near-infrared spectroscopy and
imaging of living systems, Phil. Trans. R. Soc. Lond. B 352 (1997) 643–648.
[9] R. Chandrasekhar, Radiation Transfer, Oxford, Clarendon, 1950.
[10] National Research Council, Mathematics and Physics of Emerging Biomedical Imaging,
National Academy Press, Washington DC, 1996.
170 T. Khan, A. Smirnova / Appl. Math. Comput. 161 (2005) 149–170

[11] D.T. Delphy, M. Cope, Quantification in tissue near-infrared spectroscopy, Phil. Trans. R.
Soc. Lond. B 352 (1997) 649–659.
[12] J.E. Dennis, R.B. Schnabel, Numerical Methods for Unconstrained Optimization and
Nonlinear Equation, Prentice-Hall, Englewood Cliffs, NJ, 1983.
[13] M. Hanke, J. Nagy, C. Vogel, Quasi-Newton approach to nonnegative image restorations,
Linear Algebra Appl. 316 (2000) 223–236.
[14] J. Hebden, S.R. Arridge, D.T. Delphy, Optical imaging in medicine: 1. Experimental
techniques, Phys. Med. Biol. 42 (1997) 825–840.
[15] A. Hielscher, R. Alcouffe, R. Barbour, Comparison of finite-difference transport and diffusion
calculations for photon migration in homogenous and heterogenous tissues, Phys. Med. Biol.
42 (1998) 1285–1302.
[16] T. Hohage, Logarithmic convergence rates of the iteratively regularized Gauss–Newton
method for an inverse potential and inverse scattering problem, Inverse Prob. 13 (1997) 1279–
1299.
[17] A. Ishimaru, Single Scattering and Transport, Theory (Wave Propagation and Scattering in
Random Media I), Academic, New York, 1978.
[18] M. Klibanov, T.R. Lucas, Numerical solution of a parabolic inverse problem in optical
tomography using experimental data, SIAM J. Appl. Math. 59 (5) (1999) 1763–1789.
[19] K. Levenberg, A method for the solution of certain problems in least squares, Quart. Appl.
Math. 2 (2) (1944) 164–168.
[20] H.W. Lewis, Multiple scattering in an infinite medium, Phys. Rev. 78 (1950) 526–529.
[21] D. Marquardt, An algorithm for least, squares estimation of nonlinear parameters, SIAM J.
Appl. Math. 11 (2) (1963) 431–441.
[22] S.G. Nash, A. Sofer, Linear and Nonlinear Programming, McGraw-Hill, New York, 1996.
[23] F. Natterer, F. W€ ubbeling, Mathematical methods in image reconstruction, SIAM Mono-
graphs on Mathematical Modeling and Computation, 2001.

You might also like