Sie sind auf Seite 1von 13

The finite element model for the propagation of light in scattering media:

A direct method for domains with nonscattering regions


Simon R. Arridge
Department of Computer Science, University College London, Gower Street, London, WC1E 6BT, England
Hamid Dehghani
Department of Medical Physics, University College London, Capper Street, London, WC1E 6JA, England
Martin Schweiger
Department of Computer Science, University College London, Gower Street, London, WC1E 6BT, England
Eiji Okada
Department of Electronics and Electrical Engineering, Keio University, 3-14-1, Hiyoshi, Kohoku-ku,
Yokohama, 223-8522, Japan
共Received 4 May 1999; accepted for publication 5 October 1999兲
We present a method for handling nonscattering regions within diffusing domains. The method
develops from an iterative radiosity-diffusion approach using Green’s functions that was computa-
tionally slow. Here we present an improved implementation using a finite element method 共FEM兲
that is direct. The fundamental idea is to introduce extra equations into the standard diffusion FEM
to represent nondiffusive light propagation across a nonscattering region. By appropriate mesh node
ordering the computational time is not much greater than for diffusion alone. We compare results
from this method with those from a discrete ordinate transport code, and with Monte Carlo calcu-
lations. The agreement is very good, and, in addition, our scheme allows us to easily model
time-dependent and frequency domain problems. © 2000 American Association of Physicists in
Medicine. 关S0094-2405共00兲00901-9兴

Key words: light propagation, diffusion, voids, transport equation, finite element method

I. INTRODUCTION In a previous paper8 we introduced an alternative ap-


proach that we termed radiosity-diffusion, which was specifi-
The finite element method 共FEM兲 has become well estab- cally designed for the zero-scattering case. This method is
lished in biomedical optics. We introduced this method in effectively a hybrid one, modeling the problem as a union of
Refs. 1–3, which we refer to as I–III, where we developed diffusive regions wherein the DA applies, together with clear
the method for solving the steady-state, time-dependent, and regions where light propagates rectilinearly. Comparisons
frequency-domain versions of the diffusion approximation were made therein between this model, Monte Carlo, and
共DA兲, itself derived from the more accurate transport equa- experimental data. As previously reported, the method was
tion. Recent years have seen an uptake in this approach by based on analytical Green’s function descriptions and was
other workers.4–6 limited to simple geometries. In this paper we present a
Despite the success of this method in modeling optically method for incorporating the radiosity-diffusion concept into
thick regions, it is well known that under certain conditions, our FEM model, which allows us to exploit the geometric
it is no longer valid. In particular, the presence of nonscat- flexibility of the latter. Furthermore, our previous approach
tering 共void兲 regions, such as occur in the cerebrospinal fluid was iterative, and quite slow. In this paper we show that the
共CSF兲 filled ventricles in the brain, represent a situation for method can be solved using direct methods that are much
which it is clearly inadequate. Under these circumstances, faster.
more advanced methods are required. It is tempting to utilize The remainder of the paper is organized as follows. In
methods for solving the transport equation, usually described Sec. II we briefly review the equations of light propagation
by either the Boltzmann equation, or its equivalent, the ra- in tissue and the available methods for their solution. In Sec.
diative transfer equation 共RTE兲. A very large body of litera- III we define the ‘‘void problem’’ and previous approaches
ture exists for solving this equation, and some studies have to it, as well as summarizing the radiosity-diffusion ap-
been reported using numerical methods for the Boltzmann proach. In Sec. IV we summarize the FEM approach and
equation to investigate both the very low scattering case, and show how to incorporate the radiosity method, for the
the very high absorption case in tissue, for which the DA is steady-state case. In Sec. V we explain aspects of the imple-
not valid.7 However, for the case of a ‘‘true’’ void 共i.e., one mentation, and in Sec. VI we show how to extend the analy-
with genuinely zero scattering兲 even these methods have sis to the frequency-domain, and time-domain problems. Re-
some difficulties. sults are presented in Sec. VII for the case of a nonscattering

252 Med. Phys. 27 „1…, January 2000 0094-2405/2000/27„1…/252/13/$17.00 © 2000 Am. Assoc. Phys. Med. 252
253 Arridge et al.: Domains with nonscattering regions 253

gap separating two diffusing regions as well as a nonscatter- B. Spherical harmonic expansion
ing ‘‘hole’’ 共such as a ventricle in the brain兲. We compare A well-known approach to solving the Boltzmann equa-
both to Monte Carlo results and to a production transport tion is to expand the angular variable in spherical harmonics.
code. In Sec. VIII we present some conclusions and make We express the quantities in 共1兲 as
suggestions for further improvements to the model. In Ap-

冉 冊
⬁ l 1/2
pendix A we review the ‘‘classical’’ radiosity theory and 2l⫹1
explain the differences with the radiosity-diffusion approach. ␾ 共 r,ŝ兲 ⫽ 兺 兺 ␺ l,m 共 r兲 Y l,m 共 ŝ兲 , 共5兲
l m⫽⫺l 4␲
Finally, in Appendix B we demonstrate the equivalence of

冉 冊
⬁ l 1/2
the direct method presented here to the previous iterative 2l⫹1
approach. q 共 r,ŝ兲 ⫽ 兺l m⫽⫺l
兺 4␲
q l,m 共 r兲 Y l,m 共 ŝ兲 , 共6兲

where Y l,m (ŝ) is a spherical harmonic of order l degree m,


and the normalization factor „(2l⫹1)/4␲ …1/2 is introduced
II. LIGHT TRANSPORT MODELS for convenience. If the phase function is assumed to be in-
A general model of light transport in scattering media, but dependent of the explicit angle ŝ and is written ⌰共ŝ–ŝ⬘兲, then
one that ignores polarization and coherence effects, is the 共1兲 can be expressed as an infinite set of coupled first-order
Boltzmann equation. This equation has been extensively equations. When these are truncated by assuming ␺ l,m ⫽0;
studied in the field of neutron transport9–13 and in radiation l⬎N for some N, the result is a set of (N⫹1) 2 共in three
transfer,14,15 where it is known as the radiative transfer equa- dimensions兲 first-order equations known as the P N
tion 共RTE兲. Here we very briefly review the properties of this approximation.9,11,16
equation, and outline, in general, the solution strategies that The approximations cannot satisfy the condition 共4兲 ex-
have been developed. actly. Instead, boundary conditions associated with this form
are commonly either of the Marshak,


A. The Boltzmann equation
1
We will discuss only the single-group Boltzmann equa- ␾ 共 m,ŝ兲共 ŝ–␯ˆ 兲 2 j⫺1 dŝ⫽0, j⫽1,2,..., 共 N⫹1 兲 , 共7兲
ˆs–ˆ␯⬍0 2
tion, which in the steady state is written as
or the Mark type,
„ŝ–“⫹ ␮ t 共 r兲 …␾ 共 r,ŝ兲 ⫽ ␮ s 共 r兲 冕 S n⫺1
⌰ 共 ŝ,ŝ⬘ 兲 ␾ 共 r,ŝ⬘ 兲 dŝ⬘
冕 ˆs–ˆ␯⬍0
␾ 共 m,ŝ j 兲 dŝ⫽0, 共8兲
⫹q 共 r,ŝ兲 . 共1兲
for a set of defined directions 兵 ŝ j ; j⫽1,2,..., 21 (N⫹1) 其 .
Here ␮ t (r)⫽ ␮ s (r)⫹ ␮ a (r) 共units of inverse length兲 is the The P N approximation can be further manipulated to pro-
attenuation coefficient at position r, with ␮ s (r) the scattering duce, for example, either a single (N⫹1)th-order equation,
coefficient and ␮ a (r) the absorption coefficient. ␾共r,ŝ兲 共units or a set of N! 共in three dimensions兲 coupled second-order
of inverse length cubed per steradian兲 is the number of pho- equations. The boundary conditions of these forms become
tons per unit volume at position r with velocity in angular more complex.9,12,13
direction ŝ, with q(r,ŝ) the number of source photons.
⌰共ŝ,ŝ⬘兲 is the normalized phase function representing the C. Discrete ordinates method
probability of scattering from direction ŝ⬘ to direction ŝ.
Two derived quantities that are of interest are the photon The Discrete ordinate method combines spatial discretiza-
density: tion on a mesh of discrete points 兵 ri ;i⫽1, . . . ,D 其 , and a set
of discrete directions 兵 ŝ j ; j⫽1, . . . ,M 其 , together with a cor-
⌽ 共 r兲 ⫽ 冕 S n⫺1
␾ 共 r,ŝ兲 dŝ; 共2兲
responding set of quadrature weights 兵 w j ; j⫽1, . . . ,M 其 .
The quadrature rule is used to evaluate the integral term in
共1兲, resulting in a matrix relation,
and the photon current:
T␾⫽S␾⫹q, 共9兲
J共 r兲 ⫽ 冕S n⫺1
ŝ␾ 共 r,ŝ兲 dŝ. 共3兲 where T represents the discretized streaming and removal
terms ŝ j –“⫹ ␮ t , and S represents the integral over scatter-
Several boundary conditions apply to the Boltzmann ing from other directions. This can be solved using an itera-
equation. At a bare surface, the vacuum boundary condition tive scheme.12 An acceleration was developed by Alcouffe
specifies that no photons travel in an inward direction at the using the approximate solution for direction j to generate
boundary, except for source terms modified sources for a diffusion equation for direction j
⫹1, cycled repeatedly until convergence.17
␾ 共 m,ŝ兲 ⫽0, for ŝ–␯ˆ ⬍0, 共4兲 Considerable attention has been paid to the choice of
quadrature scheme for the discrete ordinate method, and
where we use m to define a point on the domain boundary, to the specification of boundary conditions. Under the Wick–
and ␯ˆ to be the local outward directed normal. Chandrasekhar scheme, wherein Gaussian quadrature is em-

Medical Physics, Vol. 27, No. 1, January 2000


254 Arridge et al.: Domains with nonscattering regions 254

ployed, the S N equations are formally equivalent to the P N⫺1 cretized in a suitable FEM basis to give a matrix equation
equations with Mark boundary conditions.12 A well-known containing symmetric positive-definite matrices. The advan-
drawback of the discrete ordinate method is the ‘‘ray effect,’’ tage of this approach is that the matrix equation is both
which results from what is essentially the decomposition of smaller and easier to solve, although the boundary conditions
the original equation into a set of fluxes in specified direc- are less easy to express.
tions, and manifests itself in the method ‘‘missing’’ small-
scale features of the problem domain.18,19 This effect is par-
ticularly serious in the presence of void regions, as will be E. Diffusion approximation
demonstrated in Sec. VII.
The diffusion approximation 共DA兲 is the simplest non-
D. Finite element methods trivial approximation that results from considering the
second-order form of the P N approximation for N⫽1. We
In the finite element method 共FEM兲 for the Boltzmann
will make the following definitions:
equation, the angular flux is expressed in a finite-dimensional
space ␹ h
K
isotropic source: q 0 ⫽ 冕S n⫺1
q 共 r,ŝ兲 dŝ, 共18兲
␾ 共 r,ŝ兲 ⯝ ␾ h 共 r,ŝ兲 ⫽ 兺 ␾ k f k 共 r,ŝ兲 , 共10兲
k⫽1 reduced scattering coefficient: ␮ s⬘ ⫽ 共 1⫺⌰ 1 兲 ␮ s , 共19兲
where 兵 f k ;k⫽1,...,K 其 are a set of basis functions for ␹ h . 1
diffusion coefficient: ␬ ⫽ , 共20兲
When applied to the first-order form of the Boltzmann equa- 3 共 ␮ a ⫹ ␮ s⬘ 兲
tion all boundary conditions are natural and the resultant
discretized system is again expressed as a matrix relation,20 whence the DA is given by
M␾⫽q. 共11兲 ⫺“–␬ 共 r兲 “⌽ 共 r兲 ⫹ ␮ a 共 r兲 ⌽ 共 r兲 ⫽q 0 共 r兲 , 共21兲
Often it is assumed that the basis is developed separately
for the spatial and angular terms, and the measureable on the boundary 共the exitance, or inten-
sity兲 is given by
f k 共 r,ŝ兲 ⫽u i 共 r兲 ␪ j 共 ŝ兲 , 共12兲
⳵ ⌽ 共 m兲
⌫ 共 m兲 ⫽⫺ ␬ 共 m兲 ⫽ ␯ˆ 共 m兲 –J共 m兲 . 共22兲
which simplifies the derivation of the matrix elements M kk ⬘ . ⳵␯
The spatial basis is usually taken piecewise polynomial, as
The boundary condition resulting from the zeroth-order
can be the angular basis. However, if the latter is taken to be
Marshak condition can be expressed as a Robin condition:
the spherical harmonics, then the system is directly a FEM
representation of the P N approximations. ⳵ ⌽ 共 m兲
⌽ 共 m兲 ⫹2 ␣␬ 共 m兲 ⫽0. 共23兲
By taking the coupled second-order equations of the P N ⳵␯
approximation for odd N, a set of coupled diffusion-like
equations is developed with self-adjoint operators that are The value of ␣ is 1 for a matched refractive index medium,
expressed as symmetric positive-definite matrices.21,22 Alter- but otherwise depends on the refractive index mismatch at
natively, by considering the sum and difference of the angu- the boundary and is discussed in detail in II.2 More recent
lar flux in directions ŝ and ⫺ŝ, the even ␾ ⫹ and odd ␾ ⫺ studies on boundary conditions are presented in Refs. 23, 24.
partity terms can be used to define two coupled transport The DA is by far the most extensively used approxima-
equations: tion in biomedical optics. Analytical and numerical solution
ŝ–“ ␾ ⫹ 共 r,ŝ兲 ⫹ ␮ t 共 r兲 ␾ ⫺ 共 r,ŝ兲 schemes are numerous. Our approach has been based on
the finite element method 共FEM兲, which is summarized in
⫽ ␮ s 共 r兲 冕 S n⫺1
⌰ ⫺ 共 ŝ,ŝ⬘ 兲 ␾ ⫺ 共 r,ŝ⬘ 兲 dŝ⬘ ⫹q ⫺ 共 r,ŝ兲 , 共13兲
Sec. IV A.

ŝ–“ ␾ ⫺ 共 r,ŝ兲 ⫹ ␮ t 共 r兲 ␾ ⫹ 共 r,ŝ兲


III. THE VOID PROBLEM
⫽ ␮ s 共 r兲 冕 S n⫺1
⫹ ⫹ ⫹
⌰ 共 ŝ,ŝ⬘ 兲 ␾ 共 r,ŝ⬘ 兲 dŝ⬘ ⫹q 共 r,ŝ兲 , 共14兲 The problem of voids has been the subject of considerable
research in the particle transport literature.25–28 A good sum-
where mary is given by Ackroyd,13 Chap. 10. Methods have fo-
␾ ⫾ 共 r,ŝ兲 ⫽ 21 关 ␾ 共 r,ŝ兲 ⫾ ␾ 共 r,⫺ŝ兲兴 , 共15兲 cused either on using a very large number of angular terms
共discrete ordinates,19 or spherical harmonics25兲, or on using

q 共 r,ŝ兲 ⫽ 关 q 共 r,ŝ兲 ⫾q 共 r,⫺ŝ兲兴 ,
1
2 共16兲 hybrid techniques wherein the void is treated with special
approximating functions,26 or with Monte Carlo
⌰ ⫾ 共 ŝ–ŝ⬘ 兲 ⫽ 21 关 ⌰ 共 ŝ–ŝ⬘ 兲 ⫾⌰ 共 ⫺ŝ–ŝ⬘ 兲兴 . 共17兲
techniques.29 Our approach can be considered a hybrid one,
Elimination of ␾ ⫺ from 共13兲 results in the second-order but to the best of our knowledge has not been presented
even-parity Boltzmann equation, which again can be dis- before.

Medical Physics, Vol. 27, No. 1, January 2000


255 Arridge et al.: Domains with nonscattering regions 255

m⫺m⬘
d m,m⬘ ⫽ 兩 m⫺m⬘ 兩 , ŝm,m⬘ ⫽ .
兩 m⫺m⬘ 兩
As derived in Ref. 8; the contribution to the inward di-
rected photon distribution at m due to the photon current
J共m⬘兲 at m⬘, is
ŝm,m⬘ –J共m⬘ 兲 ŝm,m⬘ –␯ˆ 共m兲
␾ inc共 m,ŝm,m⬘ 兲 ⫽h m,m⬘ 2 e ⫺ ␮ a 共 ⍀̄ 兲 d m,m⬘ ,
␲ d m,m⬘
共24兲

where h m,m⬘ is a binary flag, that is one if m,m⬘ are mutually


FIG. 1. A domain ⍀ containing diffusing regions 兵 ⍀ i 其 and voids 兵 ⍀̄ j 其 . visible, and zero if they are not.
Using the definition of exitance in 共22兲, we define the
relation
A. Definition ŝ–J⫽ŝ–␯ˆ ⌫. 共25兲
Consider a problem such as shown in Fig. 1. We have a Under the assumption of the boundary condition 共23兲, we
domain ⍀ consisting of R diffusing regions 兵 ⍀ 1 ,⍀ 2 ,...,⍀ R 其 have
and V void regions 兵 ⍀̄1 ,⍀̄2 ,...,⍀̄V 其 . Let ⍀ d ⫽艛 i⫽1
R
⍀ i be
⌫ 共 m兲 ⫽ ␨ ⌽ 共 m兲 , 共26兲
the union of all diffusing regions and ⍀̄ ⫽艛 i⫽1 ⍀̄i be the
d V

union of all void regions. Thus ⍀⫽⍀ d 艛⍀̄d . Each region


where ␨⫽1/2␣. Note that this constant depends on the refrac-
has an outer boundary and any number of inner boundaries,
tive index mismatch between media and is different for a
whose union can be empty. In the case where the inner
tissue–CSF interface than for a tissue–air interface. We in-
boundary is not empty, we define ⳵ ⍀ i ⫽ ⳵ ⍀ ⫺ ⫹
i 艛 ⳵ ⍀ i for dif- terpret the incoming distribution ␾ inc as an internal source.

fusing regions, with similarly ⳵ ⍀̄i ⫽ ⳵ ⍀̄i 艛 ⳵ ⍀̄i for voids, Under the diffusion approximation, this source is represented
which are the ‘‘inner’’ and ‘‘outer’’ boundaries in the sense as an isotropic distribution whose form is obtained by inte-
of a line from the outside of ⍀ to the center. In the case grating over the angular variable ŝ. Combining 共18兲, 共24兲,
where the inner boundary is empty, it means the region does and 共26兲, we arrive at
not contain another region and we have only ⳵ ⍀ i ⫽ ⳵ ⍀ ⫹ i ,
共respectively ⳵ ⍀̄i ⫽ ⳵ ⍀̄⫹
⳵⍀⫹
i for voids兲.
1 will also be denoted by ⳵⍀.
The outer boundary q 0 共 m兲 ⫽ 冕 ⳵ ⍀̄
h m,m⬘
ŝm,m⬘• ␯ˆ 共 m⬘ 兲 ŝm,m⬘ –␯ˆ 共 m兲
2
␲ d m,m⬘

⫻e ⫺ ␮ a 共 ⍀̄ 兲 d m,m⬘ ␨ ⌽ 共 m⬘ 兲 dm⬘ . 共27兲


B. Radiosity-diffusion approach
Note that
Since the radiosity solution is essentially a steady-state
one, we will first discuss this version of the problem. The
extension to time and frequency domain cases is given below
冕 ⳵ ⍀̄
ŝm,m⬘ –␯ˆ 共 m⬘ 兲 ŝm,m⬘ –␯ˆ 共 m兲
2
␲ d m,m⬘
dm⬘ ⫽1. 共28兲
in Sec. VI. Let ⍀̄ be a void with boundary ⳵ ⍀̄, and absorp-
We want to point out that the above discussion is only
tion coefficient ␮ a (⍀̄). Let m,m⬘ 苸 ⳵ ⍀̄ be points on the void
strictly true for the case where the refractive index is
boundary, and let ␯ˆ 共m兲,␯ˆ 共m⬘兲 be their respective inward di-
matched across domain interfaces. In particular, although
rected surface normals, as illustrated in Fig. 2. We define
共25兲 is always correct for outgoing flux, from a region of a
high refractive index to a region of a low refractive index,
the same expression for incoming flux should contain an an-
gular dependency, for the mismatched case.30 In this paper
we will present results only for the case where the voids have
the same refractive index, and will address the mismatched
case in a subsequent publication.

IV. FINITE ELEMENT METHOD


In Ref. 8 we proceeded from 共24兲 by using analytical
expressions for the Green’s functions in a homogeneous slab
to develop a hybrid radiosity-diffusion model for a simple
case of two slabs separated by a uniform gap. We developed
FIG. 2. Calculation of form factor. m and m⬘ are points on the void bound-
ary, separated by distance d m,m⬘ . ␯ˆ 共m兲, ␯ˆ 共m⬘兲, are the respective surface the method using an iterative procedure that solved on suc-
normals. ŝm,m⬘ is a unit vector from m to m⬘. ␮ a (⍀̄) is the attenuation cessive sides of the gap from the radiation propagated from
coefficient in the void. the opposite side, iterating to convergence. We now develop

Medical Physics, Vol. 27, No. 1, January 2000


256 Arridge et al.: Domains with nonscattering regions 256

a direct, noniterative procedure for any general geometry, ŝm,m⬘ –␯ˆ 共 m⬘ 兲 ŝm,m⬘ –␯ˆ 共 m兲
utilizing the finite element method 共FEM兲. In Appendix B f m,m⬘ ⫽ 2 . 共36兲
␲ d m,m⬘
we demonstrate the equivalence of the two methods.
A. Purely diffusing case „no voids… Note that u i and u j are integrated over independent vari-
In the absence of voids, the finite element implementation ables in 共35兲, so that nonzero contributions are not limited to
of 共21兲 is obtained by specifying that the domain ⍀ is di- regions where these functions have overlapping support. We
vided into P elements, joined at D vertex nodes. The solution define
⌽ is approximated by the piecewise polynomial function A j 傺 ⳵ ⍀̄k ⫽supp共 u j 兲 艚 ⳵ ⍀̄k 共37兲
i ⌽ i u i (r)苸* , where * is a finite-dimensional
⌽ h (r)⫽ 兺 D h h

subspace spanned by basis functions 兵 u i (r);i⫽1, . . . ,D 其 as the part of the boundary of ⍀̄k that intersects the support
chosen to have limited support. The problem of solving for of basis function u j . Then Eq. 共35兲 can be expressed as a
⌽ h becomes one of sparse matrix inversion for which stan- matrix relation:
dard methods such as Cholesky decomposition or conjugate
gradient solvers are readily available. q0 共 ⍀̄k 兲 ⫽E共 k 兲 ⌽, 共38兲
As developed in I and II, the diffusion equation in the where the coupling matrix E(k) for void ⍀̄k is defined with
FEM framework is expressed as entries
F⌽⫽q0 ,
where
共29兲
E 共i kj 兲 ⫽ ␨ 冕 Ai
u i 共 m兲 冕
Aj
u j 共 m⬘ 兲 h m,m⬘ f m,m⬘ e ⫺ ␮ a 共 ⍀̄ 兲 d m,m⬘

F⫽K共 ␬ 兲 ⫹C共 ␮ 兲 ⫹ ␨ A. 共30兲 ⫻d n⫺1 m⬘ d n⫺1 m. 共39兲

The system matrices K, C, A have entries given by Note that the coupling is defined for both boundaries ⳵ ⍀̄⫺
and ⳵ ⍀̄⫹ .
Ki j⫽ 冕␬

共 r兲 “u i 共 r兲 –“u j 共 r兲 d n r, 共31兲
Combining all coupling matrices from all voids, E
⫽ 兺 k E(k) , we obtain an additional set of equations that are
added to the system matrix in 共30兲 to represent radiative

Ci j⫽ 冕␮

a 共 r兲 u i 共 r兲 u j 共 r兲 d
n
r, 共32兲
propagation in voids. The new structure is in the DC case
F̃⫽K共 ␬ 兲 ⫹C共 ␮ 兲 ⫹ ␨ A⫺E. 共40兲

Ai j⫽ 冕
⳵⍀
u i 共 m兲 u j 共 m兲 d n⫺1 m, 共33兲
V. IMPLEMENTATION
and the source vector q0 has terms A. Approximations


We make a few simplifications in our implementation.
q 0,i ⫽ u i 共 r兲 q 0 共 r兲 d n r. 共34兲
⍀ 共i兲 We determine h m,m⬘ only for the nodal points:

B. Diffusing medium with void regions hi j ⫽h Ni ,N j .


In the presence of voids we define a finite element mesh
for ⍀ with nodes N j 苸⍀ d only in diffusing regions. We de- This visibility is determined by testing the infinite line in
direction Ni ⫺N j for an intersection with the affine equations
fine N( ⳵ ⍀̄k )⫽ 兵 N j 兩 N j 苸 ⳵ ⍀̄k 其 as the set of void-boundary
defining each element face that is part of the void boundary.
nodes for void ⍀̄k . We define the term void-boundary ele-
The result for all pairs of nodes is a Boolean matrix that is
ments for those elements containing one or more void-
precomputed during mesh generation. Although this calcula-
boundary nodes.
tion is straightforward, it is computationally expensive if
The modification comes from the internal source terms on
there are a large number of faces, and is ammenable to ac-
the void boundaries. For each void ⍀̄k , on projection into the celeration techniques based on hierarchical decomposition
FEM basis *h , 共27兲 gives rise to additional source terms for that are commonly used in computer graphics. So far we
the nodes 兵 i 兩 Ni 苸N( ⳵ ⍀̄k ) 其 of the form have not implemented such methods. The remaining disad-

冕 冕
vantage of the nodal visibility problem is that it fails to ac-
q 0,i 共 ⍀̄k 兲 ⫽ u i 共 m兲 h m,m⬘ f m,m⬘ e ⫺ ␮ a 共 ⍀̄ 兲 d m,m⬘ ␨ count for ‘‘partial occlusion’’ of one patch for another.
⳵ ⍀̄ k ⳵ ⍀̄ k
共ii兲 Except for nodes that fall below a proximity threshold,
⫻ 兺j ⌽ j u j 共 m⬘ 兲 d n⫺1
m⬘ d n⫺1
m, 共35兲 we determine d m,m⬘ only for the nodal points:

where we have made the definition d i j ⫽d Ni ,N j .

Medical Physics, Vol. 27, No. 1, January 2000


257 Arridge et al.: Domains with nonscattering regions 257

FIG. 3. The vertex normal associated with node Ni is the normalized average
of the surface normals of all polygons meeting at node Ni .

共iii兲 We determine the factor ŝm,m⬘ –␯ˆ (m⬘ )ŝm,m⬘ –␯ˆ (m)
only at the nodal points. The vertex normal ␯ˆ (N j ) for node j
is given by the average of the surface normals of those ele-
ment faces A(N j )⫽ 兵 A i 兩 N j 苸Ai 其 that share the vertex
␯共 N 兲
兺 ␯ˆ 共 A i 兲 ␯ˆ 共 Nj 兲 ⫽ 兩 ␯共 Njj 兲 兩 .
FIG. 4. Example mesh with a clear gap region.
␯共 N j 兲 ⫽
A 苸A共 N 兲
i j outer mesh cannot propagate to the inner nodes. The pres-
ence of the extra matrix term E is the coupling factor that
This is illustrated in Fig. 3. allows the photon density to propagate.
The effect of these approximations is to consider the term It may be noted that whereas the number of nonzeros in a
¯
h m,m⬘ f m,m⬘ e ⫺ ␮ a (⍀)d m,m⬘ in 共39兲 constant over the support of row of a conventional system matrix represents the number
each node, so that we obtain of nodes to which a given node is connected, and is therefore
cos ␪ i cos ␪ j a i a j small, the coupling terms are potentially nonzero for every
f i j ⫽ ␨ e ⫺␮adi j , 共41兲 pair of nodes in the void boundary. In effect, the void can be
␲ 兩 d i j兩 2 C i C j thought of as a single large element with many nodes, in
which a different equation to the boundary equation is being
where a i , a j are the area under the support of basis functions
solved 共i.e., direct rectilinear propogation—the streaming
u i , u j , respectively, cos ␪i⫽ŝm,m⬘ –␯ˆ (m), cos ␪ j⫽ŝm,m⬘ –
term only in the Boltzmann equation兲. If a conventional
␯ˆ (m⬘ ), and C i , C j are the cardinality of the sets A(Ni ),
mesh is optimised to minimize bandwidth then the extra cou-
A(N j ) respectively.
pling will greatly increase this bandwidth. Instead we utilize
B. The proximity problem a sparse matrix structure, and a minimum degree node order-
ing scheme,33 designed to minimize fillin on 共sparse兲
The approximations above will break down if nodes i and
Cholesky factorization. Of course this ordering is irrelevant
j are too close 共an example of the proximity problem in ra-
if we use an iterative matrix solver, which is the case for
diosity theory31兲. Generally speaking, in our implementation
three-dimensional 3-D meshes.
we consider meshes with a fine resolution near boundaries.
However, we also account for proximal nodes across a void VI. TIME AND FREQUENCY DOMAIN CASES
in the following way. The time-dependent diffusion equation is
We establish a threshold ␣ for the solid angle subtended
by one face at a point on the other face, such that if a i /d i j 1 ⳵ ⌽ 共 r,t 兲
⫺“–␬ 共 r兲 “⌽ 共 r,t 兲 ⫹ ␮ a 共 r兲 ⌽ 共 r,t 兲 ⫹ ⫽q 0 共 r,t 兲 ,
⬎ ␣ 共respectively, a j /d i j ⬎ ␣ 兲 we split faces in Ai 共respec- c ⳵t
tively, A j 兲 into P ⬘ sub patches and evaluate 共41兲 on each 共43兲
subpatch:
where c is the speed of light, and 共by Fourier transforma-
ai a j cos ␪ i ⬘ cos ␪ j ⬘
f i j⫽␨ 兺兺 e ⫺␮adi⬘ j⬘ . 共42兲
tion兲, the frequency-domain equation is

冉 冊
C i C j i⬘ j⬘ ␲ 兩 d i ⬘ j ⬘兩 2
i␻
This idea again comes from radiosity.32 ⫺“–␬ 共 r兲 “⌽ 共 r, ␻ 兲 ⫹ ␮ a 共 r兲 ⫹ ⌽ 共 r, ␻ 兲 ⫽q 0 共 r, ␻ 兲 .
c
共44兲
C. System matrix structure The corresponding FEM equations are now:
It is worth commenting on the structure of the sparse sys- ⳵ ⌽共 t 兲
F⌽共 t 兲 ⫹B ⫽q0 共 t 兲 共45兲
tem matrices. In a case such as in Fig. 4, it is clear that the ⳵t
inner and outer regions are decoupled. If we represent the
nodes as a column vector with the nodes in the outer region and
ordered first, followed by those in the inner region; then the F共 ␻ 兲 ⌽共 ␻ 兲 ⫽ 共 K共 ␬ 兲 ⫹C共 ␮ 兲 ⫹ ␨ A⫹i ␻ B兲 ⌽共 ␻ 兲 ⫽q0 共 ␻ 兲 .
system matrix will be block diagonal, and a source on the 共46兲

Medical Physics, Vol. 27, No. 1, January 2000


258 Arridge et al.: Domains with nonscattering regions 258

The mass matrix B has entries given by where Dn is a ‘‘delay’’ matrix with components

Bi j⫽ 冕 1
⍀c
u i 共 r兲 u j 共 r兲 d n r. 共47兲 D ni j ⫽ 共 d i j /c 兲 n , 共55兲

In I,1 we presented a finite-differencing scheme to derive and the notation DⴰE means pointwise 共Hadamard兲 multipli-
the time-dependent intensity ⌫(t). Subsequently we have cation of two matrices.
presented efficient methods to calculate derived data types
representing transforms of the time-resolved data with filters
of the form w(t)⫽t n e ⫺st . 34,35,16 We now illustrate this
scheme for the radiosity-diffusion approach, for the case of
temporal moments 共i.e. w(t)⫽t n 兲. VII. RESULTS
First, in the frequency domain, we introduce phase delay
in the radiative propagation so that the exponential term in In the following we compare results from the radiosity-
共39兲 becomes diffusion model, the diffusion only model, Monte Carlo, and
a production discrete ordinate transport code 共DANTSYS36兲.
e ⫺„␮ a 共 ⍀̄ 兲 ⫹i ␻ /c…d m,m⬘ . 共48兲 The transport code is written for the time-independent case;
therefore only flux intensities are available. For the other
Then our modified system matrix is models we also compare mean time results. All calculations
were performed on a 266 MHz Sun Sparc station with 128
F̃共 ␻ 兲 ⫽K共 ␬ 兲 ⫹C共 ␮ 兲 ⫹ ␨ A⫹i ␻ B⫺E共 ␻ 兲 . 共49兲 MB RAM.
Beginning with For the radiosity-diffusion code we compared Dirichlet
boundary conditions with Robin boundary conditions for
F̃共 ␻ 兲 ⌽⫽q0 , 共50兲 both a matched and a mismatched refractive index. As dis-
cussed in Sec. II, the vacuum boundary conditions in the
we obtain temporal moments by differentiating in frequency, discrete ordinate model, and the Robin boundary condition
共with no refractive index mismatch兲 in the diffusion approxi-
F̃共 ␻ 兲 ⌽⫽q0 , mation are the most commensurate. We investigated differ-
ent numbers of discrete ordinates and different grid spacings
F̃共 ␻ 兲 ⌽̇⫹ 共 B⫺Ė兲 ⌽⫽0,
to achieve reliable results from the transport code. In agree-
F̃共 ␻ 兲 ⌽̈⫹2 共 B⫺Ė兲 ⌽̇⫺Ë⌽⫽0, 共51兲 ment with Ref. 7, we required adjacent grid points to be
separated by a distance no larger than one scattering mean-
]] free path. In addition, up to 30 discrete ordinates 共corre-

冉 冊
sponding to 900 angular directions兲 were needed for some
共 n• 兲 共 n⫺1• 兲 n 共 n ⬘ . 兲 共 n⫺n ⬘ • 兲
n cases, leading to very long computation times 共14 min for
F̃共 ␻ 兲 ⌽ ⫹n 共 B⫺Ė兲 ⌽ ⫹ 兺 n⬘
E ⌽ ⫽0.
S n ⫽4 up to 11.5 h for S n ⫽30兲. See below for further dis-

n ⫽2
cussion on the required angular resolution.
The above system is solved iteratively to yield the moments For the Monte Carlo results, the number of simulated pho-
共 n• 兲 tons was between 1⫻107 and 3⫻108 . The differences arise
⌽ in the need to use more photons when only small void re-
具 tn 典 ⫽ .
⌽ gions are modeled, in order to gather sufficient statistics.
Conversely, when the nonscattering region is large, the ac-
To differentiate the system matrix with respect to frequency, celeration of photons through the void leads to shorter simu-
we establish lation times. Typical times were 3 to 4 days. The refractive
⳵ n E共 ␻ 兲 共 n• 兲 index throughout the Monte Carlo 共MC兲 model was set to
in ⫽ E, 共52兲 1.4, and both the reflection and no reflection at boundaries
⳵␻ n
cases were modeled. The former is expected to more accu-
(n•)
rately match the diffusion case with Robin boundary
where, in comparison with 共39兲, E has components
conditions.2
共 n• 兲
E i j⫽␨ 冕 Ai
u i 共 m兲 冕 Aj
u j 共 m⬘ 兲 h m,m⬘ f m,m⬘ A. Case 1: Clear gap
As a first example, we consider the case of two diffusing


c 冉 冊
d m,m⬘ n
e ⫺ ␮ a 共 ⍀̄ 兲 d m,m⬘ d n⫺1 m⬘ d n⫺1 m. 共53兲
regions separated by a nonscattering gap. The example
shown is for a two-dimensional 共2-D兲 circular domain of
radius 25 mm with ␮ s⬘ ⫽2 mm⫺1 , ␮ a ⫽0.025 mm⫺1 , con-
Note that in the absence of subsplitting for the proximity taining a nonscattering ring of 1, 2, and 3 mm thickness
problem, we can represent occupying the ranges 21–22 mm, 20–22 mm, and 20–23
共 n• 兲 mm radii, respectively. The gap had absorption coefficient
E ⫽Dn ⴰE, 共54兲 ␮ a ⫽0.025 mm⫺1 . The mesh is shown in Fig. 4 for the 2 mm

Medical Physics, Vol. 27, No. 1, January 2000


259 Arridge et al.: Domains with nonscattering regions 259

case. It has 4368 elements and 2398 nodes, of which 267 lie
on the void boundary. Computation times for this case was
⬃12 s.
Results for DC intensity are summarized in Fig. 5. There
is a qualitative agreement between the data from the discrete
ordinate code and the radiosity-diffusion code, although the
former always underestimates the latter, particularly at points
farthest away from the source. However, both models show a
marked difference from the concept of modeling the gap by
a low ␮ s⬘ value in a simple diffusion model. This case,
shown in Fig. 5共a兲, is markedly in error, which stems from
the fact that the diffusion model is not valid in the low scat-
tering limit where ␮ a Ⰶ ␮ s⬘ .
A noticeable feature of these results is the ‘‘kink’’ that
occurs in the intensity profile, and the fact that this is dis-
placed as the width of the clear region increases. This kink
was at first thought to be an error since it was not reproduced
in the DANTSYS results computed with S n ⫽4. However, a
comparison with Monte Carlo showed that it was a real fea-
ture that results from photons that propagate for a long dis-
tance through the gap, when the line of sight from the source
to the detector has a significant extent through the gap 共see
Fig. 6兲. In order to reproduce this in DANTSYS it was neces-
sary to use a high number 共up to 30兲 of discrete ordinates.
When this was done, DANTSYS showed the kink too. Figure 7
shows the dependence of the DANTSYS results on the number
of discrete ordinates.
The displacement of the kink may be due to the fact that
when the gap thickness increases the area of a direct visible
region within the nonscattering region also increases. More
photons can therefore take this path and cause a distinct kink
in the intensity plots. Note also that the Robin condition case
appears to match the Monte Carlo data more closely near the
source, and the Dirichlet boundary condition is actually a
better match farther from the source. This may be due to the
choice of the boundary condition parameter, or due to poor
photon statistics in the Monte Carlo data.
For mean time of flight, only Monte Carlo results could
be compared since DANTSYS does not have a time-resolved
capability. Figure 8 summarizes these results. Mean time
data show good agreement with the MC data, as the mean
time of flight for photons is less responsive to small varia-
tions in measurement points. The standard deviation for the 3
mm case is smaller than that of 1 mm because a higher
number of photons propagated through the gap. The kink is
also seen here, since for measurements made at points near
the line of sight through the gap, there will be an increased
number of photons with a much shorter time of flight, that FIG. 5. A comparison of three models for void regions. The domain is a 25
have traveled a significant distance through the clear region. mm circle with gaps of width 1 mm 共a兲, 2 mm 共b兲, 3 mm 共c兲. In the nonvoid
region the optical parameters were ␮ a ⫽0.025 mm⫺1 , ␮ s⬘ ⫽2 mm⫺1 . The
gap ␮ a was 0.025 mm⫺1 in each case. The quantity plotted is intensity
共exitance兲 as a function of angular separation of source and detector. The
source was isotropic at angular position zero, and at a depth of 0.5 mm in all
three models. Lines are the radiosity-diffusion model with Dirichlet 共solid兲
B. Case 2: Single boundary clear hole
and Robin 共dashed兲 boundary conditions at the outer boundary and Robin
As a second example we show a circular domain, radius condition at the void boundary in both cases. Dots are DANTSYS with a
25 mm with ␮ s⬘ ⫽2 mm⫺1 , ␮ a ⫽0.025 mm⫺1 , containing a
vacuum boundary condition. Monte Carlo results are shown without reflec-
tion 共circle兲, and with reflection 共cross兲. For the 1 mm gap case, we also
concentric void radius 10 mm with no scattering and ␮ a show the simple diffusion model 共Dirichlet boundary condition兲 共triangle兲
⫽0.025 mm⫺1 . The parameters of the diffusing region are where the gap ␮ s⬘ was 0.005 mm⫺1.

Medical Physics, Vol. 27, No. 1, January 2000


260 Arridge et al.: Domains with nonscattering regions 260

FIG. 6. Photons on line of sight propagate through the gap with less attenu-
ation, and decreased mean time.

the same as in I.1 It should be noted that this case could not
be modeled using our previous iterative approach,8 since the
void has only one boundary.
Results for boundary flux and mean time are presented in FIG. 8. A comparison of mean time data 共in picoseconds兲 from the radiosity-
Fig. 9. As in case 1, the Monte Carlo and radiosity-diffusion diffusion model and the Monte Carlo model for the three gap cases as in Fig.
models are in good agreement, with the DANTSYS results un- 5共a兲, 共b兲, and 共c兲. Lines represent the radiosity-diffusion model for 1 mm
derestimating the boundary flux. 共solid兲, 2 mm 共dashed兲, and 3 mm 共dotted兲. Monte Carlo without reflection
共circle兲, and with reflection 共cross兲.

We also show internal fluxes in Fig. 10 for the discrete


ordinate code and for the radiosity-diffusion model. The
agreement between the two is remarkably good.

VIII. CONCLUSIONS
We have presented a method for directly incorporating
‘‘true’’ void regions into a FEM model of diffusive light
propagation in tissue. The void model can be thought of as a
hybrid of two physical models : diffusive propagation in re-
gions, coupled with rectilinear 共streaming兲 propagation in
nonscattering regions. We note that our method is related to
a commonly used approach in electrostatics of scattering in-
tegrals. There the domains are assumed piecewise continu-
ous and a boundary integral expression is evaluated. This
method has been applied for steady-state optical diffusion
problems.37 More generally, the scattering integral method is
an example of a boundary element method 共BEM兲, where
only surfaces of domains are discretized and their terms are
related via Green’s functions.38 Therefore another interpreta-
tion of our method is a hybrid FEM–BEM model. The ad-
vantage over a purely BEM solution is that the BEM regions
are necessarily piecewise continuous, whereas the FEM re-
gions can handle general heterogeneities.
We compared the method with a production transport
code based on the discrete ordinate method 共DANTSYS兲, and
with Monte Carlo runs. There is broad agreement between
all three results with marked difference from the naive ap-
proach of setting ␮ s⬘ →0 in the diffusion coefficient. In con-
trast to DANTSYS, our method also evaluates statistics of the
time of flight of photons. In principle, the frequency domain
case could also be modeled.
Agreement between the different methods is not perfect,
FIG. 7. Discrete ordinate derived boundary fluxes using DANTSYS for a dif-
but there are several factors that could be investigated. First,
ferent number of ordinates for the 3 mm gap case. Note the discrepancy in
the region around the direct line of sight through the gap, which is shown the value of the boundary condition coefficient is still a mat-
enlarged in the lower figure. ter of debate in biomedical optics, and its adjustment may

Medical Physics, Vol. 27, No. 1, January 2000


261 Arridge et al.: Domains with nonscattering regions 261

FIG. 10. Internal photon flux for 25 mm circle with 10 mm void 共case 2兲: 共a兲
DANTSYS; 共b兲 radiosity-diffusion. In both cases, the range was scaled to give
20 equally spaced contour intervals on a logarithmic scale from maximum to
minimum intensity.

mark’’ for the evaluation of other methods for the forward


problem, and for providing test data for the inverse problem.
FIG. 9. Boundary flux and mean time 共in picoseconds兲 for case 2 共25 mm
circle with 10 mm void兲. 共a兲 A comparison of fluxes for DANTSYS, radiosity- ACKNOWLEDGMENTS
diffusion, and Monte Carlo. 共b兲 A comparison of mean time for radiosity-
diffusion, and Monte Carlo. We thank Yiorgis Chrysanthou, and Jorge Ripoll for very
helpful discussions, and Marko Vauhkoven and David Delpy
for reading the manuscript. Funding was provided by EPSRC
generate better agreement. Second, the use of alternatives to and the Wellcome Trust.
the Lambertian cosine weighting for the emergent angular
distribution at the void boundary could be investigated. In
particular, it is worth noting that in radiosity theory it is APPENDIX A: RELATIONSHIP TO RADIOSITY
possible to include, in addition to the ‘‘diffuse–diffuse’’ cou- THEORY
pling, terms representing ‘‘diffuse–specular’’ and In Ref. 8 we pointed out the similarity of our modified
‘‘specular–specular’’ coupling.39 However, in the absence of diffusion theory to the radiosity theory of computer
a precise theory for expected angular distribution, such an graphics.32 In this technique, a scene to be visualized is con-
approach would be somewhat ad hoc. A more promising sidered as a set of surfaces, divided into a finite number of
approach is to note that our method could be combined also patches and illuminated by external sources. As well as the
with higher-order approximations to the Boltzmann equation. direct illumination from the external sources, the relative
We anticipate the radiosity-diffusion approach in FEM contribution between patches is considered to give rise to a
will be of interest to researchers attempting to image the coupled set of equations that are solved for the ‘‘brightness’’
brain using optical tomography. To use in an inverse solver, of each patch. We now briefly summarize this theory, mak-
it is necessary to specify the boundary of the voids—such ing use of the established notation as far as possible.
information might come, for example, from auxiliary imag- Let A i be an area patch on one surface and A j be an area
ing modalities such as magnetic resonance imaging 共MRI兲.40 patch on the same or a different surface 共the patches can be
However, even this may not be a restriction if the boundary the same兲. The radiosity model couples the incoming radia-
of the void region is considered to be part of the inverse tion over A i to the outgoing radiation from all points on A j
solution.41 In addition, our method may provide a ‘‘bench- that are visible, in the sense that a direct line of sight from

Medical Physics, Vol. 27, No. 1, January 2000


262 Arridge et al.: Domains with nonscattering regions 262

m苸A i to m⬘ 苸A j does not intersect another surface. The 共iii兲 Solve in each domain ⍀ i propagated to in the previ-
relationship between patches A i and A j is defined in terms of ous step, assuming zero Robin boundary conditions
a form factor, given by: on ⳵ ⍀ ⫺
i .
共iv兲
冕冕
Radiatively propagate the output intensity from each
f i j⫽ h m,m⬘ f m,m d n⫺1 m d n⫺1 m. 共A1兲 point on each domain currently solved, to every other
Ai Aj visible surface.
共v兲 Iterate until intensity converges.
Then the radiosity relationship between patches is conven-
tionally defined as The secondary sources that are generated are boundary
conditions on the internal surfaces. Thus, we can consider
⌽ i⫽ 兺j f i j ␳ j ⌽ j ⫹b i , 共A2兲 that the final system 共in the steady state兲 is just

F⌽共 ⬁ 兲 ⫽q共0⬁ 兲 , 共B1兲


where ␳ j is a term representing the reflectivity of the surface
at patch j, and b i is the emission of patch i 共resulting from (⬁)
direct illumination from an external source兲. Consideration where q is the final state of the boundary condition on all
(0)
of all patches leads to a set of simultaneous equations, which surfaces, including the initial source q . The radiative
in matrix form is propagation is entirely contained in the coupling matrices E,
共 I⫺E兲 ⌽⫽b, 共A3兲 so the iterative scheme looks like this:

where I is the identity matrix, and the components of matrix Solve F⌽共 0 兲 ⫽q共00 兲 ,
E are E i j ⫽ ␳ j f i j .
A comparison of 共A1兲 and 共A3兲 to 共39兲 and 共40兲 shows Set q共01 兲 ⫽E⌽共 0 兲 ,
that the basic relationships are the same, with the following
slight differences. Solve F⌽共 1 兲 ⫽q共00 兲 ⫹q共01 兲 ,

共i兲 Radiosity theory assumes all objects are opaque and Set q共02 兲 ⫽E⌽共 1 兲 ,
partially reflecting, but not diffusive.
共ii兲 The ‘‘reflectivity’’ ␳苸关0,1兴 of a patch in radiosity ]
considers external reflection of incoming to outgoing
radiation, whereas the surface constant ␨ in the diffu- Set q共0j 兲 ⫽E⌽共 j⫺1 兲 ,
sion theory is derived from internal reflections arising
from refractive index mismatch. Solve F⌽共 j 兲 ⫽q共00 兲 ⫹q共01 兲 ⫹¯⫹q共0j 兲 .
共iii兲 Radiosity has to consider the interaction of all sur-
We establish G⫽F⫺1 as the formal solution to the for-
faces to all others and visibility determination has to
ward problem. Then our series is expressed as
test for the intersection with all surfaces. By contrast,
our model only considers the coupling across a single j
void 共which may contain several diffusive ‘‘islands’’, q共0j 兲 ⫽ 兺 共 EG兲 n q共00 兲 . 共B2兲
such as ⍀̄3 in Fig. 1兲. n⫽0

共iv兲 In the standard radiosity theory, the shape functions u i


are constant over each patch, whereas FEM theory For the infinite sum, we have convergence since the Frobe-
requires at least first-order continuity. nius norm of EG⬍1, leading to

Although we will not make use of it here, this unification q共0⬁ 兲 ⫽ 共 I⫺EG兲 ⫺1 q共00 兲 . 共B3兲
of diffusion theory and radiosity theory will allow us to
We thus have a scheme to derive the total boundary condi-
make use of the considerable body of work on the develop-
tions in the steady state. The matrix 共I⫺EG兲 will be invert-
ment of efficient and accurate numerical methods for deter-
ible. Now consider the photon density given by the soluton
mining light interaction between surfaces.
to 共B1兲,

F⌽共 ⬁ 兲 ⫽q共0⬁ 兲 ⫽ 共 I⫺EG兲 ⫺1 q共00 兲 ⇒ 共 I⫺EG兲 F⌽共 ⬁ 兲


APPENDIX B: RELATION TO AN ITERATIVE ⫽q共00 兲 ⇒ 共 F⫺E兲 ⌽共 ⬁ 兲 ⫽q共00 兲 . 共B4兲
APPROACH
This equation gives the same system matrix as 共40兲 in the
Our previous method was as follows.
text and proves the equivalence of the two methods. How-
共i兲 Introduce a source q 0 on ⳵⍀ and solve in ⍀ 1 , assum- ever, we note that although we have developed a direct form,
ing Robin boundary conditions of zero on ⳵ ⍀ ⫺ 1 . an iterative solution method may still be of use if the dimen-
共ii兲 Radiatively propagate the output intensity from each sion of the system becomes very large. In fact, this concept is
point r j 苸 ⳵ ⍀ ⫺ ⫹
1 to surfaces ⳵ ⍀ i ‘‘visible’’ from r j to also familiar in radiosity theory, where the iterative approach
form a set of sources 兵 q i 其 . is known as progressive refinement.42

Medical Physics, Vol. 27, No. 1, January 2000


263 Arridge et al.: Domains with nonscattering regions 263

1
GLOSSARY S. R. Arridge, M. Schweiger, M. Hiraoka, and D. T. Delpy, ‘‘A finite
element approach for modeling photon transport in tissue,’’ Med. Phys.
Table B1. Definition of variables. 20, 299–309 共1993兲.
2
M. Schweiger, S. R. Arridge, M. Hiraoka, and D. T. Delpy, ‘‘The finite
Quantity Meaning Dimension element model for the propagation of light in scattering media: Boundary
and source conditions,’’ Med. Phys. 22, 1779–1792 共1995兲.
r Position vector 共L,L,L兲 3
M. Schweiger and S. R. Arridge, ‘‘The finite element model for the
m Measurement position vector 共L,L,L兲 propagation of light in scattering media: Frequency domain case,’’ Med.
constrained to a surface ⳵⍀ Phys. 24, 895–902 共1997兲.
4
Nj Position vector of a node in a mesh 共L,L,L兲 K. D. Paulsen and H. Jiang, ‘‘Spatially-varying optical property recon-
␯ˆ Outward directed surface normal struction using a finite element diffusion equation approximation,’’ Med.

冉 冊
sin ␽ cos ␸ Unit vector in S n⫺1 Phys. 22, 691–701 共1995兲.
5
R. Model, M. Orlt, M. Walzel, and R. Hünlich, ‘‘Reconstruction algo-
ŝ⫽ sin ␽ sin ␸
rithm for near-infrared imaging in turbid media by means of time-domain
cos ␽
data,’’ J. Opt. Soc. Am. A 14, 313–324 共1997兲.
␻ Modulation frequency T ⫺1 6
S. Takahashi, D. Imai, and Y. Yamada, ‘‘Fundamental 3D FEM analysis
␾共r,ŝ兲 Number of photons per unit volume L ⫺3 sr ⫺1 of light propagation in head model towards 3D optical tomography,’’
at position r with velocity in direction ŝ Proc. SPIE 2979, 130–138 共1997兲.
␮ s (r) Scattering coefficient at position r L ⫺1 7
A. H. Hielscher, R. E. Alcouffe, and R. L. Barbour, ‘‘Comparison of
␮ a (r) Absorption coefficient at position r L ⫺1 finite-difference transport and diffusion calculations for photon migration
␮ t (r)⫽ ␮ s (r)⫹ ␮ a (r) Attenuation coefficient at position r L ⫺1 in homogeneous and hetergeneous tissue,’’ Phys. Med. Biol. 43, 1285–
c Velocity of light LT ⫺1 1302 共1998兲.
q(r,ŝ) Number of photons per unit volume L ⫺4 sr -1 8
M. Firbank, S. R. Arridge, M. Schweiger, and D. T. Delpy, ‘‘An inves-
sourced at position r with velocity in tigation of light transport through scattering bodies with non-scattering
direction ŝ regions,’’ Phys. Med. Biol. 41, 767–783 共1996兲.
⌰共â,b̂兲⫽⌰共â•b̂兲 Normalized phase function representing 9
B. Davison, Neutron Transport Theory 共Oxford University Press, Oxford,
the probability of scattering from 1957兲.
10
direction â to direction b̂ A. M. Weinberg and E. P. Wigner, The Physical Theory of Neutron
⌽ Photon density L ⫺3 Chain Reactors 共University of Chicago Press, Chicago, IL, 1958兲.
J Photon current L ⫺3 11
M. C. Case and P. F. Zweifel, Linear Transport Theory 共Addison-
⌫ Exitance L ⫺3 Wesley, New York, 1967兲.
␨ Robin boundary constant
12
J. J. Duderstadt and W. R. Martin, Transport Theory 共Wiley, New York,
1979兲.
13
R. T. Ackroyd, Finite Element Methods for Particle Transport: Applica-
tions to Reactor and Radiation Physics 共Research Studies Press Ltd.,
Table B2. Other definitions. Taunton, 1997兲.
14
S. Chandrasekhar, Radiative Transfer 共Oxford University Press, London,
Quantity Meaning 1950兲.
15
A. Ishimaru, Wave Propagation and Scattering in Random Media 共Aca-
Functions: demic, New York, 1978兲, Vol. 1.
u i (r) Spatial basis function 16
S. R. Arridge, ‘‘Optical tomography in medical imaging,’’ Inverse Probl.
␪ j (s) Angular basis function 15, R41–R93 共1999兲.
d m,m⬘ Distance between m and m⬘ 17
R. E. Alcouffe, ‘‘Diffusion synthetic acceleration methods for the
ŝ m,m⬘ Unit vector from m⬘ to m⬘ diamond-difference discrete-ordinates equations,’’ Nucl. Sci. Eng. 64,
h m,m⬘ Visibility function between m and m⬘ 344 共1977兲.
18
f m,m⬘ Form factor between patches K. D. Lathrop, ‘‘Ray effects in discrete ordinates equations,’’ Nucl. Sci.
centered at m and m⬘ Eng. 32, 357–369 共1968兲.
19
Sets: K. D. Lathrop, ‘‘Remedies for ray effects,’’ Nucl. Sci. Eng. 45, 255–268
⍀k A connected set comprising 共1971兲.
20
a scattering region R. Sanchez and N. J. McCormick, ‘‘A review of neutron transport ap-
⍀̄k A connected set comprising proximations,’’ Nucl. Sci. Eng. 80, 481–535 共1982兲.
21
a nonscattering region 共i.e., a void兲 J. K. Fletcher, ‘‘The solution of the multigroup neutron transport equa-
⳵⍀k Boundary of a scattering region tion using spherical harmonics,’’ Nucl. Sci. Eng. 84, 33–46 共1983兲.
22
⳵ ⍀̄k Boundary of a void J. K. Fletcher, ‘‘A solution of the multigroup neutron transport equation
using spherical harmonics,’’ Nucl. Sci. Eng. 15, 157–179 共1986兲.
N( ⳵ ⍀̄k ) Set of nodes on the boundary 23
R. Aronson, ‘‘Boundary conditions for diffusion of light,’’ J. Opt. Soc.
of void ⳵ ⍀̄k
Am. A 12, 2532–2539 共1995兲.
A(N j ) Set of element faces that 24
D. Contini, F. Martelli, and G. Zaccanti, ‘‘Photon migration through a
share vertex N j
turbid slab described by a model based on diffusion approximation,’’
Matrices: Where defined Appl. Opt. 36, 4587–4599 共1997兲.
K System matrix for ␬ 共31兲 25
R. T. Ackroyd, J. G. Issa, and N. S. Riyait, ‘‘Treatment of voids in finite
C System matrix for ␮ 共32兲 element transport methods,’’ Prog. Nucl. Energy 18, 85–89 共1986兲.
A System matrix for boundary condition 共33兲 26
R. T. Ackroyd and N. S. Riyait, ‘‘Iteration and extrapolation methods for
B Mass matrix 共47兲 the approximate solution of the even-parity transport equation for sys-
E Void coupling matrix 共39兲 tems with voids,’’ Ann. Nucl. Energy 16, 1–32 共1989兲.
F⫽K⫹C⫹␨A Combined system matrix—DC case 共30兲 27
S. L. Scholfield, ‘‘A maximum principle for the first-order Boltzmann
F( ␻ )⫽F⫹i ␻ B Combined system matrix— 共46兲 equation, incorporating a potential treatment of voids,’’ Ann. Nucl. En-
frequency domain ergy 15, 543–551 共1988兲.
F̃⫽F⫺E Void system matrix—DC case 共40兲 28
A. K. Ziver, ‘‘Comparison of numerical solutons to predict neutron
F̃共␻兲⫽F̃共␻兲⫺E共␻兲 Void system matrix— 共49兲 streaming in shield design for reprocessing plants,’’ Prog. Nucl. Energy
frequency domain 18, 215–226 共1986兲.
29
R. S. Baker, W. F. Filippone, and R. E. Alcouffe, ‘‘The multigroup and

Medical Physics, Vol. 27, No. 1, January 2000


264 Arridge et al.: Domains with nonscattering regions 264

radial geometry formulation of the Monte Carlo/S N response matrix transport code system,’’ Manual LA-12969-M, Los Alamos National
method,’’ Nucl. Sci. Eng. 105, 184 共1990兲. Laboratory.
30 37
J. Ripoll and M. Nieto-Vesperinas, ‘‘Index mismatch for diffusive photon J. Ripoll, A. Madrazo, and M. Nieto-Vesperinas, ‘‘Scattering of electro-
density waves both at flat and rough diffuse–diffuse interfaces,’’ J. Opt. magnetic waves from a body over a random rough surface,’’ Opt. Com-
Soc. Am. A 16, 1947–1957 共1999兲. mun. 142, 173–178 共1997兲.
31
D. R. Baum, H. E. Rushmeier, and J. M. Winget, ‘‘Improving radiosity 38
A. A. Becker, The Boundary Element Method in Engineering 共McGraw-
solutions through the use of analytically determined form factors,’’ Com- Hill, Maidenhead, 1992兲.
put. Graph. 23, 325–334 共1989兲. 39
J. R. Wallace, M. F. Cohen, and D. P. Greenberg, ‘‘A two-pass solution
32
M. F. Cohen and J. R. Wallace, Radiosity and Realistic Image Synthesis to the rendering equation: a synthesis of ray tracing and radiosity meth-
共Academic, London, 1993兲. ods,’’ Comput. Graph. 21, 311–320 共1987兲.
33 40
A. George and J. W. Liu, Computer Solutions of Large Positive Definite M. Schweiger and S. R. Arridge, ‘‘Optical tomographic reconstruction in
Systems 共Prentice–Hall, Englewood Cliffs, NJ, 1981兲. a complex head model using a priori region boundary information,’’
34
S. R. Arridge and M. Schweiger, ‘‘Direct calculation of the moments of Phys. Med. Biol. 44, 2703–2721 共1999兲.
41
the distribution of photon time of flight in tissue with a finite-element V. Kolehmainen, S. R. Arridge, W. R. B. Lionheart, M. Vauhkonen, and
method,’’ Appl. Opt. 34, 2683–2687 共1995兲. J. P. Kaipio, ‘‘Recovery of region boundaries of piecewise constant co-
35
M. Schweiger and S. R. Arridge, ‘‘Direct calculation of the Laplace efficients of an elliptic PDE from boundary data,’’ Inverse Probl. 41,
transform of the distribution of photon time of flight in tissue with a 1375–1391 共1999兲.
finite-element method,’’ Appl. Opt. 36, 9042–9049 共1997兲. 42
M. F. Cohen, S. E. Chen, J. R. Wallace, and D. P. Greenberg, ‘‘A pro-
36
R. E. Alcouffe, R. S. Baker, F. W. Brinkley, D. R. Marr, R. D. O’Dell, gressive refinement approach to fast radiosity image generation,’’ Com-
and W. F. Walters, Dantsys: ‘‘A diffusion accelerated neutral particle put. Graph. 22, 75–84 共1988兲.

Medical Physics, Vol. 27, No. 1, January 2000

Das könnte Ihnen auch gefallen