7000514 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 20, NO.

2, MARCH/APRIL 2014

Modeling Light Scattering in Tissue as Continuous

Random Media Using a Versatile Refractive
Index Correlation Function
Jeremy D. Rogers, Andrew J. Radosevich, Ji Yi, and Vadim Backman

(Invited Paper)

Abstract—Optical interactions with biological tissue provide systems poses a challenge to studying length scales less than
powerful tools for study, diagnosis, and treatment of disease. When the diffraction limit (about half the wavelength). Some very
optical methods are used in applications involving tissue, scattering clever techniques have been developed to sidestep this issue,
of light is an important phenomenon. In imaging modalities, scat-
tering provides contrast, but also limits imaging depth, so models including STED, STORM, and other super resolution meth-
help optimize an imaging technique. Scattering can also be used ods [1]. These methods enable imaging or reconstruction of
to collect information about the tissue itself providing diagnos- extraordinary detail, but typically require contrast agents and
tic value. Therapies involving focused beams require scattering careful tissue preparation. Therefore, a niche remains available
models to assess dose distribution. In all cases, models of light for additional methods of quantifying fine length scales. Sev-
scattering in tissue are crucial to correctly interpreting the mea-
sured signal. Here, we review a versatile model of light scattering eral methods with potential to fill that niche rely on analysis of
that uses the Whittle–Matérn correlation family to describe the re- scattering.
fractive index correlation function Bn (rd ). In weakly scattering Many applications are critically affected by light scattering
media such as tissue, Bn (rd ) determines the shape of the power in tissue. Sometimes scattering is considered a nuisance, and
spectral density from which all other scattering characteristics are research has been devoted to optical clearing of tissue [2]. At
derived. This model encompasses many forms such as mass fractal
and the Henyey–Greenstein function as special cases. We discuss the same time, in many imaging techniques, scattering provides
normalization and calculation of optical properties including the the image contrast, for example, in confocal microscopy or op-
scattering coefficient and anisotropy factor. Experimental methods tical coherence tomography (OCT). Diffuse optical tomogra-
using the model are also described to quantify tissue properties that phy is perhaps the best example of using measurement of light
depend on length scales of only a few tens of nanometers. propagation in tissue to reconstruct tissue features that cannot
Index Terms—Biophotonics, continuous random media, mass otherwise be imaged and relies on reconstructing tissue prop-
fractal, scattering, tissue optics. erties based on transport theory due to multiple scattering [3].
Models of scattering are also important in treatment methods
such as photodynamic therapy where carefully controlling the
I. INTRODUCTION distribution of energy is crucial for successful treatment [4].
N BIOLOGY and medicine, optical techniques enable non- Clearly, methods of modeling radiative transport in tissue are
I invasive measurements of living tissue. Microscopy is par-
ticularly ubiquitous because it enables visualization of individ-
needed for a wide range of applications including treatment or
therapy, imaging, and characterization or diagnosis. One of the
ual cells and even organelles. There is a continuous effort to most widely used modeling methods is to numerically solve the
resolve ever finer detail to assess the most fundamental proper- radiative transport equation using Monte Carlo (MC) simula-
ties of living organisms. However, resolution in optical imaging tions. MC works by simulating a large number of rays as they
propagate by sampling two probability distribution functions:
1) The path length traveled between scattering events is de-
scribed by a decaying exponential associated with mean free
Manuscript received June 19, 2013; revised August 8, 2013; accepted Au- path (Beer–Lambert law). 2) The angular distribution of scat-
gust 26, 2013. Date of publication September 6, 2013; date of current version
October 25, 2013. This work was supported by the National Institutes of Health tered light for a single scattering event is called the phase func-
under Grant R01CA128641, Grant R01CA155284, and Grant R01EB003682, tion. The shape of the phase function is highly dependent on the
and the National Science Foundation under Grant CBET-1240416. The work scattering model. While MC is a powerful simulation method,
of A. J. Radosevich was supported by a National Science Foundation Graduate
Research Fellowship under Grant DGE-0824162. the results are only as good as the model it is based on.
J. D. Rogers is with the Department of Biomedical Engineering, University In this paper, we describe the use of a particularly flexible
of Wisconsin, Madison, WI 53706 USA (e-mail: [email protected]). choice for modeling tissue scattering. The model assumes that
A. J. Radosevich, J. Yi, and V. Backman are with the Department of
Biomedical Engineering, Northwestern University, Evanston, IL 60208 USA tissue can be described as a continuous random media rather
(e-mail: [email protected]; [email protected]. than discrete particles, which agrees with observations of ma-
edu; [email protected]). terial distribution such as electron microscopy. The model is
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. based on the Whittle–Matérn (WM) correlation family and is
Digital Object Identifier 10.1109/JSTQE.2013.2280999 highly flexible, encompassing several other models previously

1077-260X © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
ROGERS et al.: MODELING LIGHT SCATTERING IN TISSUE AS CONTINUOUS RANDOM MEDIA 7000514

used to describe tissue including mass fractals and the Henyey–

Greenstein (HG) phase function. While direct measurements of
the refractive index correlation function would provide the best
model of tissue, such measurements are currently not possible
with the necessary resolution of a few tens of nanometers. In the
meantime, the WM provides a good alternative. We summarize
the choices to make within this model and show how optical
properties are calculated. Finally, we demonstrate several tech-
niques where this model provides both excellent agreement with
data and also a means to quantify the distribution of length scales
well below the diffraction-limited resolution.

II. TISSUE, STRUCTURE, AND REFRACTIVE INDEX

Tissue is made up of many interconnected, arbitrarily shaped
structures that span a wide range of length scales. Many texts
can provide an overview of the components of cells or tissue [5].
For example, the extracellular matrix (ECM) is composed pri-
marily of collagen and elastin fibers ranging in size from 10
to 500 nm in diameter. Cells themselves are of various shapes
and sizes on the order of tens of micrometers. Cells are in turn
comprised of smaller structures such as membranes (10 nm) Fig. 1. A scanning electron microscope (SEM) image of colon biopsy showing
some of the structural features ranging in size from tens of nanometers to tens
and organelles including the nucleus (5–10 μm), mitochondria of microns (top). The wide range of length scales present and the complex
(0.2–2 μm), and many others. Cells also contain a cytoskeleton interconnected nature of the components mean that tissue is best modeled as a
made of filaments from 7 to 25 nm in diameter. The nucleus con- continuous random medium. Statistically, the scattering depends not on the exact
arrangement of material, but instead on the correlation B n (rd ). By calculating
tains nucleoli (0.5–1.0 μm) and DNA in the form of chromatin. the spectral density of the SEM (cropped to a square array), a medium can
The organization of chromatin is the subject of active research, be constructed that has the same autocorrelation shape as the SEM and is
but electron microscopy reveals that chromatin is found in the shown on the lower left. The WM model can be used to generate simulated
media that have a very similar appearance, an example for D = 3.7 is shown
form of heterochromatin and euchromatin and appears to have on the lower right. This illustrates how a medium can be constructed with
different densities. These form globules with length scales of a given autocorrelation function, but the reader is cautioned that the SEM is
hundreds of nanometers. not a measurement of refractive index and so direct comparison of the SEM
autocorrelation and B n (rd ) shape would not be valid.
These various structures span several orders of magnitude
in size and are composed of a wide range of materials. While
the exact refractive index of each component in living tissue While it is not yet possible to directly measure refractive
is not known, the heterogeneity of material nonetheless results index at the resolution of the smallest length scales mentioned
in a range of refractive index values with a complicated spatial above, there are ways to relate local mass density to refractive
distribution. It is this heterogeneity that gives rise to scattering. index. Since mass density can in principle be determined at high
Fig. 1 shows an image of tissue highlighting the dramatic range resolution using electron microscopy techniques, mass density
of length scales and the interconnected nature of the material. can be converted to refractive index using the Gladstone–Dale
This interconnected material means that refractive index is a relation [9], [10]:
continuous function of position and supports the modeling of
tissue as a continuous random medium like that of turbulence n = nwater + αρ (1)
[6], [7]. where ρ is the density of the solute [g/ml] and α is the refractive
When it comes to modeling light scattering, the exact distri- increment, usually α ≈ 0.17 ml/g.
bution for a particular ensemble of material is less important Future research using electron microscopy methods may en-
than the statistical properties. What matters most is the shape able direct measurement of mass density and thus refractive
of the excess refractive index correlation function Bn (rd ) dis- index at high resolution so that scattering may be more accu-
cussed below. By considering the spectral density of the SEM rately modeled and related to specific components of tissue. In
image, an artificial medium can be generated that has the same the meantime, we must be satisfied with empirical models that
correlation function as the SEM as shown in the lower left of provide good fits to observable data.
Fig. 1. A medium can also be generated from the WM model
using the method described in [8, Ch. 2] and an example is
A. Models of Tissue Scattering
shown in the lower right of Fig. 1 with remarkable similarity.
Caution should be taken in and quantitative comparison would Scattering in tissue is the result of light interacting with ran-
not be valid since the SEM does not directly represent refrac- dom variations in refractive index. To better understand scat-
tive index, but qualitatively the SEM represents the structure of tering in tissue, it is helpful to develop and evaluate models of
tissue containing a wide range length scales. light propagation in a medium comprised of a spatially random
7000514 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 20, NO. 2, MARCH/APRIL 2014

distributions of refractive index. This distribution could be in

the form of a continuous function of refractive index (like at-
mospheric turbulance) or in the form of discrete particles. Many
models of tissue scattering approximate tissue as a collection of
randomly placed spheres or spheroids [11]–[13]. This is plau-
sible considering that organelles such as nuclei, mitochondria,
lysosomes, vacuoles, and even whole cells are approximately
spherical (or spheroidal) in shape. The scattering from such iso-
lated particles can be analytically described by the Mie solution
to scattering from spheres [14], [15]. Scattering from a random
medium made up of many such spheres (i.e., a cloud of rain-
drops) is calculated by then incoherently summing the scattering
from many spheres according to the number density. However,
in many cases, it is hard to argue that the shape is spherical so Fig. 2. Example of experimental data that cannot be explained using the HG
Mie theory is inherently limited. function. Radiative transport in rat lung tissue is measured with EBS to produce
Another ubiquitous model of tissue scattering is based a probability distribution P (rs ) of the entrance/exit separation of rays scattered
in the tissue and eventually exiting. Two best fit models of P (rs ) computed
on the empirical observation that tissue scattering is usually numerically using MC are also shown based on WM (blue solid) and the special
anisotropic. That is, a very small volume of tissue scatters more case of D = 3 which is equivalent to the HG phase function commonly used to
light in the forward direction at small angles than in the back- model tissue. Fitting with a two parameter HG function produces a poor fit to
the small rs , while a three-parameter WM-based model results in an excellent
ward direction. A simple model of such anisotropic scattering fit to the experimental.
was developed by Henyey and Greenstein to describe scattering
of interstellar dust and is now referred to as the HG phase func-
inherently contain approximations. Choosing the right model is
tion [16]. Counterintuitively, the phase function has little to do
all about choosing the most appropriate set of approximations.
with the phase of a scattered ray. Instead, the term is historical
Fig. 2 shows an example where a simple model of scattering
and refers to the apparent brightness of celestial bodies such as
based on the HG function is not sufficient to explain the exper-
the moon or planets as they pass through their phases. The func-
imental data. In these data, radiative transport is measured with
tion is a normalized intensity as a function of the angle between
enhanced backscattering (EBS) at length scales ranging from
incident and scattered direction.
much less than the transport mean free path ls to much greater
Observing tissue scattering using methods such as goniome-
than ls , which for this tissue is about 200 μm. EBS measures
ter and integrating sphere, Jacques et al. found the HG function
the probability distribution P (rs ) of ray exit distances relative
provided a good model to match their observations [17]. It has
to the point of entry in a scattering medium, in other words the
since become one of the most common models of tissue scat-
spatial impulse response of diffuse reflectance [22].
tering. However, this model is limited in that it does not include
Most methods that measure radiative transport via backscat-
the dipole factor (discussed in Section IV-A) that leads to a
tering are only sensitive to separations larger than ls where the
nonmonotonically decreasing function of angle that is typically
propagation of light can be accurately modeled using the dif-
observed in scattering [11], [18] and it does not provide a phys-
fusion approximation [23]. The figure shows that the different
ical connection to the structures which generated it.
models tend to converge at these larger length scales, while
Other models of tissue scattering have been put forth and used
small distances exhibit more significant differences because in
with success and include delta-Eddington proposed by Joseph
this region, P (rs ) is dominated by low order scattering and is,
et al. [19] and discussed in detail by Prahl [20] that uses the
therefore, more sensitive to the shape of the phase function. The
sum of a Dirac-δ function and cos θ dependence to describe the
model curves are calculated by using the scattering model in MC
phase function, and the P3-approximation [21] that takes the
simulations of scattering to build a catalog of curves for ranges
first three terms in an expansion of the solution to the radiative
of parameters. Fits to experimental data are obtained by inter-
transport equation.
polating the model simulation space to produce a lookup table.
In choosing a model for tissue scattering, it is a good idea to
The dashed red line corresponds to the best fit obtained using
use the simplest model that matches the experimental method
the HG phase function, while a much better fit is obtained using
employed to avoid over fitting. For example, when using an in-
the WM model with D = 3.75 highlighting the need for more
tegrating sphere and the inverse adding doubling (IAD) method
a more advanced model for measurement methods that are ca-
[20], the experimental data are well modeled by the HG func-
pable of discriminating between different correlation functional
tion. However, as experimental methods evolve, some measure-
forms.
ments cannot be explained with such a simple model, and it is
necessary to invoke a more complex model of scattering.
III. RAYLEIGH–GANS–DEBYE APPROXIMATION
One of the most useful approximations that can be made in
B. As New Methods Emerge, Better Models Are Needed scattering from biological tissue is that of “weak” scattering. In
One challenge facing those wishing to model tissue scattering this approximation, it is assumed that the scattered field is much
is determining what model to use. Any model of scattering will less significant than the incident field. This approximation is
ROGERS et al.: MODELING LIGHT SCATTERING IN TISSUE AS CONTINUOUS RANDOM MEDIA 7000514

also known as Rayleigh–Gans–Debye (RGD) scattering, the first 1, where a is the particle radius. That is to say, the excess
Born approximation, or single scattering approximation. Here, refractive index must be small and the phase delay as the wave
we will use “RGD scattering” for brevity. This approximation traverses the particle must be much less than the wavelength.
can be used to describe discrete particles as well as continuous It is important to note that while the RGD scattering approx-
media and has been extensively investigated by Ishimaru for imation is often referred to as the single scattering approxima-
other refractive index correlation functions, so we will follow tion, this does not mean that it cannot be used to model large
his derivation for continuous random media [24]. Excellent dis- volumes such as clouds or tissue where multiple scattering takes
cussions of RGD scattering in particles are provided by Hulst place. What is required in the case of a particle is that single scat-
and van de Hulst [25] as well as Bohren and Huffman [15]. tering is valid for a single particle. When a medium is composed
A brief conceptual summary of the RGD approximation is of many such particles, light transport can be very accurately
as follows: When a plane wave traveling in free space passes computed numerically by methods such as MC [27]. The key is
through a particle, the dipoles that make up the particle are ex- to accurately compute the angular probability for each scatter-
cited and begin to oscillate reradiating electromagnetic energy ing event and then propagate rays according to mean free path
according to the Rayleigh or dipole radiation pattern [26]. In with subsequent scattering events. Interestingly, the same is true
the RGD approximation, the scattered field is insignificant rel- for continuous media.
ative to the incident field, so secondary scattering events can be The RGD approximation should, therefore, not be used to
ignored. This is equivalent to saying that the electric field arriv- model an entire region of tissue, but rather to statistically model
ing at any point inside the particle is approximately the same a representative volume just large enough to cause one scatter-
as the incident field. In other words, there is no phase delay or ing event. Multiple scattering can then be treated as decoupled
refraction inside the particle. The phase of each dipole is then events that can be summed according to the mean free path
determined by its position along the z-axis (direction of incident and scattering function using computational methods like MC.
wave). The total effect of all the dipoles can then be treated Such simulations can be made even more accurate by using the
by coherently summing the field from each dipole to obtain the Extended Huygens–Fresnel principle to account for the propa-
scattering amplitude function f (θ, φ). This summation can be gation of coherence in the case of multiple scattering, and an
mathematically written as an integral of relative excess refrac- excellent review of such treatment developed by Anderson and
tive index as a function of position nΔ (r) = (n(r) − n)/n, Thrane is provided in [28, Ch. 17].
where n is the expected value or average refractive index.
The spherical wave can be written as a complex exponential A. Validity of the RGD Approximation for Continuous Media
so that the operation takes the form of a Fourier transform. We
Unlike a particle, there is no clear boundary in continuous
are typically interested in the intensity of scattering, since this
media over which the concept of single scattering can be ap-
is the observable quantity. The intensity is proportional to the
plied. Instead, in continuous media, the single scattering ap-
square of the amplitude scattering function, which is the Fourier
proximation must be valid for a representative small volume of
transform of nΔ (r). By applying the Wiener–Khinchin theorem,
material. By representative, we mean that the volume must be
we see that the scattered intensity per unit volume as a func-
large enough to represent statistical ensemble of the medium,
tion of angle (known as the differential scattering cross section
but small enough that single scattering approximation holds. As
σ(θ, φ)) is proportional to the spectral density Φs (ks ), which
long as the volume can be defined for which small changes in the
is the Fourier transform of the autocorrelation of the excess
volume result in the same differential scattering cross section per
refractive index Bn (rd ):
 unit volume, the RGD approximation is valid. One could also
1  say that ls > Ln must be true, where ls is the mean free path and
Φs (ks ) = Bn (rd )e−i k s ·rd drd (2)
(2π)3 Ln is the outer length scale of the correlation beyond which the
where rd is a displacement vector between any two points in medium is not significantly correlated. In other words, a wave of
the medium. When the medium is statistically isotropic, Bn (rd ) propagating light must, on average, pass through enough media
is radially symmetric and can be represented as a 1-D function to be statistically representative before being scattered. This was
of scalar displacement rd . The result simplifies to the product rigorously investigated using finite-difference time domain and
of a dipole factor 1 − sin2 (θ) cos2 (φ) multiplied by the spec- the validity criterion can be written as σn2 (nkLn )2  1 [29].
tral density evaluated at ks = 2nk sin(θ/2), where k is the
freespace wavenumber: B. An Illustrative Example: Suspension of Spheres
  While we are most interested in modeling continuous media
σ(θ, φ) = 2π(nk)4 1 − sin2 (θ) cos2 (φ) Φs (ks )
  in this paper, it is instructive to follow through an example using
= 2π(nk)4 1 − sin2 (θ) cos2 (φ) a suspension of spheres. We can then compare the resulting
 ∞ scattering functions to those obtained with the Mie solution.
1 4π sin (2nk sin(θ/2)rd )
3
Bn (rd ) rd drd . (3) Incidentally, it is just such media in the form of polystyrene
(2π) 0 2nk sin(θ/2)
microsphere suspensions that are often used as tissue phantoms
Note that this relation works for a suspension of particles in research labs.
as well as continuous media. For particles, the weak scattering To begin, we must consider Bn (rd ) for a suspension of
approximation is valid when | nn  − 1|  1 and 2ka| nn  − 1|  spheres. For a sphere of radius a, the autocorrelation function
7000514 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 20, NO. 2, MARCH/APRIL 2014

Fig. 3. Comparison of the Mie solution for a suspension of microspheres to σ(θ, φ) calculated using the RGD approximation. The top two panels show the
calculation for polystyrene microspheres in water and while the 100 nm diameter spheres (upper left) show good agreement, the approximation breaks down for
400 nm spheres (upper right). However, when the index contrast is low n1/n0 = 1.02 as in the bottom row, the agreement holds for 400 nm spheres (lower left)
and even for large spheres of 3.0 μm (lower right). Calculations used a wavelength of 500 nm.

can be computed analytically, so Bn (rd ) is simply scaled by the increases a little (top right). For weaker refractive index contrast,
variance of the refractive index: the particle size can get much larger before the agreement breaks
⎧ down (bottom row). This is consistent with the requirements of
⎨ σ 2 (2a − rd ) (4a + rd ) , where rd ≤ a
2

(sphere) n
16a3 the single scattering approximation described in Section III-A.
Bn (rd ) =
⎩
0, where rd > a.
(4) IV. QUANTIFYING STRUCTURE WITH A REFRACTIVE
We must next calculate the expected value of refractive index INDEX CORRELATION FUNCTION
n and the variance σn2 of the relative excess refractive index Scattering from a medium can be computed analytically or
nΔ = n/n − 1. These values depend on the refractive index numerically for a known distribution of dipoles or, equivalently,
of the spheres ns , surrounding medium nm , and the volume a known distribution of refractive index. This is rarely if ever the
fraction Cv occupied by spheres case for biological tissue. It is, therefore, of interest to determine
n = Cv ns + (1 − Cv )nm (5) statistical average scattering from a random medium where the
 2  2 exact distribution of index is unknown, but the statistical prop-
ns nm erties are known, or can be approximated. One way to describe
2
σn = Cv − 1 + (1 − Cv ) −1 . (6)
n n the average scattering properties of a random medium is to de-
The Fourier transform of this function can be calculated ana- scribe the average shape of differential scattering cross section
lytically from a small representative volume of material. As seen in Sec-
tion III, an expected value of the differential scattering cross
3(1 − Cv )Cv (nm − ns )2 section σ(θ, φ) can be directly calculated from the refractive
Φs (ks ) =
2π 2 (Cv nm− nm − Cv ns )2 a3 ks6 index correlation function Bn (rd ).
× (ks a cos(ks a) − sin(ks a))2 (7)
A. WM Correlation Family
and substituted into (3) to calculate the differential scattering
cross section. If we plot the results and compare to the Mie The best model of tissue scattering would be based on a di-
solutions as in Fig. 3, we can see excellent agreement for very rect measurement of refractive index correlation over the entire
small particles (top left) or very small refractive index contrast range of lengths scales to which scattering is sensitive. This is
(bottom left). The accuracy begins to decrease as particle size currently not possible, and so lacking a catalog of refractive
ROGERS et al.: MODELING LIGHT SCATTERING IN TISSUE AS CONTINUOUS RANDOM MEDIA 7000514

Fig. 5. An example of the differential scattering cross section σ(θ, φ) for left
to right propagating vertically polarized light. This example is computed using
the WM function discussed below with parameter D = 3 in which case the
shape around the equator (φ = 90◦ ) corresponds to the HG phase function. The
dipole factor produces the dimples oriented in the plane of the electric field
polarization.
Fig. 4. Examples of the WM function for different values of D normalized
such that B n (rd = L n ) = 1. Inset shows the same functions on a log–log
scale.
be calculated by averaging over φ resulting in a phase function
index correlation functions in tissue, we can choose to model with a minimum at θ = 90◦ [31]. This extension of the HG
the correlation function using a functional form to approximate phase function more accurately reflects physical measurements.
the actual statistical correlation function. For example, Mourant reported goniometer measurements of
Several correlation functions have been used to model con- tissue along with fitting of the HG function, but the HG phase
tinuous random media. Chapter 16 in Ishimaru’s text describes function could not match the measurements in the backscattering
exponential correlation that leads to the Booker–Gordon for- direction [11], [18].
mula, the Gaussian model, and the Kolmogorov spectrum also
known as the von Kármán spectrum which is based on a power
law with a specific power [24], [30]. A number of papers have B. Normalization
proposed a fractal model for tissue [6], [7]. Mass fractals are The WM correlation family is useful because it can take
characterized by power law correlation functions where the mass on so many different functional forms. This does, however,
fractal dimension is related to the power. Interestingly, all of create some challenges in terms of normalizing the function. The
these models as well as the HG function can be described by the convention typically used in signal processing is that correlation
WM correlation family [30], which is the product of a power functions are normalized such that the value at the origin (rd =
law and a modified Bessel function of the second kind of order 0) is equal to 1, while in statistical disciplines, it is equal to the
(D − 3)/2, variance—in our case, the variance of excess refractive index
 (D −3)/2 σn2 . However, since power laws are unbounded at the origin, this
rd rd
Bn (rd ) = An K D −3 ( ). (8) convention is impossible to implement for D < 3. Therefore,
Ln 2 Ln
no single normalization can be used to satisfy convention in the
The fact that the WM family of correlation functions is so bounded case D > 3 and the unbounded case D < 3.
flexible and encompasses many other models makes it an attrac- The WM function has three parameters: D, An , and Ln . The
tive choice. The parameter D controls the shape of function and shape is determined by D, while the latter two can be thought
can be used to represent a wide range of plausible correlations of as scaling parameters. Ln scales the horizontal axis, while
including: An scales the vertical axis. Ln also represents a transition point
1) Gaussian: D → ∞; beyond which the function decays as an exponential. There are
2) exponential: D = 4; several options for the normalization coefficient An , and each
3) stretched exponential: 3 < D < 4; has advantages and disadvantages.
4) Kolmogorov/von Karman: D = 11/3; It should be noted that these normalization options only af-
5) HG: D = 3; fect the interpretation of the coefficient σn2 as the variance of
6) power law: D < 3. refractive index. The function Bn (rd ) is not arbitrary and ac-
Examples of the functional form for several values of D are tually represents a physical function that could, in principle, be
shown in Fig. 4. Note that when D < 3, the medium can be physically measured. In media where the best fit of Bn (rd ) is
considered a mass fractal and the parameter D takes on the for D ≤ 3, the actual physical value of σn2 is still finite, but the
special meaning of mass fractal dimension. power law only extends down to some small length scale below
Fig. 5 shows an example of the differential scattering cross which the shape is irrelevant. This is sometimes called the inner
section σ(θ, φ) (3) calculated for vertically polarized light and scale of turbulence. When this inner scale is much smaller than
D = 3. For this case, the shape around the equator at σ(θ, φ = the wavelength, the error in Φs between the unbounded model
90◦ ) corresponding to the shape of the HG phase function. function and the actual refractive index correlation function is
However, an important difference is the inclusion of the dipole small [31]. Nevertheless, it is useful to explore several options
factor. When incident light is unpolarized, the phase function can for normalizing Bn (rd ).
7000514 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 20, NO. 2, MARCH/APRIL 2014

1) Whittle and Matérn Normalization: Bn (0) = 1 for D > σn2 is also arbitrary which makes interpretation of σn2 difficult
3: A brief description of the history of the WM function and since it is no longer a material property but depends on the
some closely related functions used in modeling atmospheric chosen rm in . For a power law, a small change in the chosen
turbulence is provided by Guttorp and Gneiting [30]. These value of rm in could result in a large change in the model’s
authors describe the normalization that Whittle and Matérn used. value of σn2 and it does not necessarily correspond to the actual
In this case, the function is normalized to 1 (or σn2 ) at rs = 0 variance of the real medium.
for values of D > 3 that result in a bounded function. In other 3) Normalized to Value c at rd = Ln : Since the previous
words, σn2 is the variance for D > 3, and just a parameter for normalization can lead to a misinterpretation of the meaning of
D < 3: σn2 , one alternative is to simply normalize the model at a finite
5 −D
2 2 value of rd that is already defined in the model, namely Ln .
An = σn2 for D > 3. (9) The normalization constant c is not interpreted as variance of
|Γ( D 2−3 )|
refractive index; it is simply the value that scales the model to
The advantage of this normalization is that it is normalized match the actual medium:
at the origin when it is bounded, as any actual correlation func- c
tion should be. The disadvantage is that as D gets smaller and An = . (11)
K D −3 (1)
2
approaches 3, the narrow shape of the function forces the magni-
tude to zero at any finite value of rd . This, in turn, results in the This normalization has the advantage that the it is well behaved
total scattering cross section collapsing to zero. Additionally, for all values of D. Of course, the function is infinite at rd = 0
the function is identically zero for D = 3, and so we lose the for value of D ≤ 3 as expected. When D > 3, the function is
ability to model the HG function. Guttorp and Gneiting argue indeterminate at rd = 0, but the limit is bounded, so this poses
that (D − 3)/2 must be greater than 0 for this correlation to be no problem for modeling scattering. The disadvantage is that
valid (eliminating the need for the absolute value of the gamma relating this scaling parameter to σn2 requires an extra step, but
function). This is true in that σn2 cannot be defined when D ≤ 3, this can be easily dealt for values of D > 3:
but many interesting cases occur for these values of D, includ- 5 −D
2 2
ing a power law corresponding to mass fractals. When D ≤ 3, c = σn2  D −3  K D −3 (1). (12)
the function Bn (rd ) approaches a power law for small rd and Γ 2
2

so is infinite at the origin. When D ≤ 3, σn2 could be determined for the actual medium,
2) Normalized at a Minimum Length Scale: Bn (rm in ) = σn2 : but is not explicitly part of the model.
One possible normalization is to attempt to mimic reality. When These first three choices do not depend on scattering: they
D < 3, Bn (rd ) approximates a power law corresponding to a simply describe the properties of a medium. There are, how-
mass fractal organization of material. In nature, fractals occur ever, a few additional options for normalization that amount
that exhibit self-similarity over a range of scales and, therefore, to normalizing the spectral density independent of the scale of
are described by a power law correlation function. However, the Bn (rd ). These retain the dependence on the shape of Bn (rd ),
range of scales over which the true function closely follows a but not the magnitude.
power law is not infinite. For example, in tissue, one may observe ∞
4) Normalized Spectral Density: 0 Φs (ks )dks = 1: An-
length scales corresponding to the size of a cell. Zooming in,
one observes a similar organization in organelles, but at a length  ∞ option is to normalize the spectral density Φs such that
other
0 Φs (ks )dks = 1. In this case, the function takes the form of
scale an order of magnitude smaller. Zooming in yet further, one a Pearson distribution type VII:
might observe globule organization of macromolecules. This
5 −D
approximate self-similarity over several magnitudes of length 2π 2 2
An = . (13)
scales is what gives rise to a power law correlation function. Ln (D − 3) Γ( D 2−3 )
2
However, this power law cannot extend to infinitely small length
scales. At some small scale, only molecules are left and the very The advantage is that the spectral density is never unbounded
definition of refractive index becomes difficult to interpret. If for any value of D. One disadvantage is that because Φs is
the actual correlation function were known, the mode could be a probability function, it has no dependence on the variance
fit by defining a minimum length scale rm in and normalize the of refractive index σn2 . This choice of normalization requires
function at this value: D > 1 and has zeros, singularities, or is indeterminate at rd = 0
 (3−D )/2
 for values of D = 2, 3, 4.
rm in rm in 5) Rayleigh Scattering Normalized:
An = σn2 K D −3 . (10)
Ln 2 Ln 
σn2 2/π
Provided that rm in is chosen correctly, the value of the model An = (D /2−2) . (14)
2 Γ(D/2)
function at rm in can be close to the true value σn2 . The advantage
of this normalization is that there is an intuitively satisfying Another option is to normalize such that the scattering coef-
notion that the correlation always starts at σn2 for some arbitrarily ficient μs converges to the same value for any value of D in
small value regardless of the value of D. One problem with this the Rayleigh limit where Ln  λ. This is appealing in that as
normalization is that it introduces a fourth parameter rm in that the correlation function becomes much smaller than the wave-
must be arbitrarily chosen. Since rm in is arbitrary, the value of length, there should not be any dependence of the shape of
ROGERS et al.: MODELING LIGHT SCATTERING IN TISSUE AS CONTINUOUS RANDOM MEDIA 7000514

the correlation function (or particle). This normalization is then that is then measured in transmission and reflection. Although
equivalent to saying that all small particles look the same to Hall et al. showed that thicker tissue samples can be used if
long waves and scattering depends only on the mean free path multiple scatting is corrected using MC [33], prepared tissue
independently of the shape of the correlation function. The dis- is still required. The remaining methods are attractive because
advantage is that the scattering coefficient is indeterminate for they can be used in reflection and so have potential to be used
values of D = 2, 4. in vivo. Jacques et al. recently demonstrated quantification of
properties including μs and anisotropy g using confocal and
V. CALCULATING OPTICAL PROPERTIES FROM DIFFERENTIAL OCT methods [34]. As illustrated in Fig. 2, some methods such
SCATTERING CROSS SECTION as EBS are particularly sensitive to the shape of σ(θ, φ) and,
therefore, enable measurement of parameter D.
Armed with a method to calculate the differential scatter-
While EBS measurement of P (rs ) provides an excellent
ing cross section per unit volume σ(θ, φ), we can calculate
quantification of tissue properties and demonstrates the advan-
the optical parameters typically used to characterize a scatter-
tages of a more complex 3 parameter model, one limitation is
ing medium. These parameters include the scattering coefficient
the assumption that the medium is statistically homogeneous.
μs = 1/ls which is the inverse of mean free path ls . The scatter-
We know that tissue is often a layered medium, so methods
ing coefficient is equal to the total scattering cross section per
capable of quantifying the optical properties locally at different
unit volume and is easily obtained by integrating,
positions and depths would be of great utility. One such method
 2π  π
is based on OCT and allows assessment of properties in an imag-
μs = σ(θ, φ) sin θdθdφ. (15)
φ=0 θ =0
ing modality, a technique referred to as inverse spectroscopic
OCT (ISOCT).
The anisotropy coefficient g = cos θ is the average of the The unique advantage of ISOCT is that it is capable of imag-
cosine of the scattering angle, ing tissue structures in 3-D space, while simultaneously pro-
 2π  π viding complete quantification of the optical properties and (by
φ=0 θ =0 σ(θ, φ) cos θ sin θdθdφ
g =  2π  π . (16) assuming the WM model) the correlation functional form [35],
φ=0 θ =0 σ(θ, φ) sin θdθdφ [36]. The depth resolving capability in OCT is realized by inter-
Anisotropy determines the degree to which scattering is for- ferometry between the reference field and the sample scattering
ward directed where g = 1 corresponds to completely forward field. The primary contrast is the coherent elastic scattering from
scattering, 0 < g < 1 corresponds to scattering that is forward the tissue heterogeneity on which the WM scattering model ap-
directed, and g = 0 corresponds to isotropic scattering. Another plies. In addition, OCT requires a broadband source to eliminate
parameter commonly used to describe media such as tissue that the 2π ambiguity, providing an opportunity to analyze the scat-
are typically strongly forward scattering with values of g ≈ 0.9 tering spectrum from a scattering medium such as tissue.
is the reduced scattering coefficient, The signal in OCT can be approximated to be proportional
to the backscattering coefficient μb [see (18)]. Also the signal
μs = μs (1 − g) . (17) decay along the penetration depth is exponential and scales with
μs . Thus, the OCT image intensity can be modeled as [35]
Also of interest is the backscattering coefficient,
μb (z) −μ s ·2n z
μb = 4π · σ(θ = π). (18) I 2 (z) = rI02 Le (19)
4π
When the WM model is used, these relationships can be cal-
where I0 is the illumination intensity, r is the reflectance of the
culated analytically in terms of the model parameters. This was
reference arm, L is the temporal coherence length of the source,
shown by Sheppard [32] and later extended by Rogers et al. to
z is the penetration depth, and the mean refractive index of
include the effect of the dipole factor [31]. What is particularly
the medium is n ≈ 1.38. Further, a short-time Fourier trans-
interesting is the spectral relationship of the optical properties
form (STFT) can be performed to obtain the spatially resolved
that can be readily measured. For example, the spectral depen-
spectra, and thus, the wavelength dependence of μb and μs can
dence of μs (k) can be measured and then related to the model
be obtained. By combining the power law dependence of μb
parameter D. For example, in biological tissue, the fact that g
with the absolute values of μb and μs , the three parameters of
is large indicates that the medium contains structure large com-
the model An , D, and Ln can be determined [35]. This allows
pared to the wavelength, or kLn  1. In this regime, μs ∝ λD −4
calculation of anisotropy factor g from (16).
for D < 4.
Fig. 6 shows an example of imaging optical and physical
properties in a stratified tissue. Fig. 6(a) shows a traditional
VI. EXPERIMENTAL METHODS FOR QUANTIFICATION OCT image of rat buccal tissue. Three distinct layers can be
OF MODEL PARAMETERS
identified: keratinized epithelium, stratified squamous epithe-
There are a number of experimental methods capable of lium, and submucosa. For each layer, we assume a statistically
measuring scattering properties of tissue, for example, Go- homogeneous scattering medium and then applied the ISOCT
niometer, integrating sphere (with the IAD method), Confocal, algorithm on each layer. Fig. 6(c) and (d) shows the capability
OCT/ISOCT, and P (rs )/EBS. Of these, the first two require sig- of imaging D, the ratio of μb and μs , and the anisotropy factor
nificant tissue preparation to create a thin uniform slab of tissue g in a spatially resolved 3-D volume.
7000514 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 20, NO. 2, MARCH/APRIL 2014

Fig. 6. An example application of the model: The ISOCT method uses the
model to quantify optical and ultrastructural properties in an image map from
ex vivo rat buccal sample (adapted with permission from [35]). (a) Conven-
tional OCT images. Three layers were labeled as (1) Keratinized epithelium,
(2) stratified squamous epithelium, and (3) submucosa. (b)–(d) Pseudocolor
ISOCT images encoded with D, μ b /μ s , and g. Bar = 200 μm.

The advantage of ISOCT lies in the fact that the spatially

resolved spectrum reflects the random interference of the scat-
tering field within the resolution volume, even though the axial
resolution was sacrificed by performing STFT due to the trade-
off between the spectral bandwidth and temporal resolution. The
random interference gives rise to the granular “speckle” pattern
in the OCT B-scan image. On one hand, many treat OCT speckle
Fig. 7. Illustration of the advantage of spectral analysis in ISOCT. (a) One-
as an artifact and developed various ways to eliminate it. On the dimensional two surface reflectance located at 50 μm from the reference. Case
other hand, speckle can be informative because the random in- 1 (red): two surface separation is 100 nm. Case 2 (blue): two surface separation
terference can be modeled by Bn (rd ) for structures even below is 300 nm. (b) Simulated OCT A-line intensity from two surface separation.
(c) Spectral profile extracted at z = 50 μm by ISOCT (squares and triangles),
the image resolution limit. and beating spectra from two surfaces in solid line.
Consider a simple example where there are two reflective
surfaces along the axial direction with position of zs and zs +
Δz; the combined reflected field from these surfaces (E1 + E2 ) Further, the ISOCT spectra at z = 50 μm is numerically ex-
interferes with the reference field (Er ) which is reflected by a tracted by STFT and compared for the two cases in Fig. 7(c).
mirror located at zr . For simplicity, we arbitrarily set all the It is clear that the conventional OCT A-line signal in Fig. 7(b)
reflectance values to unity. Thus, the interference part of the has almost identical form, while the beating spectra obtained
spectral intensity is written as in ISOCT shows the difference in these two cases. While the
I(k) = 2Re(Er · E1∗ ) + 2Re(Er · E2∗ ) details for continuous media are more complex, this simple ex-
ample illustrates the basic concept that small length scales need
= 2 cos (2k(zr − zs )) + 2 cos (2k(zr − zs − Δz)) not be resolved to be quantified spectroscopically.
= 4 cos (2k(zr − zs − Δz/2)) cos(kΔz) (20)
where k is the wave number, Re(·) indicates the real part of a VII. STRUCTURAL LENGTH-SCALE SENSITIVITY
complex number, and superscript ∗ indicates the complex conju- One of the advantages of performing tissue characterization
gate. The equation shows that interference between a reference using optical spectroscopy is the ability to detect and quantify
and two reflections separated by a small phase difference (small structures smaller than the diffraction limit, even though such
z) creates spectral oscillations modulated by a beat frequency structures cannot be resolved using conventional microscopy.
determined by the two surface separation. When the separation This ability arises from the fact that scattering contrast is largely
is within the axial resolution, the OCT intensity profile cannot dependent on the spatial distribution of refractive index at struc-
differentiate (resolve) any slight changes in separation. How- tural length-scales smaller than the diffraction limit. Still, the
ever, a change in the separation that is not resolved does change question arises: which range of structural length-scales can op-
the spectra. In this way, spectral analysis provides more infor- tical spectroscopy sense?
mation about the statistical separation, which of course for a While this problem appears straight-forward, the answer is
random medium is related to the correlation function Bn (rd ). complicated in part by the fact that there is a nonlinear rela-
Fig. 7 illustrates the concept by simulating OCT and ISOCT tion between the spatial refractive index distribution nΔ (r) and
for the two surfaces. In one case, two surfaces were located σ(θ). As a result, it is not possible to independently assess the
50 μm apart from the reference and separated by 100nm; in contribution of different structural length-scales to σ(θ) in order
the second case, the separation was enlarged to 300 nm [see to determine the sensitivity range. Instead, it becomes necessary
Fig. 7(a)]. The interference spectrum is simulated from 630 to to consider the interaction between structural length-scales of
850 nm and the OCT A-line signal is reconstructed in Fig. 7(b). all sizes at once.
ROGERS et al.: MODELING LIGHT SCATTERING IN TISSUE AS CONTINUOUS RANDOM MEDIA 7000514

One way to approach such a problem is through the perturba-

tion study first presented in [37]. Under this approach, we begin
with a continuous random media specified by the WM model
(other models or actual data can easily be considered as well).
We then perturb the medium by convolving with a 3-D Gaus-
sian function in order to remove small structural length-scales
and observe how the scattering characteristics change in the
perturbed medium. This approach treats a medium as “blurred”
below a particular length scale and then determines the length
for which a notable change in optical properties occurs.
Mathematically, the volume normalized Gaussian function
can be expressed as
 3/2 
16πln(2) −4ln(2) 2
G(r) = · exp 
r (21) Fig. 8. Structural length-scale perturbation analysis for D = 3.0 and L n =
W2 W2 0.452 μm (chosen such that g = 0.9 at λ = 0.550 μm with n = 1.38).
(a) B nl (rd ). (b) Φ ls (k s ). Arrows indicate direction of increasing W l .
where W is the full-width at half-maximum. Conceptually, G(r)
represents some process (e.g., drug treatment, carcinogenesis,
etc.) which transforms the original medium by removing struc-
tures smaller than W .
Applying the convolution theorem, the modified medium can
be expressed as
nlΔ (r) = F −1 [F[nΔ (r)] · F[G(r)]] (22)
where F indicates the Fourier transform operation and the su-
perscript l indicates that lower spatial frequencies are retained.
The autocorrelation of nlΔ (r) is then [37]

Bnl (rd ) = F −1 |F[nΔ (r)]|2 · |F[G(r)]|2
 ∞
ks sin(ks r)
= 4π Φls (ks ) dks , (23)
0 r
where Φls (ks ) is the power spectral density for nlΔ (r) and can
be expressed analytically as
 2 2 Fig. 9. % change in the scattering coefficients under the structural length-scale
−k s W l perturbation analysis for D = 3.0 and L n = 0.452 μm at λ = 0.550 μm with
A l 3
Γ( D
) exp
n c 8ln (2) n = 1.38. The dotted line indicates the 5% change threshold.
Φs (ks ) = 3/2 (5−D )/2 ·
l 2
. (24)
π 2 (1 + ks2 lc2 )D /2
While (23) has no closed-form solution, it can be evaluated
(chosen such that g = 0.9 at λ = 0.550 μm with n = 1.38).
numerically in order to observe the functional form.
After perturbing this medium with Gaussian’s of different width,
Similarly, the upper bound of sensitivity can be determined
Fig. 8 shows the resulting change in shape of Bnl (rd ) and
by filtering larger structural length-scales. However, the upper
Φls (ks ).
bound is highly dependent on the shape of the correlation func-
Under the structural length-scale perturbation analysis, we see
tion, for example, on D in this model. Conceptually, the upper
a reduction in Bnl (rd ) at smaller values of rd as the width of G(r)
bound is related to the point at which the value of the corre-
widens [see Fig. 8(a)]. The location at which Bnl (rd ) diverges
lation function is insignificant in the presence of the rest of
from Bn (rd ) occurs at ∼ Wl . Conceptually, this indicates a
the correlation function. This point depends on the magnitude
loss of correlation at structural length-scales smaller than Wl .
of the function at small length scales. Further, introduction of
After Fourier transformation of Bnl (rd ) to obtain Φls (ks ), this
large length scales that scatter significantly will always affect
loss of short structural length-scales correlation corresponds to
the scattering measurements, but their presence will also change
a reduction in scattered power at higher spatial light frequencies
the correlation function at all smaller length scales, and so can-
[see Fig. 8(b)].
not be considered a perturbation. It is therefore not particularly
In order to relate the curves in Fig. 8 to observable light
meaningful to characterize the upper limit.
scattering characteristics, we first evaluate σ(θ, φ) according to
(3) and then calculate μs , μb , g, and μs according to (15)–(18).
A. Length Scale Sensitivity of EBS Fig. 9 shows the percent change in the four optical properties
As a demonstration of the shape of the functions described under the structural length-scale perturbation analysis. Defining
by (23) and (24), we begin with a general approximation of tis- a 5% change level as the threshold for detecting an alteration in
sue structure using D = 3 (i.e., HG case) and Ln = 0.452 μm the measured optical properties, we find that μs is sensitive to
7000514 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 20, NO. 2, MARCH/APRIL 2014

Fig. 11. Rendering the continuous random media used in the structural length
scale sensitivity analysis. (a) Medium realization with 1.4 NA microscope res-
olution at λ = 0.4 μm. (b) Medium realization corresponding to the sensitivity
range of P (rs , μ s ) (where pixel size corresponds to the sensitivity limit).
Fig. 10. Change in the shape of P (rs · μ s ) under the structural length-scale
perturbation analysis for D = 3.0 and L n = 0.452 μm at λ = 0.550 μm with
n = 1.38. Arrows indicate the direction of increasing W l .
cally measured medium would at best result in a blurred version
TABLE I of the actual medium with features on the order of λ/2. Fig. 11(a)
STRUCTURAL LENGTH-SCALE SENSITIVITY OF VARIOUS SCATTERING provides a demonstration of the range of length scales that a
CHARACTERISTICS IN CONTINUOUS RANDOM MEDIA
conventional microscope would visualize for a single random
medium realization. For this same realization, Fig. 11(b) shows
the range of length scales that the EBS method would detect
using a spatial resolution corresponding to the lower sensitivity
of P (rs , μs ). While the EBS method does not directly resolve
the smallest length scales, changes in the shape of P (rs , μs )
nonetheless provide a means to quantify fine structure that is
length scales above 0.082 μm, μs is sensitive above 0.029 μm, not possible with standard imaging methods.
and μb is sensitive above 0.017 μm. We note that the sensitivities provided in this section assume
Next, we extend this analysis to quantify the sensitivity of that the experimental instrument used to measure such scattering
the shape of P (rs ). In order to do this, we performed a series characteristics has sufficiently high signal-to-noise ratio to dis-
of MC simulations using the open source code detailed in [38]. tinguish a 5% change in value. Given a specific instrument noise
For these simulations, we implemented the modified scattering level, the analysis presented in this section can be repeated with a
phase functions described by Φls (ks ) and tracked the reflected different threshold value to determine the structural length-scale
intensity in bins from rs · μs = 0.005 to 5 with 0.005 resolution. sensitivity of that particular instrument.
Fig. 10 shows the shape of P (rs · μs ) calculated under the In the above numerical analysis, we used a particular oper-
structural length-scale perturbation analysis. The curves are ation by convolving the medium with 3-D Gaussian function
plotted as a function of the unitless rs · μs in order to isolate the to smooth window. The rationale is that suppression of small
change in shape from the change in μs . There is a prominent length scales without otherwise altering the correlation func-
change in the shape of P (rs · μs ) for values of rs · μs less than 1. tional form provides a reasonable comparison. However, it is
It is well known that within this regime P (rs · μs ) is extremely not be the only way to introduce the perturbation. Other meth-
sensitive to the shape of the phase function [39]. We determine ods such as adding random noise or changing the functional
the structural length-scale sensitivity limit of P (rs · μs ) by find- shape may be more appropriate depending on the application
ing the values of Wl for which the value of P (rs · μs ) at any and can be performed to assess the length scale sensitivity.
rs · μs changes by more the 5%. Using this criterion, P (rs · μs ) Experimental confirmation of the sensitivities provided in this
is sensitive to length scales above 0.018 μm. section is dependent on the fabrication of a continuous medium
A summary of the structural length-scale sensitivity of var- model whose distribution of length-scales is well controlled.
ious scattering characteristics is shown in Table I. Assuming Such phantoms have proven difficult to construct. As such, in
the best-case scenario of conventional optical microscopy with the following section, we use a superposition of polystyrene
violet wavelength illumination (λ = 0.400 μm), the diffraction micro-spheres to simulate a power-law distribution of length-
limit is at ∼λ/2 = 0.200 μm. In each case, quantification of the scales.
various scattering characteristics provides sensitivity to struc-
tures smaller than the diffraction limit.
To visualize the meaning of the length scale sensitivity, a B. Length Scale Sensitivity of ISOCT
continuous random medium that has a desired Bn (rd ) can be ISOCT can be used to measure the value of D by determining
generated using the method described by Capoglu [8]. Since the power law dependence of μb on wavelength. In this case,
conventional microscopy is diffraction limited, the microscopi- the length scale sensitivity is not exactly the same as that of
ROGERS et al.: MODELING LIGHT SCATTERING IN TISSUE AS CONTINUOUS RANDOM MEDIA 7000514

Fig. 12. Sensitivity analysis for the dependence of the measured value of D
on changes at subdiffractional length scales. Phantoms were made by mixing
a suspension of microspheres spanning a range of diameters. (a) The volume
fraction of each sphere size present in the suspension forms a power law relative
to the diameters. (b) The measured value of D for each phantom is plotted
against the value of the smallest sphere diameter present in that phantom. Error
bars indicate standard error of measurements for each phantom. Value predicted
using Mie theory is also shown. Adapted with permission from [36].

μb itself. If Bn (rd ) is perturbed as before by removing small Fig. 13. OCT and tomographic D maps from phantoms. Gray scale OCT
lengths scales, the value of μb drops quickly, but the spectral image (a) and pseudocolor D map (b) in low length scale sensitivity studies.
dependence changes more slowly requiring a change at slightly Bar = 200 μm. Adapted with permission from [36].
larger length scales to produce a measurable change in D us-
ing ISOCT, although other methods of measuring D may have
different sensitivity.
To demonstrate this experimentally, a phantom study consist- hibit no discernible differences, the increase in D is readily
ing of various sizes of polystyrene microspheres was designed appreciated for changes of length scale at 60 nm (No. 3) and
to verify the nanometer scale sensitivity. The volume fraction 80 nm (No. 4) in Fig. 13(b). These phantoms also highlight the
of the spheres is a power law relationship with their diame- fact that Bn (rd ) is a statistical property of the medium. Locally,
ters as in Fig. 12(a). To determine the smallest length scale to voxels will have a deterministic correlation function that differs
which D measured with ISOCT is sensitive, a series of phan- from adjacent voxels. The value of a particular pixel is not par-
toms was constructed by successively excluding the smallest ticularly meaningful. But upon averaging, the local correlation
microspheres included in the previous phantom. For example, functions converge to the expected value. This is the reason for
the first phantom (No.1) consisted of microspheres from 30nm the spatial variation seen in the statistically homogeneous sphere
to 1 μm. The second phantom (No. 2) consisted of microspheres suspension shown here. Spatial averaging results in more pre-
from 40 nm to 1 μm by excluding 30 nm size spheres. The third cise values, but at the cost of spatial resolution. Ideally, imaging
phantom (No. 3) consisted of microspheres from 60 nm to 1 μm is used as in Fig. 6 to map tissue structure, and then, contigu-
by excluding 30 and 40 nm spheres, and so on. ous regions or layers are segmented to report average values of
The values of D were measured with ISOCT and compared optical properties within a region or layer.
with Mie theory prediction as in Fig. 12(b). As expected, an The nanometer scale sensitivity of ISOCT in biological tis-
increase in the measured value of D was observed as smaller sue was previously demonstrated using scanning electron mi-
sphere sizes were excluded. When 40 nm spheres were removed, croscopy (SEM) [36]. As shown in Fig. 14(a), high-resolution
D was measured to be 1.78 ± 0.06 compared to that from the SEM images were taken from ex vivo human rectal mucosa as
first phantom (No. 1) 1.57 ± 0.04. The increase of 0.21 is above well as ISOCT measurements shown in Fig. 14(b). Two types
the experimental uncertainty of ±0.2 for measuring D, which of tissue components were identified: epithelium (Epi) and lam-
corroborates that the perturbation at ≈40 nm can be detected ina propria (LP). Magnified regions of each are shown in the
by measuring D with ISOCT, a value well below the resolu- insets. The image correlation functions calculated from these
tion limit for OCT. The error bars in Fig. 12(b) indicate the SEM image regions show that Epi has sharper functional form
standard error for the measurements of a particular phantom, which coincides with increased presence of small features in
while the experimental uncertainty of ±0.2 corresponds to the Fig. 14(c). Meanwhile, ISOCT measured the depth dependent
90% confidence interval for linear regression described in earlier D and Ln can be plugged back in to Bn (rd ) and plotted as in
work [35]. Fig. 14(d). Again, while SEM images do not represent refractive
The pseudocolor encoded OCT images from above phan- index, this comparison can only be made qualitatively. However,
tom studies demonstrated the advantages of the ISOCT method. the result is in qualitative agreement with SEM findings: the Epi
While the conventional OCT B-scan images in Fig. 13(a) ex- has a sharper correlation function than LP.
7000514 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 20, NO. 2, MARCH/APRIL 2014

is correlated. Once the refractive index correlation function can

be determined independently, models of scattering can be based
on a true correlation function for tissue. However, when the
function cannot be independently determined, the WM model
may provide a good alternative.
Several experimental methods have been developed that em-
ploy the WM model to quantify the optical scattering proper-
ties. These methods are shown to be sensitive to length scales
an order of magnitude smaller than the diffraction limit. While
these methods do not resolve length scales on the order of tens
of nanometers, they do provide a means to quantify changes
that occur in tissue at these lengths scales. This is of partic-
ular significance in diagnostic applications. For example, it is
well known that morphological changes occur in tissue during
carcinogenesis, but mounting evidence shows that these gross
changes are presaged by subtle changes in ECM and chromatin
organization [8].

ACKNOWLEDGMENT
The authors would like to thank Sam Norris for the SEM
image.

REFERENCES
[1] B. Huang, M. Bates, and X. Zhuang. (2009). Super-resolution fluorescence
microscopy. Annu. Rev. Biochem. [Online]. 78(1), pp. 993–1016. Ava-
ilable: http://www.annualreviews.org/doi/abs/10.1146/annurev.biochem.
77.061906.092014
[2] V. V. Tuchin, Optical Clearing of Tissues and Blood. Bellingham, WA,
USA: SPIE Press, 2006.
[3] D. A. Boas, D. H. Brooks, E. L. Miller, C. A. DiMarzio, M. Kilmer,
R. J. Gaudette, and Q. Zhang, “Imaging the body with diffuse optical
Fig. 14. ISOCT measurement from human colon biopsy. (a) SEM image of tomography,” IEEE Signal Process. Mag., vol. 18, no. 6, pp. 57–75, Nov.
a colon cross section. The resolution is 40 nm. Bar = 10 μm. (b) The ISOCT 2001.
measurement on D ± SE and L n ± SE as a function of penetration depth [4] B. Wilson, M. Patterson, and L. Lilge, “Implicit and explicit dosimetry in
where SE is standard error. The boundary between the epithelial cells and the photodynamic therapy: A new paradigm,” Lasers Med. Sci., vol. 12, no. 3,
collagen network is around 50 μm from the surface. (c) and (d) A qualitative pp. 182–199, 1997.
comparison between the correlation function obtained by the SEM image (c) and [5] R. E. Hausman and G. M. Cooper, The Cell: A Molecular Approach.
ISOCT (d). The 2-D image autocorrelation function (ACF) ±SE from SEM is Washington, DC, USA: ASM, 2004.
calculated from different regions on the image with image dimension 5 × 5 μm. [6] J. M. Schmitt and G. Kumar. (1996, Aug.). Turbulent nature of refractive-
The ISOCT index correlation functions were calculated using averaged value of index variations in biological tissue. Opt. Lett. [Online]. 21(16), pp. 1310–
D and L n from Epi and LP to reconstruct B n (rd ). Epi: epithelium, LP: lamina 1312. Available: http://ol.osa.org/abstract.cfm?URI=ol-21-16-1310
propria. Adapted with permission from [36]. [7] M. Xu and R. R. Alfano. (2005, Nov.). Fractal mechanisms of light scat-
tering in biological tissue and cells. Opt. Lett. [Online]. 30(22), pp. 3051–
3053. Available: http://ol.osa.org/abstract.cfm?URI=ol-30-22-3051
[8] A. Wax and V. Backman, Biomedical Applications of Light Scattering.
VIII. CONCLUSION New York, NY, USA: McGraw-Hill (McGraw-Hill biophotonics Series),
2009.
Models of light scattering in tissue are important for a num- [9] H. Davies, M. Wilkins, J. Chayen, and L. La Cour, “The use of the interfer-
ber of applications including imaging, therapy, and diagnosis. ence microscope to determine dry mass in living cells and as a quantitative
Many models have been previously used with good success, cytochemical method,” Q. J. Microsc. Sci., vol. 3, no. 31, pp. 271–304,
1954.
the most common being the HG phase function. Several groups [10] R. Barer and S. Tkaczyk, “Refractive index of concentrated protein solu-
have shown evidence that tissue is organized in a way that can tions,” Nature, vol. 173, pp. 821–822, 1954.
be described as a mass fractal. A model based on the WM cor- [11] J. R. Mourant, J. P. Freyer, A. H. Hielscher, A. A. Eick, D. Shen, and T. M.
Johnson, “Mechanisms of light scattering from biological cells relevant
relation family has the advantage of being flexible and actually to noninvasive optical-tissue diagnostics,” Appl. Opt., vol. 37, no. 16,
includes many previously used models as special cases. pp. 3586–3593, Jun. 1998.
The model is based on the weak scattering approximation [12] L. T. Perelman, V. Backman, M. Wallace, G. Zonios, R. Manoharan, A.
Nusrat, S. Shields, M. Seiler, C. Lima, T. Hamano, I. Itzkan, J. Van Dam,
(RGD) which allows a the intensity scattering function from J. M. Crawford, and M. S. Feld, “Observation of periodic fine structure in
random media to be calculated by taking the Fourier transform reflectance from biological tissue: A new technique for measuring nuclear
of refractive index correlation function. This approximation can size distribution,” Phys. Rev. Lett., vol. 80, pp. 627–630, Jan. 1998.
[13] A. Wax, C. Yang, V. Backman, M. Kalashnikov, R. R. Dasari, and
be made for clouds of particles or continuous random media M. S. Feld, “Determination of particle size by using the angular distri-
provided that the scattered field is weak compared to the incident bution of backscattered light as measured with low-coherence interferom-
field over paths longer than the distance over which the medium etry,” J. Opt. Soc. Amer. A, vol. 19, no. 4, pp. 737–744, 2002.
ROGERS et al.: MODELING LIGHT SCATTERING IN TISSUE AS CONTINUOUS RANDOM MEDIA 7000514

[14] G. Mie, “Beiträge zur optik trüber medien, speziell kolloidaler met- simulation of coherent backscattering in biological media containing bire-
allösungen,” Annalen der Physik, vol. 330, no. 3, pp. 377–445, 1908. fringence,” J. Biomed. Opt., vol. 17, no. 11, pp. 115 001–115 001, 2012.
[15] C. F. Bohren and D. R. Huffman, Absorption and Scattering of Light by [39] E. Vitkin, V. Turzhitsky, L. Qiu, L. Guo, I. Itzkan, E. B. Hanlon, and L.
Small Particles. New York, NY, USA: Wiley, 1983. T. Perelman, “Photon diffusion near the point-of-entry in anisotropically
[16] L. G. Henyey and J. L. Greenstein, “Diffuse radiation in the Galaxy,” scattering turbid media” Nature Commun., vol. 2, p. 587, p. 587, 2011.
Astrophys. J., vol. 93, pp. 70–83, Jan. 1941.
[17] S. Jacques, C. Alter, and S. Prahl, “Angular dependence of HeNe laser
light scattering by human dermis,” Lasers Life Sci., vol. 1, no. 4, pp. 309–
333, 1987.
[18] J. R. Mourant, T. Fuselier, J. Boyer, T. M. Johnson, and I. J. Bigio, “Pre- Jeremy D. Rogers received the B.S. degree in physics
dictions and measurements of scattering and absorption over broad wave- from Michigan Technological University, Houghton,
length ranges in tissue phantoms,” Appl. Opt., vol. 36, pp. 949–957, 1997. MI, USA, in 1999, and the M.S. and Ph.D. degrees
[19] J. Joseph, W. Wiscombe, and J. Weinman, “The Delta-Eddington approx- in optical sciences from the College of Optical Sci-
imation for radiative flux transfer,” J. Atmospher. Sci., vol. 33, no. 12, ences, University of Arizona, Tucson, AZ, USA, in
pp. 2452–2459, 1976. 2003 and 2006, respectively. He completed Postdoc-
[20] S. A. Prahl, “Light transport in tissue,” Ph.D. dissertation, Univ. Texas at toral training and continued as Research Assistant
Austin, Austin, TX, USA, Dec. 1988. Professor in biomedical engineering at Northwestern
[21] E. L. Hull and T. H. Foster. (2001, Mar.). Steady-state reflectance spec- University, Evanston, IL, USA. He is currently an
troscopy in the p3 approximation. J. Opt. Soc. Amer. A [Online]. 18(3), Assistant Professor in biomedical engineering at the
pp. 584–599. Available: http://josaa.osa.org/abstract.cfm?URI=josaa-18- University of Wisconsin, Madison, WI, USA. His re-
3-584 search interests include theoretical and numerical modeling of light scattering
[22] A. J. Radosevich, N. N. Mutyal, V. Turzhitsky, J. D. Rogers, J. Yi, as well as lens design and development of instrumentation for basic research
A. Taflove, and V. Backman. (2011, Dec.). Measurement of the spatial and application to optical metrology of cells and tissue for diagnostics.
backscattering impulse-response at short length scales with polarized en-
hanced backscattering. Opt. Lett. [Online]. 36(24), pp. 4737–4739. Avail-
able: http://ol.osa.org/abstract.cfm?URI=ol-36-24-4737
[23] T. J. Farrell, M. S. Patterson, and B. Wilson, “A diffusion theory model Andrew J. Radosevich received the B.S. degree in
of spatially resolved, steady-state diffuse reflectance for the noninvasive biomedical engineering from Columbia University,
determination of tissue optical properties in vivo,” Med. Phys., vol. 19, New York, NY, USA, in 2009. He is currently work-
pp. 879–888, 1992. ing toward the Ph.D. degree in biomedical engineer-
[24] A. Ishimaru, Wave Propagation and Scattering in Random Media. Pis- ing at Northwestern University, Evanston, IL, USA.
cataway, NJ, USA: IEEE Press, 1997. His research interests include the numerical modeling
[25] H. C. Hulst and H. Van De Hulst, Light Scattering: By Small Particles. and experimental observation of enhanced backscat-
New York, NY, USA: Dovers, 1957. tering spectroscopy for use in the early diagnosis and
[26] M. Born and E. Wolf, Principles of Optics: Electromagnetic Theory of treatment of colon and pancreatic cancers. Mr. Rado-
Propagation, Interference and Diffraction of Light, 7th ed. Cambridge, sevich received a three-year National Science Foun-
U.K.: Cambridge Univ. Press, Oct. 1999. dation graduate research fellowship in 2011.
[27] J. Ramella-Roman, S. Prahl, and S. Jacques, “Three Monte Carlo programs
of polarized light transport into scattering media—Part I,” Opt. Exp.,
vol. 13, no. 12, pp. 4420–4438, 2005.
[28] V. Tuchin, Handbook of Coherent-Domain Optical Methods: Biomedi-
Ji Yi received the B.S. degree from Tsinghua Uni-
cal Diagnostics, Environmental Monitoring, and Materials Science. New
versity, Beijing, China, in 2005, and the Ph.D. degree
York, NY, USA: Springer-Verlag, 2013.
from Northwestern University, Evanston, IL, USA,
[29] İlker R. Çapoğlu, J. D. Rogers, A. Taflove, and V. Backman, “Accuracy of
in 2012, both in biomedical engineering. He is cur-
the born approximation in calculating the scattering coefficient of biologi-
rently a joint Postdoctoral Fellow in Hao F. Zhang
cal continuous random media,” Opt. Lett., vol. 37, no. 17, pp. 2679–2681,
and Vadim Backman’s laboratory. His research inter-
Sep. 2009.
ests include using optical coherence tomography to
[30] P. Guttorp and T. Gneiting, “On the Whittle-Matérn correlation family,”
quantify tissue ultrastructural properties and explore
NRCSE-TRS, vol. 80, 2005.
functional contrast from OCT. He is also interested
[31] J. Rogers, İ. Çapoğlu, and V. Backman, “Nonscalar elastic light scattering
in optical biosensing for tumorous cell detection and
from continuous random media in the Born approximation,” Opt. Lett.,
ultraresolution multimodal imaging techniques.
vol. 34, no. 12, pp. 1891–1893, 2009.
[32] C. J. R. Sheppard, “Fractal model of light scattering in biological tissue
and cells,” Opt. Lett., vol. 32, no. 2, pp. 142–144, 2007.
[33] G. Hall, S. L. Jacques, K. W. Eliceiri, and P. J. Campagnola, “Goniometric
measurements of thick tissue using Monte Carlo simulations to obtain the Vadim Backman received the Ph.D. degree in medi-
single scattering anisotropy coefficient,” Biomed. Opt. Exp., vol. 3, no. 11, cal engineering from Harvard University, Cambridge,
pp. 2707–2719, 2012. MA, USA, and the Massachusetts Institute of Tech-
[34] S. L. Jacques, B. Wang, and R. Samatham, “Reflectance confocal mi- nology, Cambridge, in 2001. He is currently a Profes-
croscopy of optical phantoms,” Biomed. Opt. Exp., vol. 3, no. 6, pp. 1162– sor of biomedical engineering at Northwestern Uni-
1172, 2012. versity, Evanston, IL, USA, a Program Leader in can-
[35] J. Yi and V. Backman, “Imaging a full set of optical scattering properties of cer bioengineering, nanotechnology, and chemistry
biological tissue by inverse spectroscopic optical coherence tomography,” at the Robert H. Lurie Comprehensive Cancer Insti-
Opt. Lett., vol. 37, no. 21, pp. 4443–4445, Nov. 2012. tute, Chicago, IL, and a member of the Professional
[36] J. Yi, A. J. Radosevich, J. D. Rogers, S. C. Norris, İlker R. Çapoğlu, A. Staff in the Division of Gastroenterology, NorthShore
Taflove, and V. Backman, “Can OCT be sensitive to nanoscale structural University Health-Systems, Evanston, IL. He is in-
alterations in biological tissue?” Opt. Exp., vol. 21, no. 7, pp. 9043–9059, volved in translational research, which is focused on bridging these techno-
Apr. 2013. logical and biological innovations into clinical practice. His research interests
[37] A. J. Radosevich, J. Yi, J. D. Rogers, and V. Backman, “Structural length- include biomedical optics, spectroscopy, microscopy, development of analyt-
scale sensitivities of reflectance measurements in continuous random me- ical approaches to describe light transport in biological media, and optical
dia under the born approximation” Opt. Lett., vol. 37, no. 24, pp. 5220– microscopy for nanoscale cell analysis. Dr. Backman has received numerous
5222, Apr. 2012. awards, including being selected as one of the top 100 young innovators in the
[38] A. J. Radosevich, J. D. Rogers, I. R. Capoğlu, N. N. Mutyal, P. Pradhan, world by the Technology Review Magazine and the National Science Foundation
and V. Backman, “Open source software for electric field Monte Carlo CAREER Award.