DIFFUSE OPTICAL TOMOGRAPHY AND SPECTROSCOPY OF BREAST CANCER AND FETAL BRAIN Regine Choe A Dissertation in Physics and Astronomy Presented to the Faculties of the University of Pennsylvania in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy 2005 Arjun G. Yodh Supervisor of Dissertation Randall D. Kamien Graduate Group Chairperson
248
Embed
DIFFUSE OPTICAL TOMOGRAPHY AND SPECTROSCOPY OF BREAST CANCER AND
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
DIFFUSE OPTICAL TOMOGRAPHY AND
SPECTROSCOPY OF BREAST CANCER AND FETAL
BRAIN
Regine Choe
A Dissertation
in
Physics and Astronomy
Presented to the Faculties of the University of Pennsylvania in Partial
Fulfillment of the Requirements for the Degree of Doctor of Philosophy
5.15 Comparison between homogeneous and two-layer model fit . . . . . . . . . . . . . 177
xxii
Chapter 1
Introduction
1.1 Diffuse Optical Spectroscopy and Tomography
Diffuse optical tomography (DOT) and spectroscopy (DOS) are non-invasive techniques used to
measure the optical properties of physiological tissue. In the near-infrared (NIR) spectral window
of 600 - 1000 nm, photon propagation in tissues is dominated by scattering rather than absorp-
tion. Photons experience multiple scattering events as they propagate deeply into tissue (up to
10 cm). The primary chromophores in this spectral window are oxygenated hemoglobin (HbO2),
deoxygenated hemoglobin (Hb), water (H2O) and lipid. Each chromophore possesses a distinct
spectrum as shown in Fig. 1.1(a). A weighted sum of the contributions from each chromophore
which correponds approximately to the tissue absorption coefficient (µa) is also shown. Here,
the concentration of each chromophore is adjusted to a value typically found in breast tissue: the
concentration of oxygenated hemoglobin (CHbO2) is ∼24 µM, the concentration of deoxygenated
hemoglobin (CHb) as ∼6 µM, and the tissue is assumed to contain a 31 % water and a 57 % lipid
contribution. This combination leads to total hemoglobin concentration, THC = CHb + CHbO2,
1
of 30 µM, and a blood oxygen saturation, StO2 = CHbO2/THC, of ∼80% (see Chapter 2 for
more details). Notice that absorption measurements at multiple wavelengths are required to extract
the concentration of each chromophore. Fig. 1.1(b) shows a typical scattering spectra found in the
breast tissue. The tissue scattering depends on the photon random walk step in the medium. The
reduced scatttering coefficient (µ′s) which is the reciprocal of the photon random walk step length,
is often modeled within a simplified Mie-scattering approximation [185, 188], i.e. µ′s(λ) = Aλ−b
where λ is the light wavelength. Notice that scattering is 100× larger than the absorption, and the
NIR scattering spectra is relatively flat in the near infrared.
650 700 750 800 850 900 950 10000
0.05
0.1
0.15
0.2
wavelength
µ a
HbO2HbH2Olipidtotal
(nm)
(cm
−1)
(a)
650 700 750 800 850 900 950 10000
2
4
6
8
10
12
wavelength
µ s′
(nm)
(cm
−1)
(b)
Figure 1.1: Spectrum of absorption and scattering in the NIR. (a) Spectra of major chromophoreswith adjusted concentrations found in typical breast tissue: oxygenated hemoglobin (HbO2) anddeoxygenated hemoglobin (Hb), water and lipid. (b) Scattering spectra. See text for details.
Accurate retrieval of tissue properties based on the DOS and DOT measurements requires that
absorption and the scattering be decoupled. A light transport model based on the diffusion approx-
imation [121] is widely used to describe photon propagation in the NIR. Using optical measure-
ments at multiple source-detector positions on the tissue surfaces, one can reconstruct the internal
2
distribution of the absorption coefficient (µa) and the reduced scattering coefficient (µ′s) in three-
dimensions based on the transport model. Physiological images of total hemoglobin concentration,
blood oxygenation, water and lipids are then derived from this information. Thus far DOT and DOS
have generated a lot of scientific interest and have been applied in various deep-tissue applications
including breast cancer imaging [62, 69, 102, 116, 119, 144, 165, 191, 194, 213, 222, 300], brain
exploration of physiological parameters (multiple wavelength, correlation with histo-pathology
[222, 278]), (3) co-registration with other imaging modalities and (4) therapy monitoring.
1.2.3 Breast optical imaging in our laboratory
During the last decade our laboratory has made substantive contributions to problems in dif-
fuse optical tomography. We published the first experiments demonstrating tomographic recon-
struction of absorption and scattering heterogeneities [198], we were among the first groups to
show how to reconstruct the lifetime and concentration of fluorophores in tissues [199], and we
introduced a methodology whereby diffused temporal light field correlation functions are used
to reconstruct heterogeneous dynamical flow properties [24, 29]. In a different vein, we have
investigated the fundamental resolution and characterization limitations of diffuse optical tech-
niques [28, 166, 168], and we have combined the optical methods with other imaging modalities
such as ultrasound [130,299] and MRI [192–195] so structural information from the second imag-
ing modality may be optimally coupled with the optical technique to improve the accuracy of as-
signed tissue optical properties [200]. We also continue to develop theoretical tools for DOT which
have, for example, elucidated the resolution and sensitivity trade-offs associated with specific ge-
ometries and instrumentation [65], enabled experimenters to choose optimal source wavelengths
for clinical experiments [58], and clarified the conditions for which differential measurements are
well suited for quantitative tomography [192]. Finally we have consistently pushed DOT towards
11
in vivo applications, for example, in brain hemodynamics and metabolism [63, 88], in tumor biol-
ogy [182], and in breast tumor detection and characterization.
Most recently (and relevant for this thesis), we have successfully assembled a parallel-plate
soft-compression apparatus for diffuse optical tomography of breast, and we have begun clinical
studies with the instrument at the Hospital of the University of Pennsylvania (HUP) [49, 62, 81].
Our second generation parallel-plate instrumentation is unique in the community: our sources and
detectors are based predominantly on continuous-wave (CW) light, our CCD-based detection is
simple and massively parallel compared to all other instrumentation, we collect multiple views
of the breast in the device and so execute and produce full three-dimensional reconstructions,
and our soft compression parallel-plate geometry is attractive for comparison to MRI and x-ray
mammography as well as for maximizing hemodynamic contrast. Two sensible figures of merit
for DOT imaging systems are total data, and data rate (defined as a product of source-number times
detector-number times number-of-wavelengths divided by total-acquisition-time). Our current data
rates exceed those of all existing instruments, in most cases by 100× (See Table 1.1).
With this instrumentation, we demonstrated theoretically and experimentally that it is possible
to carry out 3D diffuse optical tomography with CW light. We showed that DOT reconstructions
based on measurements at multiple optical wavelengths (simultaneously) enable experimenters to
efficiently and uniquely separate scattering from absorption, as well as the contributions of one
chromophore from another chromophore [57, 58]. Our clinical pilot studies show that tumors
are detectable and characterizable with the diffuse optical method (Chapter 4, Section 4.4); they
can even be tracked during neoadjuvant chemotherapy [49] (Chapter 4, Section 4.5). In light of
searching for additional optical contrast, we show that blood flow measurement by optical method
is feasible and show distinction among different breast tumors (Chapter 4, Section 4.6).
12
Group type Ns Nd Nλ Ntotal τ data rate(Hz)this thesis (2nd gen.) CW 48 2.28×105 6 6.5×107 8.4 min 6.6×107
Schmitz et al [239] CW 25 32 4 3.2×103 0.32 s 8640Colak et al [54] CW 225 225 3 1.5×105 6 min 540
Iftimia et al [134] CW 64 64 3 1.2×104 12 min 17McBride et al [181] FD 16 15 × 3 6 4.3×103 30 s 42
Li et al [165] FD 40 9 1 3.6×102 1.5 min 4Chen et al [44] FD 12 8 2 1.9×102 NA NA
Godavarty et al [109] FD 27 128 1 3.5×103 NA NAFranceschini et al [102] FD 1 2340 2 4.6×103 3 min 2.6Ntziachristoset al [194] TD 24 8 1 1.9×102 80 s 2.4
Cubeddu et al [258] TD 1 3000 4 1.2×104 5 min 40Rinneberg et al [115] TD 1 2000 2 4.0×103 5 min 13Schmidt et al [237] TD 32 32 1 1.0×103 ∼10 min 1
Table 1.1: Comparison of instruments used in the photon migration field in terms of informationcontent. Ns: number of sources, Nd: number of detectors, Nλ: number of wavelength, Ntotal:total number of data (= NsNdNλ), τ : acquisition time. Data rate is defined as Ntotal/τ . (CW:continuous-wave, FD: Frequency-domain, TD: Time-domain method)
1.3 Fetal Brain Oxygenation Monitoring
Hypoxic-ischemic damage to the fetal brain can result in permanent neurodevelopmental impair-
ment or death [257, 263, 264]. Early detection of fetal cerebral hypoxic-ischemia is thus important
for timely intervention. Current non-invasive antepartum screening and diagnostic tools for fetal
well-being in utero include the non-stress test (fetal heart rate monitoring) and the biophysical pro-
file (fetal heart rate monitoring and ultrasound). Fetal heart rate monitoring, however, probes fetal
cerebral hemodynamics and oxygenation indirectly, and has a high false-positive rate [161, 173].
This has led to an increased number of unnecessary Cesarean sections and premature deliver-
ies [223]. Clearly, the development of devices to directly monitor fetal cerebral oxygenation and
hemodynamics could improve the specificity of antepartum tests. Near-infrared (NIR) diffuse
optical spectroscopy has the potential to non-invasively monitor fetal cerebral oxygenation and
hemodynamics in utero. NIR light is non-ionizing and the power levels used are harmless to the
13
body, making this technology safe under continuous exposure. Additionally, NIR technology can
be designed to be fast and portable and is therefore suitable in a clinical setting.
A few researchers have developed and employed NIR oximeters to monitor neonatal cerebral
oxygenation [45,77,122,127,158]. More recently, a trans-vaginal NIR fetal oximeter [73,103] has
been developed, but can only be used during labor after the amniotic membrane has ruptured.
Development of a non-invasive transabdominal NIR fetal oximeter is challenging, but if suc-
cessful could provide a direct assessment of fetal cerebral oxygenation and hemodynamics be fore
labor and delivery. The feasibility of transabdominal NIR continuous wave (CW) spectroscopy was
first explored by Ramanujam et al [227, 228] during a non-stress test. Subsequently, Zourabian et
al developed a transabdominal NIR CW oximeter with the capability to detect fetal arterial pulses
in utero [301]. In addition, theoretical and experimental tissue phantom investigations have been
performed to understand NIR photon diffusion through the fetal brain in utero [138, 270]. Col-
lectively, these studies suggest transabdominal NIR photon diffusion measurements through the
fetal brain in utero are possible. However, due to limitations of the instrumentation and analytical
models employed in these studies, it has not as yet been possible to quantify fetal cerebral blood
saturation or blood volume in utero.
To demonstrate the feasibility of transabdominal NIR spectroscopy for detecting and quantify-
ing fetal hypoxia in utero, we designed a fetal hypoxia model using pregnant ewe. We have built
a multi-wavelength NIR frequency-domain instrument with the capability to perform NIR pho-
ton diffusion measurements through tissue over a wide range of source-detector separations and a
two-layer numerical diffusion model to accurately quantify the fetal cerebral blood saturation. We
have shown good agreement between fetal blood saturation determined by the transabdominal NIR
method, and arterial and venous fetal blood saturation quantified from fetal blood samples using a
where ξs and θs are the amplitude and phase of the source coupling and ξd and θd are the amplitude
and phase of the detector coupling. In the multi-spectral method, the unknowns are chromophore
concentrations and the scattering factors. For instance, when the optical properties of the homo-
geneous liquid phantom made with India ink and Intralipid (as described in Chapter 3, Section
3.8.1) are to be recovered, the unknowns are Cink, A and b. When the bulk optical properties of
25
physiological tissue (with the assumption of homogeneous medium) is under investigation, the un-
knowns are CHb, CHbO2, CH2O, Clipid, A and b. Since the calculated fluence rate Φc(λ, r) is often
described in terms of µa and µ′s, one needs to decompose the initial guess using Equation 2.4 to
solve the forward problem. (See Appendix Section 3.8.1 for extinction coefficients.)
positive source
extrapolatedboundary
negative image
source fiberdetector fiber
z=0z0
zb
zbz0+
z
n in
outn
z=-zp
z=-zb
Figure 2.3: Source and image configurations for extrapolated boundary condition.
The analytic solutions of the forward problem for semi-infinite and slab geometry with extrap-
olated boundary conditions are widely used in most DOS analyses. For the extrapolated boundary
condition [5, 76, 94, 121], the fluence rate is set to zero at an extrpolated boundary located at a
distance zb outside the medium i.e. Φ(ρ, z = −zb) = 0. The method of images depicted in Figure
2.3 can be used to construct the solution by placing a negative image source at the opposite side
of the extrapolated boundary. In this extrapolated boundary case, the position of negative image
source is at −zp = −(z0 + zb) since the source is placed at z0 (1/µ′s inside the medium). The
configuration of source and image for the semi-infinite and slab geometry is depicted in Figure 2.4.
The analytic frequency-domain solution for the semi-infinite geometry (Figure 2.4(a)) with
26
z=0
z=dz=2dz=3d
z=4d
z=-2dz=-d
z=-3d
z=-zz=0z=z 0
p
z
(a) (b)positive imagenegative image
sourcesource
Figure 2.4: Source and image configurations for measurement Geometry of (a) Semi-infinitemedium and (b) Slab.
extrapolated boundary conditions is [217, 218]
Φ(ρ, z) =vS
4πD
(
exp(ik√
ρ2 + (z − z0)2√
ρ2 + (z − z0)2− exp(ik
√
ρ2 + (z + zp)2√
ρ2 + (z + zp)2
)
, (2.7)
where k2 = −vµa+iωD , z0 = 1
µ′s, zp = z0 + 2zb, zb = 2
3µ′s
1+Reff1−Reff
, Reff = −1.44/n2 + 0.71/n +
0.668 + 0.064n. z is the axis perpendicular to the medium, and ρ is the radial distance parallel
to the medium, as can be seen in Figure 2.3 and Figure 2.4(a). The filled circle represents the
light source displaced 1µ′s
inside the medium [203] and the open circle is the imaginary negative
source displaced an equal distance (zb) away from the extrapolated boundary where the fluence
rate becomes zero. This solution is used in Chapter 3, Section 3.7 for bulk optical properties
assessment of tissue phantom, in Chapter 4, Section 4.3 in calculating bulk optical properties of
healthy breasts, and in Chapter 5, Section 5.3 for preliminary assessment of clinical data.
The analytic frequency-domain solution for the slab geometry (Figure 2.4(b)) with extrapolated
27
boundary conditions is [55, 217]
Φ(ρ, z) =vS
4πD
(
m=∞∑
m=−∞
exp(ik√
ρ2 + (z − z+,m)2√
ρ2 + (z − z+,m)2−
m=∞∑
m=−∞
exp(ik√
ρ2 + (z − z−,m)2√
ρ2 + (z − z−,m)2
)
,
(2.8)
where z+,m = 2m(d+2zb)+ z0, z−,m = 2m(d+2zb)− 2zb− z0 for m = 0,±1,±2, ..., and d is
the slab thickness. The slab solution is constructed with a series of imaginary sources arising from
two air-tissue boundaries (Figure 2.4(b)). In practice, we use up to 5th term for m. The remaining
terms are usually quite small. This solution is used for analysis in Chapter 4, Section 4.3.
The presence of source detector coupling coefficients in the multiple source detector configu-
ration requires attention: differential and absolute approaches will be described in the following
sections. These approaches have distinct forms of χ2. Once χ2 is defined, various optimiza-
tion functions can be used to find the combination of unknowns which minimize χ2. Usually the
Nelder-Mead simplex method [189, 224] is utilized for our spectroscopy approaches.
2.3.1 Differential approach
When perturbative physiological measurements (Chapter 5, Section 5.2) are possible, the source
coupling coefficient problem becomes considerably simpler. Since the probe is fixed in the same
position under the same condition throughout the perturbation, the source detector coupling coef-
ficients are normalized out.
Typically we use χ2 of the form
χ2 =
Nλ∑
w=1
Ns∑
s=1
Nd∑
d=1
∣
∣
∣
∣
Φm(λw, rs, rd)
ΦRm(λw, rs, rd))− Φc(λw, rs, rd))
ΦRc (λw, rs, rd))
∣
∣
∣
∣
2
, (2.9)
where Nλ, Ns, and Nd are the number of wavelengths, sources and detectors respectively.
28
2.3.2 Absolute approach
Where there is no reference measurement, one needs to explicitly consider the effect of source
and detector coupling coefficients. In particular, the retrieval of optical properties of the matching
fluid of the 2nd generation parallel-plate DOT instrument (described in Chapter 3, Section 3.6.2)
relies on the absolute approach. Raw data of amplitude and phase shift is shown with respect to
the source and detector separations in Figure 2.5. The presence of source and detector coupling
coefficients leads to significant deviation of data points from the expected amplitude and phase
(normalized respctively to the measured values) for a homogeneous medium. This deviation does
not arise for the one source and one scanning detector configuration (Chapter 3, Section 3.6.1).
1 2 3 4 5 6−6
−5
−4
−3
−2
−1
0
ρ (cm)
log
(r2 *A
mp
litu
de)
(a)
1 2 3 4 5 6−0.5
0
0.5
1
1.5
2
ρ (cm)
Ph
ase
(rad
)
(b)
Figure 2.5: Effect of coupling coefficients on measured amplitude and phase of homogeneousmedium. (a) Amplitude and (b) Phase plotted versus source detector separations. Amplitude andphase are measured on a homogeneous matching medium using frequeny-domain, multiple sourcedetector instrument (Chapter 3, Section 3.6.2). Significant deviation from semi-infinite solution(solid line) shows the effect of the coupling coefficients on data.
An implicit way to estimate the source and detector coupling coefficients while fitting for the
chromophore concentration and scattering factors was developed. In this method, we defined χ2
29
to be normalized with the measured fluence rate,
χ2 =
Nλ∑
w=1
Ns∑
s=1
Nd∑
d=1
∣
∣
∣
∣
Φm − ΦcΦm
∣
∣
∣
∣
2
. (2.10)
We use Nelder-Mead simplex method to update the unknowns. At each iteration, when Cl, A
and b are updated, they are decomposed into µa(λ) and µ′s(λ) by Equation 2.4. For these µa(λ)
and µ′s(λ), coupling coefficients are estimated in the following two-step process. First, detector
coupling coefficients are estimated by summing over all the sources.
ξd(µa, µ′s) =
1
Ns
Ns∑
s=1
Ac(µa, µ′s)
Am(2.11)
θd(µa, µ′s) =
1
Ns
Ns∑
s=1
θc(µa, µ′s)− θm, (2.12)
where the calculated amplitude is Ac = |Φc| and the calculated phase is θc = arg(Φc). Then,
source coupling coefficients are estimated by summing the ratio of adjusted fluence rate over all
the detectors.
ξs(µa, µ′s) =
1
Nd
Nd∑
d=1
Ac(µa, µ′s)
ξd ·Am(2.13)
θs(µa, µ′s) =
1
Nd
Nd∑
d=1
θc(µa, µ′s)− θm − θd. (2.14)
These source detector coupling coefficients depend on µa(λ) and µ′s(λ) through Φc, which depends
on the updated unknowns at each iteration. The iteration continues until χ2 reaches the stopping
criterion.
30
2.4 Imaging
Two different imaging software packages were developed and employed in our laboratory for
breast cancer imaging. The first software, which will be referred to as “Nonlinear Rytov Itera-
tive Method (NRIM)”, was developed by Dr. Monica Holboke [128, 129]. NRIM has mainly been
used for analyzing the tissue phantom measurements to characterize DOT instrumentations [62].
For clinical data analysis, use of NRIM turned out to be challenging mainly because of its exces-
sive computational memory requirements; NRIM computes the Jacobian (weight matrix) explic-
itly. Another approach, TOAST (Time-resolved Optical Absorption and Scattering Tomography)
developed originally by Arridge et al [13] is based on the nonlinear conjugate gradient method;
it has significant memory overhead reduction compared to NRIM. Therefore a collaboration with
Dr. Simon Arridge’s group was initiated. The CW version of TOAST was adapted to our imag-
ing geometry and further developed at UPENN to include the multi-spectral approach [57, 58].
Furthermore, envelope-guided spatially variant regularization [56] and source-detector coupling
coefficient [162] fitting were incorporated.
In the following section, the details of NRIM, the single-spectral TOAST (original version),
and the multi-spectral TOAST (MTOAST) are described. Explicit forms of equations are presented
to bring out clear distinction among the algorithms.
2.4.1 Nonlinear Rytov Iterative Method
NRIM [128, 129] was first developed using a single-spectral approach. The unknown variables to
fit are the absorption coefficient µa and the reduced scattering coefficient µ′s. Furthermore, this
code was customized for breast imaging wherein we take reference measurement of matching fluid
(Intralipid/India ink). It is also suitable for differential measurement before and after contrast agent
31
injection. The following flow chart (Figure 2.6) shows more of the details for the inverse problem
(which can be compared with that of TOAST in the following section).
In diffuse optical community, quantitative imaging using continuous-wave (CW) method has
been controversial because of nonuniqueness of CW solution to the photon diffuse equation [9].
Since there is no unique solution, crosstalk between absorption and scattering is quite significant,
which in turn affects the accuracy of derived physiological parameters. Yet the CW approach offers
several advantages over frequency-domain or time-domain approaches: good signal to noise, sim-
plicity of technology, low cost per detection channel, and (most importantly for us) being readily
adaptable for lens-coupled CCD data acquisition to yield high spatial information for reliable 3D
reconstruction. Some research groups have devised a scaling approach [208, 287] to circumvent
this problem, which depends on their instrument and reconstruction geometry.
Corlu et al has shown that a multi-spectral approach helps to overcome the nonuniqueness
problem in CW imaging [58]. His implementation of this multi-spectral approach to nonlinear
Conjugate Gradient inverse algorithm (TOAST) is well-documented in the reference [57] with the
methodology to find optimal CW wavelengths.
The distinct differences between single-spectral and multi-spectral TOAST are (1) Definition
of unknown, (2) χ2 form, and (3) gradients of χ2 with respect to unknowns. The unknowns are
typically chromophore concentration Cl, A, and b (l = 1, . . . L).
The multi-spectral χ2 is different from single-spectral χ2 in that there is an additional summa-
tion over the wavelength, and the spectral constraint governs the optical properties µa(λ, r) and
µ′s(λ, r) which changes Φ(λw, rs, rd). The multi-spectral χ2 is further modified to include the
41
envelope-guided spatial variant regularization term [56],
χ2 =1
2
Nλ∑
w=1
Ns∑
s=1
Nd∑
d=1
(
ln
(
Φm(λw, rs, rd)
ΦRm(λw, rs, rd)
ΦRc (λw, rs, rd)
Φc(λw, rs, rd)
))2
+Q (2.31)
where Q =∑N
k=1 γ(rk)||µ(rk) − µ0(rk)||2, γ(rk) is given by Equation 2.22 and µ stands for
solution vector (either µa, D or both).
The gradient of χ2 is then calculated with respect to multi-spectral unknowns Cl, A and b.
∂χ2
∂Cl
∣
∣
∣
∣
rk
=∂χ2
∂µa
∂µa∂Cl
= εl(λ)∂χ2
∂µa
∣
∣
∣
∣
rk
(2.32)
∂χ2
∂A
∣
∣
∣
∣
rk
=∂χ2
∂µ′s
∂µ′s∂A
= λ−b(rk)∂χ2
∂µ′s
∣
∣
∣
∣
rk
(2.33)
∂χ2
∂b
∣
∣
∣
∣
rk
=∂χ2
∂µ′s
∂µ′s∂b
= −A(rk)λ−b(rk)ln(λ)∂χ2
∂µ′s
∣
∣
∣
∣
rk
(2.34)
where ∂χ2
∂µais given by Equation 2.23 and ∂χ2
∂µ′sby
∂χ2
∂µ′s
∣
∣
∣
∣
rk
= 3D2Ns∑
s=1
Nd∑
d=1
ln
(
Φm(rs, rd)
ΦRm(rs, rd)
ΦRc (rs, rd)
Φc(rs, rd)
)
1
Φc(rs, rd)∇G(rd, rk) · ∇Φc(rk, rs).
(2.35)
For breast imaging, a geometric constraint is utilized to give a good initial guess for the re-
construction. This additional technique only involves the changes in the input field to MTOAST,
and not the internal algorithmic changes. The gist of the geometric constraint is to fix the optical
properties of matching fluid based on the frequency-domain measurements and not to allow the
matching fluid region to update its optical properties. Only the breast region is allowed to update
its multi-spectral unknowns such as CHb, CHbO2, CH2O, and A. The specific methodology used
for our breast imaging application is described in Chapter 4, Section 4.4.2 in detail.
42
Chapter 3
Experimental Techniques
3.1 Introduction
In this chapter, experimental techniques are outlined. The choice of light sources, detectors and
electronic components depends on the physiological application. First the light sources and detec-
tors used in the DOS/DOT field are introduced. Then, the frequency-domain homodyne system
which is used in all experimental applications in this thesis is described. Specification of the
CCD-based instrument is described in the separate section. In the next section, the methods to
test electronical components and to characterize the whole system for dynamic range are outlined.
Then the specific instrumentation used for each application is described in detail. For breast cancer
imaging, there were two parallel plate instruments: Frequency-domain scanning system and CCD-
based hybrid system. For fetal oximetry, a multi-wavelength, multi-optode frequency-encoded
DOS instrument was developed. Validation and characterization of the system as a whole was
carried out using tissue phantoms.
43
3.2 Components
3.2.1 Light sources
Three types of light sources are used for DOS and DOT applications : white light, light emitting
diodes (LED), and laser diodes.
White light sources have gained more attention recently, because the importance of multiple
wavelength schemes has been recognized [19, 273]. The spectrum of most white light sources
extends over the visible range and into the near-infrared. Tungsten incandescent lamps radiate light
by heating tungsten filaments. Halogen lamps are essentially efficient high performance tungsten
lamps, with iodine or bromine in the fill gas to return evaporated tungsten to the filament. Arc
discharge lamps operate by ionizing xenon or mercury gas with a short high-voltage pulse and then
allowing the capacitor to discharge through the now-conductive gas. Typically white light sources
are utilized with a monochromator or a series of spectral filters.
In light emitting diodes, free electrons moving across a diode junction combine with holes
from the P layer. In this process, photons are generated. The photon energy is set by the energy
drop between the conduction band and the valence band.
In laser diodes (or diode lasers), a population inversion is induced in the P-N junction region.
Stimulated emission results from this population inversion and a net light amplification and lasing
is achieved by an optical cavity formed with the reflective coatings at opposite ends of the crystal.
Laser diodes are used widely for DOT applications. They are highly monochromatic and respond
rapidly to variations in the driving current.
44
3.2.2 Detectors
In the following sections, various detectors commonly used in our field are described with the focus
on the characteristics of signal and noise.
3.2.2.1 Photodiode
P
I
N
DepletionRegion
-+
+ -
+ -
(a)
P
N
DepletionRegion
+ -
+
+
-
-
(b)
Figure 3.1: Photodiode structure of (a) PIN diode and (b) Avalanche Photodiode. Reproduced fromreference [111]
The photodiode is a robust and inexpensive detector for relatively high light levels. Light with
energy greater than band-gap energy hits the photodiode, excites electrons into the conduction
band, and creates a hole in the valence band. An electric field in the depletion layer drives electrons
to the N layer and holes to the P layer. The external circuit collects this photocurrent. This process
is illustrated in Figure 3.1(a) for PIN photodiode. PIN photodiode includes an intrinsic region
between the P and N layers, which results in the expansion of depletion region and thus broader
spectral response. PIN photodiode does not have internal gain mechanism. The need of external
45
amplifier introduces substantial noise. The current is propotional to incident power in normal
operating range.
3.2.2.2 Avalanche Photodiode (APD)
The avalanche photodiode is a high-speed, high-sensitive photodiode with an internal gain mecha-
nism through a reverse-bias voltage. The electron-hole pairs are generated from exposure to light
with higher photon energy than band gap energy. A reverse voltage applied to the PN junction of
the APD causes the electrons to drift towards N layer and holes towards P layer. An avalanche
effect occurs when the electron-hole pairs acquire sufficient energy to create additional pairs by
colliding with the crystal lattice as shown in Figure 3.1(b). The gain of APD could reach up to
100 [111].
3.2.2.3 Photomultiplier Tube (PMT)
Photomultiplier tube (PMT) is a sensitive detector which amplifies the input light signal 105− 106
times with almost no additional noise. It is usually selected for high speed or low light level
detection. It is suitable for single photon counting applications when the rate photons strike the
photocathode is below 100 MHz.
Figure 3.2 illustrates how the PMT converts light into a detectable electric signal. The PMT
first converts the photon which strikes a photocathode into a photoelectron by the photoelectric ef-
fect. Amplification is performed through a dynode chain, which consists of 8 to 12 metal plates. A
potential of around 100 volts applied between the photocathode and the first dynode accelerates the
photoelectron into the dynode. Upon striking the dynode, 5 - 6 secondary electrons are produced.
These electrons are accelerated into the next dynode by a potential difference and get multiplied
46
faceplate
PhotocathodeDynodes
Anode last dynode
vacuum
stem
stem pin
focusing electrode
LIGHT
secondary electron
Figure 3.2: Schematic of typical photomultiplier with some electron trajectories (in red). Repro-duced from Hamamatsu manual [120].
by 5 - 6 times each. These cascading effect produces 105 − 106 electrons at the anode.
3.2.2.4 Charge Coupled Device (CCD)
The charged coupled device (CCD) is a solid state sensor with a wafer of silicon crystal. When
the silicon is exposed to light, the photoelectric effect generates electrons from the silicon bonds.
These free electrons are collected by CCD electrodes (as shown in Figure 3.3(a)) at the interface
created by positive surface potential to form an extremely thin but very dense inversion layer. The
amount of charge deposited in the inversion layer is often described by the hypothetical concept
of the potential well. In Figure 3.3(b), the charge transfer mechanism is illustrated using a three-
phase CCD structure as an example [20]. In a three-phase CCD, three sets of electrode strips make
one pixel where the charge accumulates on the electrode biased more positively than the other
two. The charge is transferred by making the adjacent electrode potential raised while the first
lowered. Figure 3.4 shows the CCD chip structure [169] consisted of parallel and serial electrodes
and amplifier on a silicon chip. The channel stops restrain electrons from moving along the length
of electrode, thus defining the extent of the pixel. The charges in potential wells columns are
47
shifted in parallel onto a serial register. Then the charges in the serial register gets shifted out
onto the output node. At this stage, on-chip hardware binning can be incorporated to increase the
signal-to-noise ratio. The charges are then amplified by an on-chip amplifier and digitized.
inversion layer
depletion layer
p-type semiconductor
oxide
electrode
(a)
0 V +V 0 V0 V
+V 0 V0 V
+V 0 V0 V
+V
0 V
potential wellcontaining charge
(b)
Figure 3.3: (a) Single CCD electrode, (b) Charge transfer mechanism of CCD electrodes (three-phase CCD system). Reproduced from reference [20].
3.2.2.5 Image intensifier
An image intensifier is a vacuum tube device consisting of a photocathode input, microchannel
plate, and the phosphor screen. The electrons due to photoelectric effect are accelerated and multi-
pled through microchannel plate (MCP) via mechanism analogous to photomultiplier, to the phos-
phore screen where the light is released upon striking the coating. The image intensifier can be
combined with CCD to detect in ultra-low-light conditions and resolve extremely short temporal
phenomena by biasing the photocathode with respect to the MCP.
48
1 2 3 1 2 3 1 2 3 1 2 3
one pixel
silicon chip
output amplifier
parallel transfer electrodes
serial output register electrodes
channel stop diffusion
123123123123
Figure 3.4: CCD chip layout for three-phase CCD system. First, charges collected at each electrodeare shifted along parallel electrodes onto serial register electrodes. Then the charges are shifted tothe output amplifier. Reproduced from reference [169].
3.2.2.6 SNR comparison of detectors
The dynamic range of the detector is characterized by a low limit set by the noise equivalent power
(NEP) from dark current, a middle linear region composed of various noise sources, and a high
limit set by saturation (Figure 3.5). It is important to identify and characterize the linear region
of system to optimize measurements for given application. The characterization of this region in
the overall electronic-optical system will be described in the following section 3.5. In this section,
intrinsic limitation arising from the choice of detector is closely examined.
When the incident light photons arrive at the detector, not every photon produces electrons.
This factor is commonly characterized by the quantum efficiency η (number of electrons produced
/ incident number of photons), which is given in percentage. Also, for APD and PMT, it is given
49
Noise Equivalent Power
Linear response region
Saturation
Input Power (W)
DetectorResponse (V)
Figure 3.5: Characteristic response of detector to the input light power
as the detector sensitivity in [A/Watt]
Sk =IkPin
, (3.1)
where Ik is the photocurrent [A] and Pin is the incident light power in [Watt]. The quantum
efficiency η and the detector sensitivity Sk are related by
η =hc
λqSk, (3.2)
where h is Plank’s constant (6.63×10−34 J×s), c is the speed of light (2.99×108 m/s), q is electron
charge (1.60× 10−19 coulomb), and λ is the light wavelength (in [m]).
There are three major noise factors in photodiode and photomultiplier: shot noise due to in-
cident photon statistics, dark current induced noise, and thermal noise. The shot noise associated
with current flow across a potential barrier is given by ishot =√2qBI [A] where I is the average
50
dc current [A] and B is the bandwidth [Hz] [201]. The shot noise due to the photocurrent Ik is
iphoton =√2qBIk, where the shot noise is in fact proportaional to
√nphoton, where nphoton is the
equivalent number of photons after degradation caused by imperfect conversion caused by quantum
efficiency η. Since Ik is proportional to the incident light power, the photon shot noise increases
with respect to the increase of Pin (in square root fashion). The lower end of this relation is limited
by the noise from the dark current Idark, which is present even there is no input light. The param-
eter is usually given in terms of Noise equivalent power (NEP), where PNEPSk =√2qBIdark.
These noise factors are given at the point where no amplification has yet taken place, such as at the
cathode of PMT or before the avalanche effect in APD. Thermal noise is due to thermal agitation of
electrons within a resistance and given by ithermal =√
4kTB/R, where k is Boltzman’s constant
(1.38× 10−23 J/K), T is absolute temperature [K], and R is the load resistance [Ω]. Thermal noise
usually sets the lower limit on the noise present in an electronic circuit.
To calculate the signal-to-noise ratio (SNR) of a PMT, we consider the signal and noise after
the amplification process of the detector. The amplified current is then Isignal = GIk = GSkPin
where G is gain and the noise is given as
inoise =√
2qBG2Ik + 2qBG2Idark. (3.3)
Here, the thermal noise component is ignored since PMT is a current source. SNR is then calcu-
lated via Isignal/inoise. This simplified relationship applies to PMT with the assumption that gain
is quite large and there is minimal additional noise coming from the amplification process (i.e.
dynode chains). In photon counting mode, one could approximate dark noise contribution to be
zero since the discriminator threshold could be set up to reject dark count (i.e. shot-noise limited).
51
However, in the case of the APD, it is more complicated due to its different amplification
mechanism. Dark current in the APD is categorized into surface leakage current Ids which does
not go through the multiplication process and an internal current Idg generated inside the silicon
substrate. The non-uniformity of ionization induces “excess noise” during the multiplification
process. The Equation 3.3 becomes
iAPDnoise =
√
2qBG2F (Ik + Idg) + 2qBIds +4kTB
R(3.4)
where F is excess noise factor related to gain G. In PIN photodiode, usually the thermal noise
plays a major role in the noise contribution since there is no gain mechanism.
The calculation of signal and noise of the CCD takes a different form as follows. There are
four major noise sources associated with the CCD; dark current noise, read noise, shot noise and
fixed-pattern noise. Due to the thermal energy within the silicon lattice of the CCD, dark current
is generated. The statistical fluctuation of dark current (dark current noise) can be reduced by
cooling the CCD with thermoelectric coolers (TEC) or liquid nitrogen. Read noise is the random
noise associated with the process of quantifying the electronic signal on the CCD, especially from
on-chip pre-amplifier. Shot noise arises from the inherent variation of the incident photon flux.
Fixed-pattern noise is a dominant noise at relatively high light level, resulting from differences in
sensitivity among pixels or photo-response nonuniformity.
Within the light level below where the saturation begins, the signal-to-noise ratio can be calcu-
lated by
SNRCCD =ηNpt
√
ηNpt+ (Ndt)2 +N2r
(3.5)
where η is the quantum efficiency, Np is the photon flux (photons/pixel/second), t is the integration
52
time (seconds), Nd is the dark current (electrons/pixel/second, in short form e−/p/s), and Nr is
the read noise (electrons RMS/pixel, e−/p) due to the on-chip amplifier. Given Pin at one pixel,
Np = λPin/h/c.
For detectors commonly used in our laboratory, SNR for incident power level can be calculated
using the specification information summarized in Table 3.1. SNR of APD, PMT and CCD is
plotted with respect to the incident light power in the horizontal axis in Figure 3.6. The bandwidth
B was assumed to be 1 Hz. The load registance of APD is order of 1 GΩ. The APD inside
the module C5331-04 is S2384. The excess noise index is 0.3, therefore the excess noise factor
F = G0.3. Also, as a quick comparison, PMT with improved quantum efficiency in NIR region
was considered with photon-counting mode (PMT: III-V. η = 15 % in NIR, G = 106, maximum
cathodecurrent = 1 pA.).
type model NEP (W/√Hz) sensitivity η gain Idark
APD S2384 4 × 10−13 0.5 A/W 32 % 30 1 nA (after gain)PMT R928 1.3 × 10−16 0.02 A/W 3 % 107 30 nA (anode)type model Nr(e
−1/p) full well η exposure time Nd(e−1/p/s)
CCD Versarray 8 200,000 e−1 37 % 500 ms 0.5
Table 3.1: Parameters needed for estimating signal-to-noise ratio for detectors commonly used inour instrument. Noise equivalent power (NEP), sensitivity and quantum efficiency are given forλ = 800 nm.
It should be noted that incident light power in Figure 3.6 is per unit detector, i.e. area of
detection was not considered into the calculation yet. For each detector, the incident power range
up to the onset of saturation is shown. The lower limit of APD (S2384) is governed mainly by
the dark current noise component. The cross-over to photon shot noise limited region happens
around 10−10 W. In reality, APD does come with built-in amplifier as a module (C5331-04), whose
significant noise factor was not considered in this calculation. Therefore its SNR in the high light
53
10−18
10−16
10−14
10−12
10−10
10−8
10−1
100
101
102
103
104
Incident light power (W)
Sig
nal
/No
ise
PMT: R928APD: C5331−04CCD: VersarrayPMT: III−V
Figure 3.6: Signal-to-noise ratio (SNR) vs incident light power. SNR of PMT, APD and CCDcommonly used in our instrument are estimated based on the specification given in Table 3.1.PMT with improved photocathode material is shown as PMT: III-V. When SNR = 1, the detectorperformance is limited by the noise. Since the high end of linear range is limited by the saturationpoint, only the incident light power range up to the saturation is shown for each detector (i.e.maximum CCD SNR is 500.).
power range appears to be quite high. PMT R928 in our laboratory is used in analog mode (i.e.
not photon counting mode) and not cooled. Thus, the lower light is limited due to the presence
of dark current noise, but it quickly goes into photon shot noise limited linear region from 10−15
W to its saturation ∼ 2 × 10−10 W (saturation due to anode linearity limit). This large dynamic
range enables us to employ various combination of multiple source detector separations needed for
imaging geometry. The specific CCD that we are using for breast cancer imaging has high quantum
efficiency. However, the maximum SNR is limited by its linear full well, which is the number of
electrons that the individual CCD pixels can hold before spilling over into adjacent wells.
Figure 3.6 gives an impression that CCD will be definitely the detector of choice. However, one
54
should not use this figure assuming the same amount of incident light falls onto each detector unit
in the real scenario. In the experimental setting, one usually would like to use the same source light
power onto the turbid medium of interest and see how the different detectors respond. The incident
light power Pin for unit detector depends on detector-specific parameters such as the collection
area. That is,
Pin = |Φ(rs, rd, µa, µ′s)| ·Adet · floss (3.6)
where Adet is the detector area, floss is the loss factor associated with inefficiency from fiber
coupling and factor relating to N.A. of detector [93], and |Φ| is the fluence rate intensity in medium
(with optical properties µa, µ′s) with source positioned at rs, detector positioned at rd.
If we assume a homogeneous medium with typical optical properties measured at source de-
tector separations of 6 cm, |Φ| ∼ S × 10−6 [Watt/cm−2] (Equation 2.7). Also, if we assume S= 1
mW and general loss factor floss to be 0.05 for simplicity, then one can see the area of the detector
plays a big role in determining Pin. We usually use 6 mm diameter fiber for PMT R928 and 3
mm diameter fiber for APD C5331-04. For CCD, 0.3 cm pixel size is assumed which is similar to
what we use in our measurement. The bandwidth B was taken as 2 Hz (to match the usual CCD
exposure time of 500 ms). From this detection area differences, incident power and SNR for the
signal level are summarized in Table 3.2. SNR of PMT is larger than that of CCD and much larger
than that of APD.
Even though this calculation is more of back-of-the-envelope type, it qualitatively matches
with our experimental observations. In the real experimental set-up, the use of APD for 6 cm
source detector separation is not so practical. This can be seen from the incident power being
very close to its NEP. With more inefficient coupling, Pin can easily become similar to NEP.
On the other hand, PMT has been used extensively for large source detector separations without
55
type model coupling detection area (cm−2) Pin SNRAPD S2384 3 mm fiber 0.07 3.5× 10−12 W 8PMT R928 6 mm fiber 0.28 1.4× 10−11 W 660CCD Versarray 0.3 mm pixel 4.5× 10−4 9.0× 10−14 W 180
Table 3.2: Example of SNR calculation of our commonly used detectors in the standard breastmeasurement for λ = 800 nm. Depending on the detection area, the incident input power (Pin tothe detector varys and thus affect SNR.
difficulties. Note that Pin on a CCD pixel is much lower than that of APD or PMT due to its area
even though the capability of CCD to detect low incident power seems to be better than PMT due
to its excellent quantum efficiency. (Area was not from the physical CCD-chip dimension, but the
magnified imaging area by the lens typically found in the experiment.) However, SNR retrieved is
still reasonable for the geometry that we are currently using. One could argue that by averaging
over many number of pixels to cover larger collection area may vastly improve the SNR. However,
each pixel is limited by its read noise (which is higher than several photons) and averaging over
pixels with no response may not necessarily improve SNR by many-fold as expected. Whereas in
PMT, larger collection area can be utilized to enhance the light detectability since it is not limited
by read noise.
3.3 Frequency-domain Homodyne System
3.3.1 Frequency-domain System Schematic
The schematic of the homodyne frequency-domain system is illustrated in Figure 3.7. The local
RF oscillator provides a modulated waveform for the laser diode and the reference signal for the
in-phase and quadrature (I&Q) demodulator. The RF modulated light from the laser diode is
coupled to the turbid medium through optical fibers. Depending on the multiplexing method,
56
RF local oscillator
LPFLPF
Turbidmedium
Detector
I&QDemodulator
Computer
Splitter
AmplifiersAttenuators
Filters
AD board
Laser Diode
Optical switch
Figure 3.7: General Homodyne frequency-domain system (LPF: low pass filter). RF local oscilla-tor provides the reference signal to I&Q demodulator while modulating the laser diode. The source,detector positions and wavelengths are multiplexed using a combination of optical switches. Thenthe attenuated signal through the turbid medium is detected by the detector, goes through properamplification and filtering. I&Q demodulator with low pass filter extracts amplitude and phaseinformation which are recorded in computer through analog-to-digital (AD) board.
optical switches or optical combiners are used for wavelength and optode position changes. The
attenuated signal after propagating through the turbid medium is detected by either a PMT or an
APD. The signal goes through bandpass filter and amplifirs before reaching the I&Q demodulator.
The I&Q demodulator computes the amplitude and phase of the detected signal with respect to the
reference signal. The details of I&Q demodulator are described in the Section 3.3.3.
3.3.2 Laser diode modulation
In RF modulation of laser diode, one needs to first choose the driving circuit which can drive
the laser diode in continuous wave mode with good stability. The driving circuit depends on the
57
polarity of the laser diode and the monitor photodiode. There are four common types of laser diode
configurations depending on the polarities of the laser diode and photodiode connection as shown
in Figure 3.8. The photodiode embedded within the laser diode circuitry is used for regularizing
the laser diode output by feedback circuit. In naming the style, we are following the conventions
used by Thorlab, Inc. in their catalog [256].
cathode
anode
(a)
LD
PD
Style A
(b)
LD
PD
Style B
(c)
LD
PD
Style C
(d)
LD
PD
Style D
(e)
Figure 3.8: (a) Diode polarity, (b) Laser diode style A, (c) Laser diode style B, (d) Laser diodestyle C, (e) Laser diode style D. LD stands for laser diode and PD stands for photodiode.
Most of our laser diodes were modulated using laser diode driver chips from Sharp: IR3C01
for the laser diode style B and IR3C07 for the laser diode style C.
laser style Rr Cr Lr Rc Cc Rv
IR3C01 B 47Ω 45pF (f=70MHz) 330µH 20Ω 22µF 20kΩ22pF (f=140MHz)
IR3C07 C 47Ω 45pF (f=70MHz) 100µH 10Ω 22µF 100kΩ22pF (f=140MHz)
Table 3.3: Parameters for laser diode RF modulation
Rr, Cr and Lr are the parts essential to RF modulation. If one omits Rr, Cr and Lr from the
circuit in Figure 3.9, the laser diode will operate in CW mode. Rr = 47Ω was chosen to give 50
Ω impedance when combined with internal resistance of 3 ∼ 5 Ω of powered laser diode. Cr is
58
1
2
3
4
7
6
5
8
IR3C01
RF in
Rr Cr Lr
Laser diode(style B)
Rc
Cc
-12 V
Rv
+5 V
TTL High = ONTTL Low = OFF
LDPD
(a)
1
2
3
4
7
6
5
8
IR3C07
RF in
Rr Cr Lr
Laser diode(style C)
Rc
Cc
Rv
+5 V
TTL High = ONTTL Low = OFF
LDPD
(b)
Figure 3.9: RF modulation diagram for laser diodes (a) IR3C01 for laser diode style B, (b) IR3C07for laser diode style C. Without Rr, Cr and Lr, the laser diodes operate in CW-mode.
chosen to match the RF input frequency, i.e. f ∼ 12πRC where R = 50Ω (Cr = 45pF for f = 70
MHz, Cr = 22pF for f = 140 MHz). Lr is added to prevent the RF signal from passing into
the laser diode chip. Rc, Cc and Rv values vary depending on the laser diodes. Rv is a variable
resistor with maximum resistance value given by the Table 3.3. Therefore, the tabulated values are
not absolute and are included only as a reference.
3.3.3 I&Q demodulator
The heart of the homodyne frequency-domain detection is the I&Q demodulator (Figure 3.10).
Suppose the reference signal isREF = Ar sin(ωt) and the detected signal isDET = Ad sin(ωt+
θ), where θ is the phase shift with respect to the reference. In the I&Q demodulator, the reference
signal is split into two, one of which goes through a 90o phase shifter. The detected signal is
also split into two, but they do not go through the phase-shifter. The multiplication of unshifted
59
90 shifto
0 shifto
Reference signal
Detected signal
I(t) Q(t)
Figure 3.10: Schematic of In-phase and Quadrature (I&Q) demodulator.
reference and detected signals occur at in-phase section.
I(t) =Ar2
sin(ωt) · Ad2
sin(ωt+ θ) + I0
= A cos(θ)−A cos(2ωt+ θ) + I0 (3.7)
where A = ArAd8 and I0 is the DC offset when the RF is absent. In the quadrature section, 90 o
shifted reference and unshifted detected signals get multiplied.
Q(t) =Ar2
cos(ωt) · Ad2
sin(ωt+ θ) +Q0
= A sin(θ) +A sin(2ωt+ θ) +Q0 (3.8)
where Q0 is the DC offset when the RF is absent. The low pass filter following at I and Q
output selects out the high frequency signal components. Therefore, IDC = A cos θ + I0 and
60
QDC = A sin θ + Q0 are the final output signal arriving at data acquisition point. The amplitude
and phase of the detected RF signal is then
A =√
(IDC − I0)2 + (QDC −Q0)2 (3.9)
θ = tan−1(
QDC −Q0IDC − I0
)
(3.10)
It is a standard practice to measure the DC offset I0 and Q0 by blocking the light input to the
detection system.
3.4 Continuous-wave CCD System
In this section, specification of CCD camera used in the second generation CCD-based parallel
plate instrument (Section 3.6.2) is discussed. A 16-bit thermoelectically cooled CCD array (Ver-
sArray:1300B, Roper Scientific) is operated at 500 ms exposure time, 2×2 on-chip binning at
high-sensitivity amplifier setting for the region of interest covering an imaging area of 16 cm ×
11 cm onto 570 × 400 binned pixels. The specification is summarized in Table 3.4. Binning of
the image increases SNR at the cost of resolution. On-chip binning which happens at the hardware
level yields higher SNR given exposure time. This feature can also be used to decrease exposure
time for the same level of SNR. The use of ROI in conjunction of binning makes the CCD trans-
fer time shorter than that of the full frame (450 ms for our ROI, 1.8 second for full frame). The
spectral response of this CCD is good for near-infrared range. The quantum efficiency is above
60 % from 380 - 850 nm with its peak around 550 nm (92 %), decreasing towards longer wave-
length (30 % for 950 nm). (Back-illuminated device has better quantum efficiency compared to the
front-illuminated device.)
61
CCD sensor E2V CCD36-40; back-illuminated device (with AR coating)CCD format 1340 × 1300 imaging pixels; 20 × 20-µm pixels
100% fill factor; 26.8 × 26.0-mm imaging areasingle pixel full well 200,000 e− - 300,000 e−
Dark current 0.1 e−/p/s at -40oC operationNonlinearity < 2%
Nonuniformity ≤ 4% over entire CCD areaDynamic range 16 bitsFrame readout < 1.8 seconds for full frame at 1 MHzDark current < 0.5 e−/p/s at -50 oC
Operating temperature -40 oC with TEC (backfilled)Thermostating precision 0.05oC over entire temperature range
Table 3.4: Specification of CCD array, VersArray:1300F
For our CCD, dark current noise is 0.1 e−/p/s (electron/pixel/second) and read noise is 8 e−/p
for the high-sensitivity amplifier. For 500 ms exposure time, the dark current noise is 0.05 e−/p
which is considerably smaller than the read noise. The saturation point is when the potential well
is full (∼200,000 electrons), which corresponds to 1.5× 10−13 W. Since the read noise defines the
limitation of the signal, the dynamic range of CCD is obtained by usually dividing linear full well
by read noise. In order to take full advantage of the dynamic range, an appropriate A/D converter
is selected for each camera. The gain is defined in terms of electrons/ADU (analog-to-digital unit).
Suppose the gain is set to 1×, then 1 ADU corresponds to ∼3 electrons (linear full well / 216).
However, for this particular CCD, the conversion factor was set as 4 e−/ADU, 2 e−/ADU, and
1e−/ADU (which is referred as gain setting). Depending on the gain setting, the saturation due to
the full potential well may not happen before it is limited by digitization limit (216 ADU). For the
example presented in Table 3.2 (i.e. Pin = 4.5× 10−14 Watt), assuming 70% quantum efficiency,
the number of electrons deposited in a pixel is 63,000 e− which corresponds to 31,500 ADU for
medium gain. The order of magnitude of this matches with the experimentally measured value.
62
3.5 System Characterization
Once the instrument is designed and assembled, the characteristic dynamic range needs to be quan-
tified. It is important to first identify the dynamic range of major electronic components, since the
instrument design goal is to build the system limited by the optical components (i.e. detectors),
not by the electronic components. Then the dynamic range of the whole system including the
electronic and optical components is measured to characterize the system. This methodology is
demonstrated in the following sections using a one detector channel frequency-domain homodyne
system.
3.5.1 Dynamic range of electronic components
Attenuator
Power Splitter
70 MHz Oscillator
I&Q Demodulator
LPF LPF
Referencesignal
Amplifiers,etc.
AD board
Computer
Inputsignal
Figure 3.11: Schematic of electronic dynamic range measurement set-up using electronic attenua-tor.
63
In determining the dynamic range of electronic components, we started by testing the response
of I&Q demodulator with input signal variation. Then we added amplifiers one by one to see if
there is any degradation due to addition of amplifiers.
The schematic of the test set-up is described in Figure 3.11. A 70 MHz local oscillator signal is
split into two. One provides the reference signal for the I&Q demodulator. The other is attenuated
and serves as the detected signal to the I&Q demodulator. Amplitude response measured for each
RF input signal is plotted for three cases in Figure 3.12: (1) only I&Q demodulator was present, (2)
I&Q demodulator and two amplifiers were present and (3) I&Q demodulator and three amplifiers
were present. The dynamic range defined by I&Q demodulator is RF input of -60 dBm to -5 dBm.
The addition of amplifiers do not limit this range, but merely shifts the range to lower RF input by
its gain factor (i.e. With two amplifier, the dynamic range is -120 dBm to -65 dBm).
-140 -120 -100 -80 -60 -40 -20 0 20
-60
0
-20
-40
-80
-100
20
RF input signal (dBm)
Measured amplitude (dBm)
2 amplifiers3 amplifiers
IQ only
Figure 3.12: Effect of amplifier addition on the electronic dynamic range. 2 amplifiers: test of twoamplifiers and I&Q demodulator, 3 amplifiers: test of three amplifiers and I&Q demodulator
64
The I and Q output contains the DC offset (I0, Q0) as in Equation 3.10. As shown in Fig-
ure 3.13, offset subtraction improves the amplitude dynamic range especially at the low end and
significantly improves the phase stability.
−40 −30 −20 −10 0 10−40
−30
−20
−10
0
10
RF input (dBm)
Am
plit
ud
e (d
Bm
)
before offset subtractionafter offset subtraction
(a)
−40 −30 −20 −10 0 100
50
100
150
200
250
RF input (dBm)P
has
e (d
egre
e)
before offset subtractionafter offset subtraction
(b)
Figure 3.13: Effect of offset subtraction on dynamic range. (a) Amplitude vs RF input, (b) Phasevs RF input. Offset subtraction improves the dynamic range at low RF input and the phase stability.
3.5.2 Dynamic range of the whole system
The dynamic range of the whole system combining the electronic and optical components is as-
sessed using the set-up illustrated in Figure 3.14. The light emitted from light source is reduced by
a mechanical attenuator and then split into two beams by 1×2 fiber splitter. One beam goes to an
optical power meter and the other one goes to the detection part of the optical system. The signal
response of the detection system to the attenuated light signal is plotted against the light signal
measured by the optical power meter in Figure 3.15(a). Figure 3.15(b) shows the deviation of the
measured value from the fitted linear line over the dynamic range. Typically, we choose the range
by the deviation of ± 1%. In this example, the dynamic range is -50 ∼ -10 dBm. Note this range
65
is smaller than the electronic dynamic range of previous section. However, this dynamic range can
be extended further with use of attenuators.
OpticalPowermeter
1x2 fiber splitter
Detectionelectronics
Light Source
Mechanical attenuator
Figure 3.14: Amplitude linearity test set-up
The phase accuracy is tested using the set-up shown in Figure 3.16(a). The optical fiber end
mount and lens are separated by the focal length of the lens so that the light is parallel after exiting
the lens. Then another lens focuses the beam into fiber coupled to the detector. The mount and first
lens are moved as a unit while the distance between the two lenses is varied. The change in the
distance between the two lenses produces a change in phase. Since the modulation frequency is 70
MHz, the expected phase change is 0.84o/cm. Typical linear response in phase at fixed amplitude
within the instrument dynamic range is shown in Figure 3.16(b).
66
−70 −60 −50 −40 −30 −20 −10 0
−60
−40
−20
0
20
40
60
Power (dBm)
Am
plit
ud
e (d
Bm
)
measuredfit
Dynamic Range
−50 −40 −30 −20 −10−3
−2
−1
0
1
2
3
Power (dBm)
Err
or
(%)
Dynamic Range
Figure 3.15: Amplitude linearity Test data. (a) Amplitude vs power in linear scale, (b) Amplitudeerror (%)
The 1st generation parallel plate transmission DOT instrument is utilized for quantification of
bulk properties of healthy female breasts [81] (Chapter 4, Section 4.3). The instrument uses three
wavelengths - 750nm, 786nm and 830nm, and employs a scanning, fiber-coupled PMT detector
(R928, Hamamatsu) for detection. The system is characterized by a noise equivalent power of
≈ 0.1pW/√Hz, a linearity in amplitude of less than 1%, and a phase drift of 0.25o over 80 dB
(with use of attenuators). We calibrated the instrument over a broad range of input powers in order
to extend this range. For transmission measurements the signal variation is typically only ≈30
dB. In clinical measurements we utilize a single source position due at the center of the scanning
region. The lasers are amplitude modulated at 140 MHz to produce diffuse photon density waves
(DPDWs) in the medium. The amplitude and phase of the DPDW is recovered using a homodyne
IQ-demodulation scheme [288] (Section 3.3.3).
67
Detectionelectronics
Light Source
Mechanical attenuator movable
lens unit lens
0 2 4 6 8 10 120
2
4
6
8
10
12
expected phase shift (degree)
exp
erim
enta
l sh
ift
(deg
ree)
Figure 3.16: (a) Phase linearity test set-up, (b) Phase linearity test data. Measured phase shift (indegree) vs expected phase shift (in degree) at fixed amplitude within the linear dynamic range.
For the in vivo measurements the volunteer lies in the prone position and her breast is inserted
into a small tank filled with a matching-fluid solution of Intralipid and india-ink mixture (see Figure
3.17). The detector is scanned along the output plate glass surface. The source is attached to the
movable compression plate which applies a gentle compression to the breast. Usually the range of
compression is 4.5 cm to 7 cm. It takes ≈ 15 minutes to acquire data from a 9.6 cm (x) by 4.8 cm
(y) scan region with 153 (17×9) points. We note that the feedback from volunteers was generally
positive; in comparison to the X-ray mammography, they felt that the soft-compression of the DOT
instrument did not cause discomfort. Figure 3.18 shows the detailed electronics of the set-up. We
took two sets of data for each volunteer: 1) From the tank filled with the matching fluid, without
the breast; this data enables us to normalize for instrument response for imaging purposes, 2) From
the tank filled with the matching fluid and the volunteer’s breast. The Intralipid/ink solution helps
to reduce the detrimental effects of breast boundaries by acting as a matching medium.
68
Electronics
X−Y scanningstage
Detector Fiber
Source Fiber
Intralipidbox
Compressionplate
Source on opposite plane
Image View
Detector grid
Figure 3.17: A sketch of the prototype clinical table. The volunteer lies in the prone position withher breast inserted into the tank through an openning on the bed. Soft compression is applied onthe source plane and detector scans a 2D grid on the opposite plane. Image view shows the sourcedetector positions as seen in the data.
Based on the experience gained by the first generation parallel plate instrument, the second gener-
ation instrument was developed with the focus on imaging [62]. A schematic of the DOT instru-
ment is shown in Figure 3.19. The hybrid system combines frequency-domain (FD) remission and
continuous-wave (CW) transmission detection. This parallel-plane DOT system has been exten-
sively characterized for breast cancer imaging using various tissue phantoms and a healthy female
volunteer [62] (Chapter 3, Section 3.7.2) and used extensively for breast cancer imaging (Chapter
4, Section 4.4 and 4.5).
The table was designed so that when the female subject lies on it in the prone position, both
her breasts are inside a breast box (60 × 22 × 23 cm) extending underneath the table. A breast
is typically centered and softly compressed between a movable compression plate and a parallel
viewing window with cranio-caudal compression. The compression distance varied from 5.5 - 7.5
cm, depending on breast size. The breast box was coated black and designed to hold the matching
fluid which has optical properties similar to human tissue. The matching fluid was made with
Liposyn III (30 %, Abbott Laboratories, Chicago, IL) as a scattering agent and India ink (Black
India 4415, Sanford, Bellwood, IL) as an absorption agent. The optical properties of the matching
fluid were fixed so that µa(786nm) = 0.05 cm−1 and µ′s(786nm) = 8 cm−1. The absorbance of
the ink solution was checked after each measurements with a spectrophotometer (Ocean Optics,
USB2000).
Four laser diodes at 690, 750, 786 and 830 nm were sinusoidally intensity modulated at 70
MHz. The modulation depth for each wavelength was 86 %, 99 %, 72 % and 67 % respectively.
Later on, 650 and 905 nm CW laser diodes were added to the instrument (along with optimization
of acquisition time from 12 minutes to 8.4 minites). Combination of optical switches (DiCon Fiber
70
(a)
Source Plane Detector Plane
12.8 cm
6.4 cm
15.6 cm
9.0 cm3mmSource
FD detectors
detectors984
(b)
Figure 3.19: Schematic of parallel plate diffuse optical tomography instrument. (a) Frequency-domain (FD) remission and continuous-wave transmission measurements are performed simulta-neously on a female subject lying in the prone position. (b) The source plate contains 45 sourcepositions and 9 FD detectors. 984 detection points with 3 mm spacing are selected from CCD datafor image reconstruction.
Optics) making 6× 2 connection was used for wavelength switching. The light output was relayed
to a 2× 48 switch (DiCon Fiber Optics, GP700) in order to access the various light source positions
on the compression plate. The compression plate had 45 fibers in 9 × 5 grid with a spacing of
1.6 cm (12.8 cm × 6.4 cm). The optical fibers embedded in the compression plate was of 200
µm diameter (FIS). Time-multiplexing was used for both wavelength and position switching. In
particular, wavelengths are multiplexed for given source position.
71
Nine 3 mm diameter fiber bundles were connected to the frequency-domain detection module,
interlaced in a 3 × 3 grid on the compression plate as shown in Figure 3.19(b). A homodyne
technique [288] was utilized to extract the amplitude and phase of the detected remission signal.
The electrical signal from the avalanche photodiode (APD, Hamamatsu C5331-04) is amplified
(Mini-Circuits ZFL-500LN, 24 dB), filtered by a band pass filter (Mini-Circuits BLP-70), and
then amplified again (Mini-Circuits ZFL-500HLN, 19 dB). An I&Q demodulator (Mini-Circuits
ZFMIQ-70D) with a series of low pass filters (Mini-Circuits SLP-1.9 and 100 Hz RC circuit)
extracts the amplitude and phase by comparing the signal with the reference signal driving laser
diodes [288]. The frequency-domain measurements are used for accurate quantification of bulk
properties of human tissue and matching fluid, thus improving our initial guess for the image
reconstruction.
A lens-coupled 16-bit CCD camera system was used for collecting CW transmission data with
an anti-reflection coated glass viewing window. A lens (Nikkor AF 50 mm F/1.4D) relayed the
detection plane (inner glass window in contact with breast) onto the CCD chip (2.68× 2.6 cm). To
reduce ambient light, a long-pass color-glass filter (630 nm, CVI Laser Inc.) was placed in front of
the lens and a light shielding box was placed surrounding the space between the viewing window
and the camera. A thermoelectrically cooled CCD pixel array (Roper Scientific, VersArray:1300F,
1340 × 1300 pixel) was used for light detection, with 1140 × 800 pixel region of interest (ROI)
and 2 × 2 on-chip binning, resulting frame size of 570 × 400 pixels covering a detection area of
16 × 11 cm (as described in Section 3.4).
72
Figure 3.20: Schematic of (a) the trans-abdominal probe and (b) the NIR frequency domain instru-ment to which the probe is coupled.
3.6.3 Frequency-encoded DOS instrument for fetal oximetry
We designed a multi-wavelength, multi-separation NIR frequency-domain instrument and probe
for trans-abdominal NIR spectroscopy [50] (Chapter 5, Section 5.2). Figure 3.20 shows a schematic
of the trans-abdominal probe and the NIR frequency-domain instrument to which the probe is cou-
pled. The probe was designed as a linear array consisting of 2 detector fibers and 6 source fibers.
This probe is capable of performing NIR photon diffusion measurements at a total of 12 source-
detector separations ranging from 1.8 to 9.5 cm. However, in this study, NIR photon diffusion
measurements at 8 source-detector separations ranging from 1.8 to 4 cm (Figure 3.20(a)) were
73
sufficient for retrieval of fetal cerebral blood saturation. The instrument consists of a light source
module and two detection modules. It is described in detail by Yu et al. [293]. Laser diodes at 675
nm, 786 nm and 830 nm are intensity modulated with three different local radio frequency (RF)
oscillators operating at around 70 MHz. The light output from each laser diode are combined using
a fiber combiner into a single optical fiber. In this manner, the single optical fiber can simultane-
ously deliver light output from the three frequency encoded laser diodes [288,293]. The total light
output from the single optical fiber was 15 mW (3 mW at 675 nm, 9 mW at 786 nm and 3 mW at
830 nm respectively). An optical prism switch was used to direct the light output from the single
optical fiber to the six source fiber positions on the trans-abdominal probe.
The detection module consists of an avalanche photodiode (APD), and two amplifiers, with a
band pass filter between them. The output from the detection module is connected to a demodu-
lation unit, which consists of three In-phase and Quadrature-phase (I&Q) demodulators. The in-
phase and quadrature-phase signals are low-pass filtered, digitized, and then converted into diffuse
wave signal amplitude and phase. Total acquisition time for the NIR photon diffusion measure-
ments at the three wavelengths and 12 source-detector separations was 1 second.
This instrument has been extensively tested in tissue phantom and rat brain studies [293]. The
noise equivalent power is less than 10 pW/√Hz and the dynamic range is greater than 70 dB
(amplitude errors < ± 1% and phase error < 1o.). The instrument has very low inter-channel
crosstalk (< -80 dB) and good long-term stability (amplitude errors < ± 1% and phase error < 1o
in 30 minutes).
74
3.7 Validation with Tissue Phantoms
Liquid phantoms serve as the gold standard to test the instrument and the algorithm using a com-
bination of absorbers and scatterers with known spectra. The titration of absorbers is useful for
testing the µa response of the overall system. Similarly, the titration of scatterers tests the µ′s re-
sponse. Indocyanine Green (ICG) is helpful for checking the spectral response since its spectra
has a well-defined peak in the NIR spectrum. Ultimately, titration of hemoglobin simulates the
global response from human tissue. The recipe for constructing liquid phantoms is described in
Section 3.8.1. However, solid phantoms are useful in simulating more tissue-like geometry and do
not deteriorate as fast as the liquid phantom. For imaging purposes, a compressible solid phantom
was developed as described in Section 3.8.2.
3.7.1 1st generation parallel plate DOT system phantom test
Typical raw data of breast is shown in Figure 3.21, overlayed with the forward data generated
from a fit based on the semi-infinite homogeneous medium analytic solution and a fit based on the
finite difference method (FDM). FDM fit was based on two regions of different optical properties
(breast and matching fluid regions) as shown in Figure 3.22(b). Particularly, NRIM described
in Chapter 2, Section 2.4.1 was modified to calculate bulk optical properties by summing up the
weight matrices only within the breast region and not allowing the matching fluid region to update
optical properties. The dispersion of amplitude and phase from the semi-infinite fit is mainly due
to the boundary effect, as can be seen from the FDM fit.
The scanning-based 1st generation instrument has been characterized by extensive phantom
studies [64, 82–85]. In particular, the phantom experiment was devised to best model the actual
breast measurement geometry as described. Balloons filled with different concentrations of ink
75
5 5.5 6 6.5 7 7.50
0.005
0.01
0.015
0.02
r (cm)
Am
p (n
orm
)FD Meas Semi−inf
(a)
5 5.5 6 6.5 7 7.5−0.5
0
0.5
1
r(cm)
Pha
se (
rad)
Semi−infFD Meas
(b)
Figure 3.21: (a) Amplitude and (b) phase vs source detector separation from a healthy female breasttissue showing the measured data (Meas) and fits with semi-infinite (Semi-inf) and finite differencemethod (FD). FD simulates the spread of measurement from semi-infinite due to boundary effect.
and Intralipid (Figure 3.22(c)) were used as a simulated breast. The balloons were inserted into
the breast box filled with the matching fluid to fill a volume occupied by the average breast tissue
volume found in the clinical measurements. A large fraction of balloon was kept above the match-
ing fluid level to simulate the chestwall. The background properties of the matching fluid were µa
= 0.05 cm−1 and µ′s = 8 cm−1 at 786 nm. The retrieved bulk optical properties of fifteen balloon
phantoms using FDM fit are fairly linear in comparison with the expected value as shown in Figure
3.23. Even though there are some deviation, it shows a fairly linear relation.
3.7.2 2nd generation paralell plate DOT system phantom test
First, the frequency-domain part of the instrument was tested on accuracy of retrieving optical
properties of homogeneous medium. Typical matching fluid optical properties fitted from RF semi-
infinite solution with coupling coefficient as described in Section 2.3.2 is presented in Figure 3.24.
The response of the DOT instrument to overall µa change of the medium was tested by ink titration
76
x
y
x
y
20cm
10cm
2.7cm
Background Measurement Geometry
Volunteer Measurement Geometry
11cm
6cm
air(y=-3.6)
air
z
y
source white detector
glass
z
y
source white
detector glass
2.7cmair(y=-3.6)
6cm 6cm
x
y
2.7cm
Baloon Measurement Geometry
11cm
6cm
air
z
y
source white
detector glass
2.7cm
6cm
balloon filled with intralipid and ink held in the tank.
Figure 3.22: Measurement geometry for three cases: (a) background, (b) breast, and (c) tissuephantom are shown with the corresponding boundaries. The shaded regions show the estimatedbreast volume in the segmentation process. The source boundary is diffuse white reflecting and thedetection plane is clear glass.
to the matching fluid (i.e. the tank was filled with matching fluid of varying µa with the method
described in Appendix Section 3.8.1). Figure 3.25 shows good linear response of DOT instrument
up to µa = 0.08 cm−1, which is well within the breast optical properties range.
Then, to test the imaging capability of our instrument, various imaging phantoms were devel-
oped for breast cancer imaging application as shown in Figure 3.26. The first two are mostly for
differential type of measurements, each object simulating tumor with higher absorption or scatter-
ing than the rest of the tissue. Figure 3.26(b) was aimed to provide the flexibility to change optical
77
0.02 0.04 0.06 0.080.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
µa real
µ a rec
on
5 10 15 20
5
10
15
20
µs’ real
µ s’ rec
on
Figure 3.23: Linear response of Scanning DOT to optical properties variation. µa variation (left)and µ′s variation (right)
properties of the object easily by changing liquid solutions and pumping into it. Figure 3.26(c)
shows a silicone-based phantom mimicking a compressed breast with chestwall extension.
We evaluated the resolution performance of our system using point spread function (PSF)
measurements. We used small strongly absorbing point-like objects, specifically 0.4 cm diame-
ter sphers with µa = 2.0 cm−1 and µ′s = 8 cm−1. Two such objects were arranged at r1 = (-2.5,
3.0, 0.0) and at r1 = (2.3, 3.0, 0.0). Reconstructions were performed for a range of regularization
constants lc from 0.001 to 1000. Typically, NRIM (Chapter 2, Section 2.4.1) was used. The re-
constructed µa(786 nm) images from lc = 0.1 are depicted in Figure 3.27. The cross-sections are
evaluated at the FWHM and they are dependent on the regularization constants. There is gener-
ally a trade off between resolution and image noise. By considering the ratio between the object
contrast and the image noise (CNR), the optimal regularization constant can be chosen. For this
experiment, a 3D point spread function with FWHM of 1.1 × 1.13 × 1.1 cm is obtained.
We also evaluated the ability of the system to image an extended tissue volume and to re-
construct a heterogeneous medium with multiple objects. We dispersed 15 cubic silicone tissue
phantom objects of µa = 0.2 cm−1 and µ′s = 8 cm−1 (0.8 cm edge) throughout the measurement
78
1 2 3 4 5 610
−6
10−4
10−2
1 2 3 4 5 60
1
2λ=690 nm µ
a=0.038, µ
s’=8.4 cm−1
1 2 3 4 5 610
−6
10−4
10−2
1 2 3 4 5 60
1
2
1 2 3 4 5 610
−6
10−4
10−2
1 2 3 4 5 60
1
2
1 2 3 4 5 610
−6
10−4
10−2
ρ (cm)
No
rmal
ized
Am
plit
ud
e
1 2 3 4 5 60
1
2
ρ (cm)
Ph
ase
(rad
)λ=750 nm
λ=786 nm
λ=830 nm
µa=0.056, µ
s’=7.6 cm−1
µa=0.057, µ
s’=6.7 cm−1
µa=0.049, µ
s’=7.1 cm−1
Figure 3.24: From top to bottom, 690, 750, 786, and 830 nm. Left to right, amplitude vs separationsand phase vs separations.
volume filled with the matching fluid of µa = 0.05 cm−1 and µ′s = 8 cm−1. The reconstructed
µa(786 nm) images with lc = 0.1 is shown in Figure 3.28 with isosurfaces at µa = 0.092 cm−1. The
reconstructed objects are well separated in all three dimensions. The spatially variant regulariza-
tion has provided fairly consistent sensitivity and resolution throughout the volume.
To evaluate the ability to characterize the concentrations of different chromophores in a tissue
phantom, two 3.2 cm3 flow cells (Figure 3.26(b)) were used. One was filled with an ink contrast
(higher µa than the background matching fluid) and the other was filled with an ICG contrast. Mul-
tiple measurements were performed with the titration of ICG while fixing the ink concentration.
The reconstructed µa at all 4 wavelengths were combined to calculate concentration map of ink and
ICG respectively. The concentration images sucessfully distinguish between the ink and the ICG
79
0.04 0.06 0.08 0.1 0.12 0.140.04
0.06
0.08
0.1
0.12
0.14
Expected µa (cm−1)
Ret
riev
ed µ
a (cm
−1)
Figure 3.25: µa at 786 nm measured and quantified using DOS is plotted with respect to theexpected µa calculated from spectrophotometric measurement.
as shown in Figure 3.29. Moreover, the integrated signals (volume× concentration [cm3µM ]) are
accurately retrieved for ICG concentration variations within the physiological region.
Previous tissue phantoms are all small objects (less than 3.2 cm3) suspended inside the match-
ing fluid. However, in vivo breast measurement involves tissue which has different constituents
from the liquid suspended in the matching fluid. This is simulated by using a breast shaped sili-
cone phantom with an embedded object with tubings as shown in Figure 3.26(c). One can change
the optical properties of the embedded object (which simulates a cancer) by flowing different con-
centration of Intralipid/ink solution. Figure 3.30(a) is the reconstructed images with the reference
measurement being the same silicone phantom at different titration stage (smaller ICG concentra-
tion) in the embedded object. The location of the reconstructed absorber matches well with the
expected location. (Confirmed by slicing the silicone phantom in half.) This simulates the case for
extrinsic contrast injection where the breast without the injection serves as the reference. In this
80
(a) (b) (c)
Figure 3.26: Photos of (a) two 1 cm3 silicone cube, (b) hollow cylindrical container with tubingsenabling changes of liquid inside (flow cell), and (c) breast shaped silicone with extension for thechestwall
case, source coupling coefficients problem diminishes since the signal difference is only coming
from the injection. Figure 3.30(b) is the reconstruction result with the matching fluid as the ref-
erence measurement, with geometric constraint approach. This reconstruction is most similar to
our in vivo breast measurement, where it relies on the intrinsic contrast (Chapter 2, Section 2.4.3
and Chapter 4, Section 4.4.2). The quality of reconstructed image is not as good as the extrinsic
contrast simulation case (Figure 3.30(a)), but the location is fairly accurate. We are in the process
of refining the reconstruction technique to improve image accuracy.
3.8 APPENDIX: Tissue Phantom Recipes
3.8.1 Liquid Phantoms
The absorbers typically utilized in constructing liquid tissue phantoms are ink, ICG, Hb, HbO2,
H2O and lipid. Except the ink spectrum, well established spectra of the absorbers are available in
OMLC homepage (http://omlc.ogi.edu/spectra/, accessed January 2001). In Figure 3.31, the water
and lipid spectras are presented in µa [cm−1]. The spectra of Hb and HbO2 are shown in extinction
81
y (cm)
y (cm)x (cm)
x (cm)
x (cm)
y (cm)
z (cm)
z
0.090.080.070.060.05(cm )-1
8 6 4 2 0-2-4-6-8
5
-10
10
-50
105
0
-5
0
5
-5 0 5
50-5
-8-6-4-2 0 2 4 6 8
z (cm)246
2 4 6
Figure 3.27: Reconstructed µa map of two point spread function like objects with high absorption
coeffiencient ε [cm−1 M] (Figure 3.32). ICG spectra changes depending on the suspension medium
(water or plasma) as shown in Figure 3.33. The absorption spectrum of India ink (Figure 3.34(a))
is measured with the spectraphotometer. It is presented in absorbance Ab(λ) = − log I(λ)I0(λ)
where
I is the transmitted intensity through the absorbing sample of 1 cm thickness and I0 is the initial
light intensity. India ink is typically utilized for its monotonical spectra makes it easy to produce
expected absorption when testing µa response of the instrument at one wavelength. General con-
figuration for liquid tissue phantom measurement in semi-infinite geometry is shown in Figure
3.34(b).
According to van Staveren’s Mie theory approximation [262], µ′s of 10 % Intralipid is given by
82
x (cm) y (cm)
z (cm)
Figure 3.28: Reconstructed µa map of fifteen 1 cm3 objects suspended in the matching fluid
the following
µs = 2.54× 109 × λ−2.4 (3.11)
g = 1.1− 0.58× 10−3λ (3.12)
µ′s = µs(1− g) (3.13)
where λ is in [nm] and µs and µ′s are in [cm−1]. µ′s of 0.8 % Intralipid usually used for our
experiments are presented in Figure 3.35. (There is a consistent discrepancy between this expected
value and the measured scattering value. This may be attributed from the difference in scattering
83
Ink & ICG flowable phantom
Figure 3.29: Respective concentration maps of ICG and ink based on reconstructed µa.
medium we are using (Liposyn III) and van Staveren used (Intralipid).)
Recipe : Standard Ink/Intralipid solution
(a) Calculation of Ink concentration for desired absorption
First, one needs to dilute the ink concentration to either 1 % or 10 %. If the total liquid volume
is less than 10 l, 1 % dilution is adequate. For the optical mammography, we need 26 l of the
matching fluid, so we need the ink of 10 % dilution.
The absorbance of the ink solution to be used should be measured with the spectrophotometer,
ideally on the day of the measurement. Typical dynamic range of spectrophotometer is up to
84
4
13x 10
−3
4
13x 10
−3
Figure 3.30: Reconstructed µa (cm−1) image of a silicone breast shape phantom with an embeddedabsorber. (a) Using a measurement on the same silicone phantom with smaller absorption in theabsorber as the reference, (b) Using a measurement on the matching fluid as the reference.
absorbance of 4. Diluting the ink solution to 0.1 % of the original concentration decreases its
absorbance to measurable level. Measure the ink absorbance spectra with spectrophotometer with
water as a reference sample.
If the measured absorbance is Ab(λ), µa(λ) = ln(10)×Ab(λ). The conversion factor ln(10)
stems from the fact that bases of log are different (i.e. Ab = log10I0I and µa = ln I0I ). This µa is
that of 0.1 % diluted version of the original concentration. Therefore, one needs to multiply 10 for
1 % ink and 100 for 10 % ink.
From the titration relation Cbefore×Vbefore = Cafter×Vafter (C: concentration, V : volume)
can be modified as the following, since µa is proportional to the concentration, µbeforea ×V before+
µinka ×V ink = µaftera ×V after. Especially when calculating the amount of absorber concentration
to be used for desired total µa, one needs to take water absorption into the consideration.
(b) Calculation of Intralipid concentration for desired scattering
µ′10% Intralipids can be obtained from the Equation 3.11. Usually, 20 or 30 % Intralipid solutions
are readily available. Therefore, one needs to scale µ′s values accordingly. The volume of Intralipid
to be used can be calculated by µ′befores × V before + µ′intralipids × V intralipid = µ′afters × V after.
85
600 700 800 900 10000
0.1
0.2
0.3
0.4
0.5
wavelength
Water
abso
rpti
on
co
effi
cien
t(cm
−1)
(nm)
(a)
600 700 800 900 10000
0.01
0.02
0.03
0.04
0.05
wavelength
abso
rpti
on
co
effi
cien
t
Lipid
(nm)
(cm
−1)
(b)
Figure 3.31: µa spectra of (a) water and (b) lipid
(c) Example
Suppose that we want to make a solution of µa = 0.05 cm−1 and µ′s = 8 cm−1 at 786 nm for
total volume of 26 l. We will use 10 % diluted ink and 30 % Intralipid. How much volume of
ink and intralipid solution are needed? Suppose the measured absorbance of 0.1 % ink solution is
Ab(786nm) = 1.50. Then in order to get the absorbance of 10 % ink, a factor of 100 should be mul-
tiplied. Also, there is already absorption due to water to be taken into account, i.e. µH2Oa (786nm)
0× (26000− V 30% intralipid) + 305× V 30% intralipid = 8× 26000
V 30% intralipid = 682ml.
If one wants to increase the absorption of above solution to µa = 0.07 cm−1, then the extra
amount of ink to be added can be calculated by
0.05× 26000 + µinka × V new ink = 0.07× (26000 + V new ink)
V new ink = 1.5ml.
87
600 700 800 900 10000
0.5
1
1.5
2x 10
5
wavelength(nm)
exti
nct
ion
co
effi
cien
t (c
m−1
/M) ICG
ICG in waterICG in plasma
(a)
600 650 700 750 800 850 9000
0.2
0.4
0.6
0.8
1
wavelength (nm)
No
rmal
ized
ab
sorb
ance
ICG in waterICG in plasma
(b)
Figure 3.33: (a) Extinction coefficient of ICG (10 µM concentration). (b) Normalized spectra ofICG in water and plasma.
Recipe : ICG solution
The ICG preparation used for clinical study is prduced by Akorn, Inc, under the tradename of IC-
Green. It is usually available in 25mg IC-Green accompanied by 10ml ampules of sterile aqueous
solvent. Since the molecular weight of ICG is 775, this combination makes 3.2 mM solution. For
ICG injection study, we use 0.25 mg/kg dose. However, depending on the preparation variation,
the actual concentration may be different. For spectrophotometric measurement, diluting 45 µl of
this ICG solution into 10 ml water (∼14 µM ) gives the measurable absorbance.
For example, suppose that we want to make a solution of ∆µICGa = 0.1 cm−1 at 786 nm for
800 ml total volume. There is a significant shift in ICG spectra with increase of concentration.
Figure 3.33(a) is the extinction coefficient for ∼ 10 µM concentration range. For our purpose,
this spectra can be utilized: the extinction coefficient ε at 786 nm is 110210 cm−1/M . How much
88
650 700 750 800 850 900 9500
0.5
1
1.5
2
2.5
3
wavelength (nm)
abso
rban
ce
(a) (b)
Figure 3.34: (a) Absorbance spectra of India ink and (b) General configuration for liquid phantommeasurement in semi-infinite geometry
volume of 3.2 mM ICG solution is needed to make ∆µICGa = 0.1 cm−1 of 800 ml total volume?
3.2× V ICG =∆µICGa
ln(10)× ε × total volume
V ICG = 100 µl.
Recipe : Blood phantom
First, determine the concentration of human blood by spectrophotometer measurement. Usually
1% dilution of blood results in good SNR range of spectrophotometer. Use the measurements at
541 nm, 550 nm and 576 nm to determine the total hemoglobin concentrations Ct. Assuming
the blood sample is a mixture of oxygenated and deoxygenated hemoglobins, decompose the con-
centrations based on three wavelengths using known extinction coefficient of Hb and HbO2 from
Figure 3.32(a).
89
600 700 800 900 10000
2
4
6
8
10
12
wavelength (nm)µ s′ (
cm−1
)
µs′ of 0.8% Intralipid
Figure 3.35: Intralipid µ′s base on van Staveren [262]
Then determine the total volume of blood model (Vt). When diluting the blood, Phosphate
Buffered Saline (PBS) should be used to maintain blood cells at pH=7.4. (The oxygen saturation
of blood diagram shifts with ∆pH = 0.2 difference.) The volume of Intralipid is determined by the
same way as described before. The volume of the blood samples is then determined by Vafter =
Cafter × Vt/Ct.
For deoxygenation, either yeast or N2 can be used. Yeast induces complete deoxygenation
since it consumes the oxygen. 3-5 grams per liter is adequate. Full oxygenation is obtained by
bubbling O2 into the blood model. NIR probe should be placed sufficiently far away from the
bubbling O2, as shown in Figure 3.36. Typical StO2 response measured by DOS is presented in
Figure 3.37. StO2 increased fairly fast when the oxygen was turned on and decreased slowly due
to yeast when the oxygen was off. Large downward arrows indicate the time when the blood was
added to the solution. A cycle of oxygen on and off was repeated at each blood titration. Around 0
% and 90% StO2 were achieved at each oxygen off and on period respectively. The last cycle StO2
90
was noisy since the addition of blood increased absorption significantly as to affect the signal-to-
noise of measurements.
Figure 3.36: Configuration for blood phantom measurement in semi-infinite geometry.
The type of blood (such as whole blood, red blood cells and so on), type of anti-coagulation
agent and expiration date can be found on the cover of the blood bag (from American Red Cross).
For research purposes, we get the blood cells which are not suitable for transfusion. The expiration
date is based on the transfusability. If possible, one needs to get the most recently expired blood.
The blood should be stored in the refrigerator at 1-6 oC. It is usually good up to one month. (Blood
cells can live up to 4 months. However, by the time we get the blood from American Red Cross,
it already consists with portion of dead cells. Live cells make up 10 % of total after 4 months.)
Sampling different sites of the blood bag is possible using sampling site coupler or 3-way stop
cock.
When handling blood, one should wear lab coat, safety glasses and gloves for possible spill or
splash. Also, one handling blood needs to get the Hepatitis shots in advance as well as to take the
training course.
When the experiment is over, dispose needles and pipet tips in the bio-hazard sharps container.
91
Figure 3.37: StO2 response with oxygen supply to blood phantom. StO2 increased fairly fast whenthe oxygen was turned on and decreased slowly due to yeast when the oxygen was off at eachblood titration stage. Large downward arrows indicate the time when the blood was added to thesolution.
Dispose other materials in a bio-hazard bag. The bio-hazard bags need to be autoclaved before
pick-up. Add bleach to the containers exposed to blood and blood model itself. Then drain through
the sink. At UPENN, the Principal Investigator should contact the building administrator to have
SMI (infectious waste hauler) come pick up bio-hazard materials. Also, he has to have an exposure
control plan.
3.8.2 Solid Phantoms : silicone
RTV 12 silicone product (GE Silicones) consists of two parts : RTV 12A (base compund) and
RTV 12C (Curing agent). RTV 12A is 80 % Polydimethylsiloxane, 10 % MQ Resin, Benzene
and Toluene. RTV 12C is 5 % Dibutyl Tin Oxide solution, 20 % Ethyl Silicate 40, 20 % Amino-
propyltriethoxysilane, 5 % 1,2,4-Trimethylbenzene and Naphtha (mineral spirits 60 %). It is quite
92
harmful. Especially one should avoid inhalation and contact with skin and eyes. It is also flamm-
bale with flash point at 32 oC.
The weight ratio of 20:1 of RTV 12A : RTV 12C is recommended by the manufacturer. How-
ever, increased 10:1 volume ratio is used for making silicone model to accelerate the curing rate.
The density of silicone is similar to water, so weight ratio approximates volume ratio. RTV 12
product makes a bit cloudy silicone when mixed. Other products like RTV 615 makes clear sili-
cone. RTV 12 was chosen because of the price being the lowest. Recommended curing time is 72
hours, but 24 hours are enough to pull it out of the mold.
As an absorbing agent, carbon black (Raven 5000 Ultra II, Columbian Chemistry Company) is
used. The mean particle size of the carbon black is around 8 nm. Titanium Dioxide (TiO2) is used
as a scattering agent. We used TiO2 from Sigma (T-8141) for most of the solid phantoms.
Silicone adheres to most of the plastic materials. Even if it does not adhere, sides in contact
with the container do not cure as fast as the exposed side to the air. Therefore we built several kinds
of mold that can be taken apart easily after the silicone is cured. Teflon tape (PTFE Teflon tape :
McMaster Carr) is applied on the surface of the mold to easily separate the silicone when cured.
The amount of absorbing and scattering agent to induce the desirable absorption and scattering
was determined empirically. 9 silicone phantoms of the same breast shape (Figure 3.26(c)) with
varying concentration of carbon black and TiO2 were constructed and measured with 1st generation
parallel-plate DOT instruments (frequency-domain). The absorption formula was extracted from
the linear relationship between phantoms under µa = 0.1 cm−1 as shown in Figure 3.38. However,
we did not get the linear relationship with variation of TiO2 amount and µ′s, so one should not
blindly believe the optical properties expected from the recipe. Unlike the liquid phantom, there
are many places the nonlinearity could arise and thus a lot more work is needed. It is provided as a
93
reference point to start and I recommend anyone using this recipe to measure the optical properties
of these tissue phantoms using independent means.
0 2 4 6 8 100
0.05
0.1
0.15
0.2
carbon black (mg)
abso
rptio
n co
effic
ient
(cm
−1)
Figure 3.38: Silicone (RTV-12) recipe : Carbon black amount vs µa
Recipe: Silicone
Initially, we have tried making the silicone phantom by mixing absorption and scattering agents
into RTV12A. Since RTV12A is too viscous, the agents settle down to the bottom of the cured
silicone. Then we tried mixing agents into RTV12C. This has resulted in homogeneous mixing.
One can calculate the amount of carbon black by
total carbon =µa + 0.0068
0.0203× total volume (ml)
1650 (ml).
The recipe is in a complicated form since it was derived from the fixed breast shape of 1650 ml.
Assuming we use 8 mg carbon in 10 ml solution, volume of carbon solution = total carbon×108 .
94
The amount of TiO2 for scattering is
total TiO2 = 3.6 (g)× µ′s7.5× total volume (ml)
1650 (ml).
Mix total TiO2 into 100 ml of RTV12C (using breast mold).
Example
Breast shape phantom with µa = 0.045 cm−1, µ′s = 7.5 cm−1, Volume = 1650 ml.
1. In 10 ml vial, put 10 ml of RTV12C (←(a)). In 100 ml container, put 100 ml of RTV12C
(←(b)).
2. Put 8 mg of carbon black into (a). In this case, ‘zeroing’ with (a) on a scale and add carbon
black directly into (a). Put 3.6 g of TiO2 into (b).
3. Vortex (a) and (b) for 30 seconds at full speed.
4. Ice bath sonicate (a) and (b) for 1 hour. Ice bath : put ice cube in the sonicator. (because
RTV 12C has very low flash point) Sonication helps breaking the aggregated particles down.
5. Meanwhile, measure 50 ml of RTV12C into a beaker and measure 1500 ml of RTV12A into
a mixing pale.
6. After the sonication, pour 3.2 ml of (a) and all of (b) into the beaker with RTV12C. Mix
well.
7. Pour the mixture of RTV12C into RTV12A. Mix really well. Using high speed mixer is not
recommendable since it traps a lot of air bubbles.
8. Pour into a mold. (pour almost up to the brim.)
95
9. Place the mold inside the descicator or bigger container connected to vacuum pump to pump
the air bubbles out.
10. Take the phantom out after one or two days.
* underlined quantities are to be calculated for different optical property using the formula.
* One needs to use glass bottles for RTV12C. It reacts with plastics.
96
Chapter 4
In vivo Diffuse Optical Tomography of
Breast
4.1 Introduction
The clinical DOT experiments and results on in vivo female breasts are presented in this chapter.
(The clinical motivation for spectroscopical imaging of breast cancer is described in Chapter 1.)
To familiarize the reader with clinical terminology of breast cancer cases, one section outlines
the breast tumor classification, diagnostic and treatment procedure. In the following section, we
quantified the optical properties of healthy in vivo breasts and their dependence on demographic
parameters to assess the inter-patient variation. A tumor contrast index was derived based on
three-dimensional DOT of breast with tumor and was compared with MRI and pathology. Then,
the treatment monitoring capability of DOT was explored for a patient going through neoadjuvant
chemotherapy with comparison to MRI. Lastly, the feasibility of optical blood flow measurements
97
for breast cancer diagnosis and detection was explored and its impact on oxygen metabolism esti-
mation of breast cancer is discussed.
4.2 Breast cancer: clinical side
4.2.1 Classification of breast disease
The classification and the description of breast disease in terms of pathology is summarized from
[164].
There are four categories in benign breast disease which may mimic carcinoma: inflammations,
fibrocystic changes, proliferative breast disease, and benign tumors. Inflammations of the breast
include acute mastitis, periductal mastitis, mammary duct ectasia, fat necrosis and granulomatous
mastitis. Inflammations are rather uncommon. Among these, acute mastitis happens mostly in the
lactating period. Fat necrosis is the liquification of fat and formation of mass related to trauma,
previous surgical intervention, or radiation therapy. Fibrocystic changes include cysts (abnormal
membrane sac containing semitranslucent, turbid fluid), fibrosis (scarring due to rupture of cysts
into the adjacent stroma) and adenosis (abnormal formation or enlargement of glandular tissue).
These changes mimic carcinoma by producing palpable lumps, mammographic densities of calci-
fications, or nipple discharge. Proliferative breast disease such as epithelial hyperplasia, sclerosing
adenosis, and small duct papillomas increase the risk of cancer. Fibroadenoma is a new growth
composed of both fibrous and glandular tissue, which is the most common benign breast tumor.
Fibroadenoma usually occurs in young women.
The malignant breast adenocarcinoma is divided into two groups: In situ (noninvasive) car-
cinoma and invasive (infiltrating) carcinoma. In situ carcinoma includes ductal carcinoma in situ
98
which is the most common type of noninvasive carcinoma and lobular carcinoma in situ. Inva-
sive carcinoma includes invasive ductal carcinoma which is the most common type in all breast
Table 4.2: Average optical properties and physiological parameters from the histograms.
103
10 20 30 40 50 60 700
10
20
30
40
50
60
70
80
90
100
B. Volume (µ M)
B. S
atu
rati
on
(%
)
Figure 4.2: Blood saturation vs blood volume with the dashed lines indicating the ranges for normaltissue from the mean and standard deviation of the healthy breast tissue.
It might be expected that tumors and other diseased tissue are distinguished by the relative
value of their total hemoglobin concentration and blood oxygen saturation. For example, malignant
tumors might be expected to have high blood volume with a low oxygen saturation since both a
higher blood content and higher metabolism are necessary to achieve tumor growth in proliferating
tumor cells [255]. Figure 4.2 shows blood saturation plotted vs blood volume for each breast.
The dashed lines indicate the range of blood volume and blood saturation for normal breasts from
Table 4.2. The error bars for each individual are obtained by the standard deviation of repeated
measurements of the same breast. In order to use endogeneous contrast effectively in DOT, tumor
tissue properties should lie in one of the “other” regions defined on this plot.
The normal state of the tissue covers a wide range of optical properties and it could be expected
that the heterogeneity of the tissues would induce a similar variation within an individual. This
could reduce tumor contrast. The ability to recover all the available physiological information is
crucial in order to relate parameters such as the blood volume and blood saturation to increase
specificity. Ideally an imaging instrument should combine a large number of wavelengths with
104
0 10 20 30 0.9
1
1.1
Minutes
amp
/<am
p>
Breast
0 10 20 30 0.9
1
1.1
Minutesam
p/<
amp
> Intralipid
Figure 4.3: Normalized amplitude measured on breast and Intralipid sample by visiting same 17points ten times every two minutes.
abundant spatial information.
4.3.2 Physiological Noise
Apart from the well characterizable noise due to electronics, optics and positioning of the sources
and detectors, there is as additional noise in the measurements due to the changes in the physio-
logical state of the tissue during the measurements. Respiration, movement, heart beat, blood flow
downstream from the hanging breast are some factors that contribute to this “noise”.
In order to estimate the effect of physiological noise on our measurements, the scanning de-
tector was modified to visit the same point ten times every two minutes. This was repeated for
seventeen different points. The results are shown in Figure 4.3 where we plot time series of nor-
malized amplitude (phase not shown) obtained from measurements on an Intralipid sample and on
a breast tissue in vivo. We find that signals from the Intralipid sample are stable within 1 − 2%
whereas on the breast tissue the dispersion is up to 5− 10%. This provides us with an estimate of
the physiological noise in our experiments.
We also performed repeated measurements of the same breast with minimal movement of the
105
15 20 25 30 35 4010
20
30
40
50
B. V
olu
me
(µ M
)
BMI (kg/m2)15 20 25 30 35 400
20
40
60
80
100
B. S
atu
rati
on
(%
)
BMI (kg/m2)
Figure 4.4: Left: Blood Volume vs BMI with a decaying exponential fit (correlation coefficient0.42), Right: Blood Saturation (correlation coefficient 0.03) vs BMI
breast. By comparing repeated measurements, we find that the average standard deviation is 11%
for µa, 4% for µ′s, 4% for blood saturation and 5% for blood volume (see for example error bars in
Figure 4.2). These values are consistent with the variations in the amplitude and phase observed in
Figure 4.3.
4.3.3 Demographics and Optical Properties
As mentioned above, it might be expected that optical properties would show a variation with
demographics. Pogue et al [219, 220] used a similar system geared towards imaging and reported
that blood volume had a correlation with body mass index (BMI) which is related to the weight and
height of an individual. They did not report any strong correlation between any other quantities
and BMI or age. Cerussi et al [38,39] reported weak correlation between blood volume and b with
age (as well as some other quantities that are not available to us).
Our findings are shown in Figure 4.4 for correlations with BMI. We see a similar correlation
of blood volume to BMI as reported by Pogue et al [219, 220]. A higher BMI indicates more
tissue fat content. In compositional studies a higher fat content correlates with a lower blood
106
15 20 25 30 35 404
6
8
10
12
14
µ s ′ (cm
−1)
BMI (kg/m2)
Figure 4.5: µ′s at 830nm vs BMI with a decaying exponential fit. Other wavelengths show similartrends. The correlation coefficient is 0.46.
10 20 30 40 50 60 7010
20
30
40
50
age
B. v
olu
me
(µ M
)
10 20 30 40 50 60 700
20
40
60
80
100
age
B. S
atu
rati
on
(%
)
Figure 4.6: Left: Blood Volume vs age, Right: Blood Saturation vs age
content [75, 105, 163, 255, 279, 286]. The correlation coefficient is significanly higher for blood
volume than blood oxygen saturation (0.42 vs 0.03).
We also observed a similar correlation of BMI and µ′s as shown in Figure 4.5. Cerussi et
al [38, 39] showed that the scattering power and µ′s change with the fat content. BMI is also a
measure of the tissue fat content hence this result is in agreement with the their reasoning.
Figure 4.6 shows the correlation with age. Our results again indicate an agreement with Pogue
et al [249]; we do not see any clear correlations.
Our study had a different sensitivity than Cerussi et al who showed that there is considerable
107
change in breast properties with age. Their main result is that older breast tissue has a different
water and lipid content which effects the scattering and absorption properties of the tissue. Their
instrument was particularly sensitive to this aspect because it measured mainly the outer ∼1 cm of
the breast tissue and had many wavelengths which allowed accurate derivation of the wavelength
dependence of the scattering. Our results, by contrast, sample a larger tissue volume in transmis-
sion geometry and has a vast spatial information rather than spectral information. Therefore, we
sample the fatty tissue as well as the nodules and vasculature extensively. This is a weakness for
our measured properties in terms of correlating them with demographics, age and hormonal sta-
tus, however, it is imperative to establish these properties in our geometries since most imaging
systems rely on sampling a large volume of breast tissue. It has previously been reported that the
differences in the acquisition geometry changes the computed bulk properties in extensive studies
by Cubeddu et al [60, 61].
4.4 Breast Cancer Imaging
4.4.1 CCD Raw data
For three-dimensional reconstruction of optical and physiological parameters for breast cancer
patients, the second generation parallel-plate CCD-based hybrid instrument was utilized (Chapter
3, Section 3.6.2). The sequential nature of data acquisition produces a set of 2D continuous wave
CCD intensity maps. An example of a set of reference and in vivo breast raw data with a 786 nm
light source is shown in Figure 4.7. The two-dimensional CW intensity images taken with CCD
for three central sources of source plate (Figure 3.19(b)) is shown. There is a significant difference
in the source power between source positions and wavelengths due to non-uniform response of our
108
switch and fiber coupling. Each intensity image shown in Figure 4.7 is scaled by its own minimum
and maximum range. The absolute values between images are therefore not comparable.
The intensity images of the matching fluid in lower row of Figure 4.7 are characterized by its
concentric distribution of intensity with the maximum intensity occuring at the source position.
The intensity images of typical in vivo breast in upper row of Figure 4.7 show significant deviation
from concentric distribution observed from homogeneous matching fluid medium. This deviation
arises from the boundary effect between the breast and the matching fluid as well as the breast
heterogeneity (especially at the boundary where the breast is touching the detection glass).
source 23source 22 source 24
breast measurements
matching fluid measurements
16 cm
11 cm
Figure 4.7: Measured CW intensity images of breast (upper row) and matching fluid (lower row)at source position 22, 23, and 24 (three central sources in the middle row of source plane shown inFigure 3.19(b)).
109
A quick composite look at the attenuation level of the raw breast data with respect to the homo-
geneous reference data can be offered by constructing a transillumination picture. This approach
is inspired by the earlier transillumination works [16,35,108,131,178,183,209,225,247,272,275]
where widebeam or plane-wave illumination were used. We define the detected intensity at detec-
tion position, rd, due to source number, s, to be I(s, rd). We construct a two-dimensional tran-
sillumination picture by summing the contribution of all sources, and normalizing with reference
data I0(s, rd) summed over all the sources, i.e.
T (rd) = −log(
∑Nss I(s, rd)
∑Nss I0(s, rd)
)
. (4.1)
Transillumination may reveal the signal contrast due to the presence of the tumor (Figure 4.8(a),
compare with subject # 68 results in Figure 4.13), but most of time it shows sensitivity to the
presence of surface blood vessels near detector plane (Figure 4.8(b)). In the latter case, the transil-
lumination picture can aid in identifying image artifacts due to surface structures. Black ellipsoidal
lines show the ellipsoid-approximated breast boundaries; outer line is the largest extent of breast
and inner line is the breast touching the detection plane. The attenuation level is higher in breast
than in the matching fluid for T > 0 and vice versa. In Figure 4.8(a), the tumor is located in
middle inner quadrant of the right breast, which correponds to the high attenuating enhancement
in this view. Note the attenuation level is relatively low inside the breast and T ∼ 0 far outside
of the breast, since the region is occupied by the matching fluid. The effect of perturbation due to
presence of breast is seen outside of the breast boundary due to its diffusive nature (i.e. T 6= 0 at
the boundary). In Figure 4.8(b), the tumor is located in middle outer quadrant of the left breast,
which correponds to the left outer side of the breast outline. The nipple is shown at the middle
110
lower and the blood vessel nearby the detection plane is clearly seen, which was also observed in
an outline photograph.
−8 −4 0 4 8
−2
0
2
4
6
8−0.3
−0.2
−0.1
0
0.1
0.2
0.3
(a)
−8 −4 0 4 8
−2
0
2
4
60.2
0.4
0.6
0.8
1
1.2
(b)
Figure 4.8: Transillumination pictures of two breasts. Black ellipsoidal lines show the ellipsoid-approximated breast boundaries; outer line is the largest extent of breast and inner line is the breasttouching the detection plane. All dimensions are in centimeter.
4.4.2 3D DOT reconstruction method
Three-dimensional (3D) images are reconstructed from the data using the multi-spectral recon-
struction method with Envelope-guided spatially variant reconstruction described in Chapter 2,
Section 2.4.3.
Specifically, the geometric constraint is applied as following. We define the unknowns to be
CHb, CHbO2, CH2O, Clipid and A. We fix the b value according to bulk values obtained from the
frequency-domain measurements. Since our particular imaging geometry involves space occupied
by breast and matching fluid, image segmentation of breast and matching fluid was used through-
out the calculation. The breast region was approximated as a 3-dimensional half-ellipsoid based on
its outline in the photo taken prior to measurement scan. To assign background and initial values to
the two regions, the bulk optical properties of matching fluid were derived from frequency-domain
111
reference measurements made on the box when it was completely filled with matching fluid. The
breast optical properties were derived from frequency-domain measurements in contact with breast
surface. A homogeneous semi-infinite analytic solution of the frequency-domain diffusion equa-
tion with multi-spectral approach [81] was utilized to fit directly for the bulk CHb, CHbO2, CH2O,
A and b in each region. The scattering power, b, was allowed to take on different values in the breast
and matching fluid, respectively. Usually, µbackgrounda inside the breast was then fixed as combi-
nation of 31% water or bulk CH2O estimated from frequency-domain measurement and 57% lipid
absorption (from literature [163, 279, 286]). µbackgrounda for the matching fluid region was fixed
at the fitted bulk µa of the matching fluid. µbackgrounda inside the breast was then fixed as combi-
nation of 15 % water (estimated from frequency-domain measurement of bulk CH2O) and 57 %
lipid absorption (from literature [163, 279, 286]).For the initial guess, CHb, CHbO2and A were as-
signed to the breast and the matching fluid based on the bulk measurements, e.g. zero hemoglobin
concentration in the matching fluid region. After the reconstruction of CHb(r), CHbO2(r) and
A(r), 3D images of total hemoglobin concentration (THC(r) = CHb(r) + CHbO2(r)), blood oxy-
genation saturation (StO2(r) = CHbO2(r)/THC(r)) and scattering µ′s(r, λ) = = A(r)λ−b(r) were
constructed.
4.4.3 Image orientation of 3D DOT reconstruction
Figure 4.9 shows the orientation of the three-dimensional reconstructed DOT image. A series
of slices along the y axis are arranged from left to right, from the source plane to the detector
plane, respectively. Each slice represents a 16 × 11 cm image in x-z plane, with the caudal-
cranial view (i.e. from feet to head, same as the CCD camera view). The orientation of each
image is such that the left side of the image slice is lateral (towards outer side of breast) and
112
right side is medial (towards middle of the breasts) for the left breast, and vice versa for the right
breast. For convenience of presentation, slices at selected spatial intervals are presented. Since the
reconstructed data on FEM nodes is interpolated to regular grid of 0.2 cm spacing, each slice has
0.2 cm of pixel size in the x and z directions.
Figure 4.9: Orientation of three-dimensional reconstructed DOT images. Caudal-cranial slice s(foot to head) were arranged left to right, from source to detector plane. The left side of eachimage is lateral and right side is medial.
4.4.4 3D reconstruction images
Representative three-dimensional reconstructed images are shown for three malignant cancer cases
and one benign tumor case. For these cases, 31 % water and 57 % lipid concentration were assumed
and b distribution was fixed according to fitted value from FD measurement. We then reconstructed
CHbO2, CHb andA. The first malignant cancer case shows a cancer near nipple area and the second
malignant case shows a cancer far away from the nipple. The third malignant cancer case shows
two cancers, of which one was subject to core biopsy. In the benign fibroadenoma case, DOT did
not detect the mass in the quandrant where mass was present. For each case, summary of radiology
113
and histopathology report is presented to give information about the tumor. Radiologists who have
experiences in breast cancer MRI compared DOT and MRI and confirmed the position of lesion.
Subject #103 was a 53 year old postmenopausal female with a 2 × 2 × 2 cm subareolar mass.
The cancer was located near nipple of the right breast as shown in Figure 4.10. The mammogram
measured a 1.5 cm mass, ultrasound measured the mass to be 1.1 × 1.3 × 1.8 cm, and MRI
measured it to be 2.2 × 2.1 cm. A representative sagittal MRI containing the nipple is shown
in right subfigure of Figure 4.10. (The enhancement near the chestwall was deemed to be DCIS,
but histology did not find DCIS.) The histopathology analysis after mastectomy revealed invasive
ductal carcinoma behind the nipple of 2 × 2 × 2 cm size and benign proliferative breast within the
axillary tail of 5 × 3.4 × 2 cm size.
Invasive Ductal Carcinoma
Benign Proliferative Breast Disease
5cm
Photo (caudal-cranial view)
medial lateral
Figure 4.10: Tumor location of subject #103 (near nipple). Left: location of tumor in frontal view,Center: location of tumor in caudal-cranial breast outline photograph by CCD, Right : sagittalMRI showing the enhancement of tumor due to high uptake of gadolinium.
Figure 4.11 shows the reconstructed DOT images of total hemoglobin concentration (THC),
blood oxygen saturation (StO2) and reduced scattering coefficient (µ′s) at 786 nm. There is an
enhancement of THC and µ′s near nipple area which correponds to radiologist-confirmed cancer
position. StO2 does not show any enhancement corresponding to the mass, but the StO2 value near
the nipple area is lower than near the chestwall.
114
Total hemoglobin concentration (µM)
28
56
invasive ductal carcinoma proliferative disease?
Blood oxygen saturation (%)
50
100
µs′ at 786nm (cm−1)
5
35
16 cm
11 cm
∆ y = 1 cm sourceplane
detectorplane
Figure 4.11: 3D reconstructed images of THC, StO2 and µ′s at 786nm of subject #103, right breastwith ductal carcinoma
Subject #68: Adenocarcinoma (malignant)
Subject #68 was a 41 year old premenopausal female with an adenocarcinoma in her right breast.
According to X-ray mammography, ultrasound and MRI, the cancer was 4 cm in size and located
in the middle inner quadrant (3 o’clock). In our CCD camera view, the cancer is located at the left
side of nipple near the chestwall as shown in Figure 4.12(b).
High THC and µ′s enhancement of∼2 cm size spans from y = 1 cm to 5 cm, corresponding to 3
o’clock position in Figure 4.13. There are slight shift between radiologist-confirmed mass position
and DOT-enhanced position. However, this could rise from the difference in the compression
scheme between the MRI and DOT as well as the variation in breast positioning.
Subject #111 was a 59 year old female with invasive ductal carcinoma in her right breast. She had
115
(a) (b)
Figure 4.12: Tumor location of subject #68. (a) In frontal view, the tumor is located in middleinner quadrant about 3-4 o’clock of the right breast. (b) In caudal-cranial CCD view, it correpondsto the left side of the breast. All dimensions are in centimeter.
two masses at 2 o’clock (2.0 cm) and at 10 o’clock detected by X-ray mammography, ultrasound
and MRI. Ultrasound-guided core biopsy with 15 gauge needle on 10 o’clock mass was performed
8 days before optical measurement, which found invasive ductal carcinoma with high grade nuclei.
In the optical measurement, bruise due to recent core biopsy was observed in lower outer quadrant
of the right breast (Figure 4.14).
In THC and µ′s images (Figure 4.15), two distinct enhancements corresponding to 2 and 10
o’clock are observed. The extension of 10 o’clock enhancement through the middle sections in y
direction could be attributed to the bruise effect induced by core biopsy. An interesting distribution
of StO2 is noticeable where the biopsied area near the detection plane has much lower oxygen
saturation.
Subject #95: fibroadenoma (benign)
Subject #95 was a 34 year old premenopausal female with a fibroadenoma in her right breast. X-
ray mammography detected a mass at 9 o’clock, 4 cm from the nipple of size (0.6 × 1.2 × 2.0
cm) as shown in Figure 4.16. The histopathology analysis done after the lumpectomy found 1.5
116
Total hemoglobin concentration (µM)
6
13
Adenocarcinoma
Blood oxygen saturation (%)
50
100
µs′ at 786nm (cm−1)
5
25
Figure 4.13: 3D reconstructed images of THC, StO2 and µ′s at 786nm of subject #68, right breastwith adenocarcinoma
cm fibroadenoma.
The reconstructed images in Figure 4.17 do not show noticeable contrast near the mass. This
may be due to lack of vasculature around the fibroadenoma. However, some fibroadenoma may
develop a vasculature. We need to measure more benign cases to assess the feasibility of using
optical contrast to distinguish between benign and malignant cases. There is an image artifact
appearing near source plane, which warrants the improvement of reconstruction scheme in the
future.
4.4.5 Comparison between single and multi-spectral approach
Subject #69 was a 56 year old postmenopausal female with invasive and in situ ductal carcinoma
in her left breast (Figure 4.18). X-ray mammography detected a mass behind the nipple of 1.5 cm
117
Figure 4.14: Tumor location of subject #111. (a) In frontal view, the tumors are located at 2 and 10o’clock of the right breast. (b) In caudal-cranial CCD view, 2 o’clock mass correponds to the leftside and 10 o’clock mass corresponds to the right side of the breast. The biopsied 10 o’clock massis associated with core-biopsy induced bruise extending through lower outer quadrant, as shown indark color in this outline photograph. All dimensions are in centimeter.
spiculated mass associated with pleomorphic calcifications (category 5). MRI also found spicu-
lated mass of 2 cm in size in the subareolar position. The histopathology analysis done after the
mastectomy reported firm 2.1 cm area with adjacent but separate area of 0.4 cm, which are invasive
and in situ ductal carcinoma with local lobular features.
For comparison between single-spectral and multi-spectral reconstruction approaches we re-
constructed CH2O as well as CHbO2, CHb and A for subject #69. (Descriptions of single-spectral
and multi-spectral approaches are in Chapter 2.) The contrast arising from the water concen-
tration has been observed by other groups and its physiological implication has been empha-
sized [140, 259]. Our previous analysis was concentrated in reconstructing CHbO2, CHb and A
due to limited number of wavelengths. However, it is important to consider the water concentra-
tion in the reconstruction and it is our future direction to increase light sources sensitive to water
region. For the existing data, the reconstruction of additional parameter CH2O was explored in the
context of comparing single-spectral and multi-spectral approaches. Lipid concentration was fixed
as 57%. The reconstructed images of THC, StO2, µ′s at 786nm and CHbO2using single-spectral
approach are presented in Figure 4.19 and those using multi-spectral approach are presented in
118
Total hemoglobin concentration (µM)
15
30
invasive ductal carcinoma biopsied carcinoma
Blood oxygen saturation (%)
50
100
µs′ at 786nm (cm−1)
5
30
sourceplane
detectorplane ∆ y = 1 cm
11 cm
16 cm
Figure 4.15: 3D reconstructed images of THC, StO2 and µ′s at 786nm of subject #111, right breastwith multiple ductal carcinomas.
Figure 4.20. In both cases, invasive ductal carcinoma is detected by the enhancement of THC and
µ′s. However, the reconstructed water concentration using the single-spectral approach show nega-
tive concentration, which was probably to compensate a false increase in THC. The multi-spectral
method provides a more robust water concentration within the physiological range.
4.4.6 Tumor contrast
In order to quantify tumor contrast, an average (p) and a standard deviation (σp) was calculated
THC, StO2, µ′s in each image. The tumor region was defined by p > p + 2 × σp, since values
greater than 2 × σp have a 95% chance to be different from the average with the assumption of a
gaussian distribution. For StO2, threshold of p < p−2×σp was considered based on tumor hypoxia
hypothesis; this did not yield any tumor region. Tumor regions estimated from THC and µ′s were
119
Figure 4.16: Tumor location of subject #95. (a) In frontal view, the tumor is located at 9 o’clockof the right breast. (b) In caudal-cranial CCD view, 9 o’clock mass correponds to the right side ofthe breast. All dimensions are in centimeter.
averaged to define the average tumor region. Relative THC (rTHC = THCtumor/THCnormal)
was calculated by averaging THC in the tumor region and outside the tumor region as defined
above. Relative µ′s (rµ′s) and relative StO2 (rStO2) were defined in the same way. rTHC error
bars were estimated based on standard deviation of THCtumor and THCnormal. Relative µ′s (rµ′s)
and relative StO2 (rStO2) error bars were defined in the same way. We define an optical index
(OI) which is a combination of the tumor contrast by OI = rTHC·rµ′srStO2
to explore the possibility
of maximizing optical contrast. The concept of the optical index was first introduced in our com-
munity by Tromberg group [40]. In other imaging modality, it is a common practice to devise an
index to maximize the contrast. Individual values of rTHC, rStO2, rµ′s and OI for breast cancer
images shown in previous section as examples are summarized in Table 4.3.
Table 4.3: Tumor contrast rTHC, rStO2, rµ′s and Optical index OI for individual subjects arepresented.
120
Total hemoglobin concentration (µM)
18
36
Fibroadenoma O
Blood oxygen saturation (%)
50
100
µs′ at 786nm (cm−1)
5
15
Figure 4.17: 3D reconstructed images of THC, StO2 and µ′s at 786nm of subject #95, right breastwith fibroadenoma.
The tumor contrasts (rTHC, rµ′s, rStO2) from twenty-one subjects with carcinomas are plot-
ted in a bar graph in Figure 4.21. The standard errors for total of 22 carcinomas are shown as error
bars in Figure 4.21. Note that rStO2 is around 1 and does not vary much. Combined contrast
index OI is 2.63±0.29.
Using three-dimensional DOT, we observed high tumor contrast compared to adjacent normal
tissue for 21 subjects with breast carcinomas. High THC contrast of tumor has also been reported
by other groups [70, 140, 144, 145, 221, 222, 297, 300], and has been supported by histopathologic
analysis of microvessel density counts [222]. The scattering contrast is still illusive since no careful
comparable histopathologic analysis has not been done for scattering contrast and also due to the
absorption and scattering crosstalk issues. Nevertheless, the hypothesis that the increased number
of scattering organells due to proliferation of cells and the increased fibrosis would increase the
scattering support our finding. Some groups also see some scattering contrast [70], but not to
121
Figure 4.18: Tumor location of subject #69. (a) In frontal view, the tumor is located behind thenipple of the left breast. (b) In caudal-cranial CCD view, it correponds to slightly above the nipple.
the extent we found. Since our CW wavelengths were not optimal, we do expect the presence
of absorption-scattering crosstalk in our reconstructed data. The modification of instrument to
incorporate optimal wavelengths are currently underway.
In the next section, the dynamic changes in the DOT-derived physiological tumor contrast with
a treatment (neoadjuvant chemotherapy) will be presented.
4.5 Neoadjuvant Chemotherapy Monitoring
4.5.1 Introduction
Jakubowski et al [140] have recently demonstrated the capability of diffuse optical spectroscopy
(i.e. as opposed to diffuse optical imaging) for monitoring neoadjuvant chemotherapy in a breast
cancer patient. (The clinical motivation for neoadjuvant chemotherapy monitoring is described in
Chapter 1, Section 1.2.) This important paper introduced a new clinical application to the field.
However, quantification of breast cancer properties from spectroscopic data alone requires assump-
tions about tissues (e.g. homogeneous media, etc.) which reduce the fidelity of their results [26].
In addition, the remission measurement geometry used in their experiments is primarily useful for
palpable, near-surface tumors.
122
Total hemoglobin concentration (µM)
15
50
invasive ductal carcinoma
Blood oxygen saturation (%)
50
100
µs′ at 786nm (cm−1)
5
20
Water concentration (%)
−50
100
Negative water concentration
Figure 4.19: 3D reconstructed images of THC, StO2, µ′s at 786nm and CHbO2of subject #69, left
breast with invasive carcinoma using single-spectral method. Lipid concentration was fixed as 57%. Note the negative water concentration at the cancer site.
The potential of diffuse optical imaging as a chemotherapy monitoring tool has not yet been
explored. We present a case study which demonstrates the feasibility of this approach. We have
utilized a 4-wavelength near-infrared hybrid DOT instrument with continuous wave transmission
and frequency-domain remission detection for breast imaging [62]. Our subject had a locally
advanced invasive ductal carcinoma, and underwent 4 cycles of Adriamycin plus Cytoxan and 4
cycles of Taxotere prior to surgery. DCE-MRI were performed at 3 time points throughout the
therapy. After completion of Adriamycin cycles, we tracked the subject with DOT at 3 time points.
Three-dimensional DOT images of total hemoglobin concentration, oxygenation and scattering
123
Total hemoglobin concentration (µM)
17
26
invasive ductal carcinoma
Blood oxygen saturation (%)
50
100
µs′ at 786nm (cm−1)
5
20
Water concentration (%)
25
50
Figure 4.20: 3D reconstructed images of THC, StO2, µ′s at 786nm and CHbO2of subject #69, left
breast with invasive carcinoma using multi-spectral method. Lipid concentration was fixed as 57%.
were reconstructed. We found that tumor volume and total hemoglobin tumor-to-normal contrast
decreased over the course of neoadjuvant chemotherapy. Furthermore, tumor volume changes
measured by DOT showed good correlation with DCE-MRI measurements of the same subject.
4.5.2 Methods
4.5.2.1 Neoadjuvant chemotherapy & MRI protocol
A 35-year-old premenopausal Caucasian female underwent neoadjuvant chemotherapy for invasive
ductal carcinoma in her left breast at the Hospital of the University of Pennsylvania. The therapy
124
0
0.5
1
1.5
2
2.5
3
Tu
mo
r/N
orm
alrTHC rStO
2 rµ
s’ OI
Figure 4.21: Tumor contrast of 22 carcinomas. rTHC, rStO2 and rµ′s are the relative ratiobetween tumor and normal. OI is an optical index defined as rTHC·rµ′s
rStO2.
consisted of 4 cycles of doxorubicin (Adriamycin, 60 mg/m2) plus cyclophosphamide (Cytoxan,
600 mg/m2, regimen denoted as AC) followed by docetaxel (Taxotere, 100 mg/m2, regimen de-
noted as T ). Herein we will refer to the treatment as “AC” followed by “T ”. Each cycle was
taken at 3 week intervals. She participated in the MRI research study CALGB150007/150012 :
“Contrast Enhanced Breast MRI and Correlative Science Studies to Characterize Tumor Response
in Patients Undergoing Neoadjuvant Chemotherapy for Locally Advanced Breast Cancer (I-SPY)”
(PI: L. Esserman, MD, MBA). The timing diagram for chemotherapy and imaging measurements
(MRI and DOT) is provided in Figure 4.22.
DCE-MRI measurements were performed at the following time points: one week before chemother-
apy (pre-chemotherapy), week 12 following completion of AC, but prior to initiation of Taxotere
(T ) therapy, and week 23 following the completion of Taxotere therapy, but prior to surgical tu-
mor removal (mastectomy). MRI of the breast was performed at 1.5T (General Electric, Signa,
Milwaukee, WI) using in-house sagittal compression receiver coils [135]. At each time point,
the DCE-MRI measurement consisted of sagittal high-resolution thin section three-dimensional
125
-0 3 6 9 12 15 18 21 24
Chemotherapy
Adriamycin+Cytoxan TaxotereSurgery
weeksImaging
MRI DOT
Figure 4.22: Neoadjuvant chemotherapy timing diagram. Four cycles of AC (Adriamycine +Cytoxan) therapy were followed by four cycles of Taxotere therapy, and then by a mastectomy.Arrows indicate timing of MRI and DOT measurements.
T1-weighted spoiled gradient echo imaging of the affected breast performed before, and twice
Figure 4.23: Tumor location. (a) According to X-ray mammogram, ultrasound and MRI, theprimary tumor was located at the 12 o’clock position in the left breast. (b) Photo of breast outline incaudal-cranial view (feet to head view). The location of the tumor was approximated by palpationprior to DOT measurement and is indicated by a circle.
127
4.5.2.3 MRI Data Analysis
Tumor measurements by MRI were performed through analysis of enhancing tumor pixels. Sub-
traction imaging (post-gadolinium minus pre-gadolinium) was employed as required. Maximum
intensity projections in the sagittal and axial planes were obtained to facilitate accurate tumor mea-
surement in three orthogonal dimensions. A radiologist with breast MRI experience measured the
tumor in three orthogonal planes. Tumor volume was then computed assuming an ellipsoid shape.
4.5.2.4 DOT Transillumination
The transillumination pictures of breast at different chemotherapy time points are presented in
Figure 4.24. The transillumination picture offers a quick composite look at the attenuation level of
the raw breast data with respect to the reference data as described in Section 4.4.1. It also enabled
us to identify surface features which may correlate with image artifacts.
After 4th Chemotherapy
After 5th Chemotherapy
After 7th Chemotherapy
Transillumination at 830 nm
16 cm
11 cm
Figure 4.24: Transillumination of breast at 830 nm normalized with respect to reference measure-ment, offering a composite view of attenuation level.
128
4.5.2.5 DOT Data Analysis: 3D Reconstruction
The details of the data analysis scheme is described in previous Section 4.4.2. The reconstruction
volume was a 16 cm× 7.5 cm× 16 cm region, extending into the chestwall area. In this volume, a
finite element mesh with 58087 nodes was used by the finite element method based forward solver
to calculate Φc. Φm was constructed by sampling and smoothing the CCD data on a 41 × 24
grid (total 984 detection points, Figure 3.19(b)) with 3 mm spacing for each source. Hemoglobin
concentrations CHb(r), CHbO2(r), and the scattering prefactor A(r) were chosen as unknowns to
be reconstructed while other variables such as water concentration, CH2O(r) and lipid concentra-
tion, Clipid(r) and scattering power, b(r) were fixed as described below. For image segmentaion
using a geometric constraint, the breast region was approximated as 3D ellipsoid based on Figure
4.23(b). µbackgrounda inside the breast was then fixed as a combination of 15% water (estimated
from frequency-domain measurement of bulk CH2O) and 57% lipid absorption. µbackgrounda for
the matching fluid region was fixed at the fitted bulk µa of the matching fluid. The unknowns to be
reconstructed were CHb, CHbO2and A. After the reconstruction of unknowns at each voxel r, 3D
images of THC(r), StO2(r) and µ′s(r) at 786 nm were constructed.
4.5.2.6 Image correlation analysis between MRI and DOT
As shown in Figure 4.25(a), there are three standard orientation of views in tomography. Since
the compression schemes of MRI (sagittal) and DOT (axial) are different, a true one-to-one image
comparison without distortion is not possible.
In order to compare the tumor positions obtained by MRI with those obtained by DOT, one
must derive a transformation relating coordinates in the MRI axial image to coordinates in the DOT
axial image. We developed a simple scaling transformation for this purpose, which by its nature
129
Coronal
Sagittal
Axial
x
yz
(a)
YMRI
ZMRI
rMRItumor
YDOT
ZDOT
rDOTtumor
Sagittal view
max
max
max
max
(b)
Figure 4.25: (a) Orientation of MRI view. (b) A sagittal view of MRI (top) and DOT (bottom)showing differences in compression. A linear scaling transformation scheme was utilized to findthe tumor center with respect to the nipple in DOT image (rtumorDOT ) from MRI rtumorMRI .
cannot account for all of the deformations arising from breast compression in orthogonal directions.
The schematic of the transformation is illustrated in Figure 4.25(b). First, the breast dimensions
of corresponding ‘central’ image slices (i.e. slices containing the nipple) were compared. We
derived linear scale factors from the ratio of breast length and breast width in these corresponding
central slices. In particular we defined scale factors α = XmaxDOT /X
maxMRI , β = Y max
DOT /YmaxMRI , and
γ = ZmaxDOT /ZmaxMRI , where Xmax
i , Y maxi , Zmaxi is the longest linear dimension of the breast in the
X , Y , Z direction of the ith (MRI or DOT) image.
130
We located the tumor center and tumor boundaries in the MRI image by finding the region with
high intensity due to large gadolinium uptake. We then rescaled the tumor center coordinate rtumorMRI
to rtumorDOT by multiplying by scaling factors (i.e. α, β and γ). The tumor boundaries are defined
in this rescaled fashion as well. Importantly, the tumor in the DOT image lies approximately
within the volume defined by the rescaling based on pre-chemotherapy MRI. We note this scaling
approach assumes linear deformation and does not account for elastic differences between tumor
and surrounding tissue. Future models will be developed to account for elastic variations and
deformation due to the different compression schemes.
4.5.2.7 DOT tumor volume estimation
To define a tumor volume in the DOT images, a threshold based on the standard deviations of the
reconstructed parameters was introduced. First, DOT parameters were decomposed into µa and µ′s
at 690, 750, 786 and 830 nm. Then for each image (THC, StO2, CHb, CHbO2, and µa, µ′s at the
4 wavelengths) an average (p) and a standard deviation (σp) was calculated. The tumor region was
defined by p > p+2×σp, since values greater than 2×σp have a 95% chance to be different from
the average with the assumption of a gaussian distribution. The average tumor volume (and stan-
dard deviation) was calculated by averaging all tumor regions defined by all images except StO2
(where the variation did not exceed 2 × σp). Relative THC (rTHC = THCtumor/THCnormal)
was calculated by averaging THC in the average tumor region and outside the tumor region as
defined above. rTHC error bars were estimated based on standard deviation of THCtumor and
THCnormal.
131
4.5.3 Results
According to the X-ray mammography, ultrasound, and MRI, the primary cancer was located at the
12 o’clock position in the left breast as shown in Figure 4.23. The mammography and ultrasound
reported adjacent multiple masses, the largest being 2.1× 2.2× 2.1 cm in size at pre-chemotherapy
time points. Pre-chemotherapy DCE-MRI measured the tumor size to be 5.3 × 2.2 × 2.7 cm.
MRI has been shown to be more accurate in depicting the size and overall extent of tumor than
mammography or ultrasound [30, 92, 186]. Tumor size estimation by DCE-MRI at different time
points is summarized in Table 4.4. The approximate tumor location in prone position assessed by
palpation was ∼ 8.5 cm from the nipple along the breast contour. In our camera view (caudal-
cranial view), this tumor location corresponds to a distant upper central position from the nipple as
shown in Figure 4.23(b).
Time Tumor size Tumor volumepre-chemotherapy 5.3 × 2.2 × 2.7 cm 16.5 cm3
after completion of AC cycles 3.8 × 1.4 × 1.8 cm 5.0 cm3
after completion of T cycles 3.0 × 1.6 × 1.3 cm 3.3 cm3
Table 4.4: Tumor size measured with DCE-MRI at different time points during neoadjuvantchemotherapy (AC : Adriamycin + Cytoxan, T : Taxotere). Tumor volume was estimated byassuming ellipsoidal tumor shape.
The transillumination images in Figure 4.24 show high signal attenuation at the surface blood
vessel and around the upper central region particularly for data taken after the 4th chemotherapy
cycle. This surface blood vessel was first identified through observation while positioning and was
confirmed with DCE-MRI images. The transillumination picture is quite sensitive to such surface
features.
Three-dimensional DOT images of total hemoglobin concentration (THC), reduced scattering
132
coefficient (µ′s) at 786 nm and blood oxygen saturation (StO2) are presented in Figures 4.26, 4.27,
4.28 respectively. In each figure, images from left to right correspond to 3D DOT image slices
taken from source to detector planes. The DOT images corresponding to different time points
within the chemotherapy cycle, are arranged from top to bottom. Colorbar scales are fixed from 5
to 15 cm−1 for µ′s at 786 nm, and from 50 to 100% for StO2. However, colorbar scales for THC are
not fixed and are adjusted at each time point to maximize the THC color contrast between tumor
and normal region. The THC scale was adjusted so its maximum value was twice its minimum
value; this preserves the percentile changes in the colorbar.
In the THC images (Figure 4.26) after the 4th chemotherapy cycle, a high THC region is
found in slices near the source plane (1 - 3 cm deep from surface) and near the upper central
part of the breast. The lesion position corresponds to the tumor location estimated initially by
palpation at 12 o’clock and 8.5 cm away from nipple (as shown in Figure 4.23). After the 5th
chemotherapy cycle, the tumor region is still identified by THC contrast near the source plane and
in the upper central region. However, the contrasted region appears smaller than the corresponding
region after the 4th chemotherapy cycle. Also, the average THC decreased significantly (i.e. from
21.4 ± 1.4 µM to 9.1 ± 0.5 µM). The THC distribution after the 7th chemotherapy cycle is more
homogeneous throughout the slices compared to previous chemotherapy cycles. Within the original
tumor margins, the high THC region shifts towards outside of the tumor, leaving a relatively low
THC region occupying most of the tumor extent. The average THC increased slightly between 5th
and 7th chemotherapy from 9.1 ± 0.5 µM to 12.5 ± 0.6 µM.
Figure 4.27 exhibits a similar trend. It shows higher µ′s values in the tumor region, and this
region with high µ′s shrinks over the course of treatment. The µ′s range does not vary as dramati-
cally as the THC, and gives a higher contrast ratio between tumor and normal tissue. Slices near
133
15
30After 4th Chemotherapy Total Hemoglobin Concentration ( µM )
Invasive Ductal Carcinoma
6
12After 5th Chemotherapy Total Hemoglobin Concentration ( µM ) x
z
8
16After 7th Chemotherapy Total Hemoglobin Concentration ( µM )
16 cm
11 cm
∆ y = 1 cm
source plane(head)
detector plane(feet)
Figure 4.26: Three-dimensional reconstructed total hemoglobin concentration images. Imageslices from source to detection plane are presented at 1 cm intervals in caudal-cranial view, fromleft to right. DOT images corresponding to after 4th, 5th and 7th chemotherapy were arrangedfrom top to bottom.
the detection plane are affected by artifacts related to large vessels. These artifacts bear close re-
semblance to the transillumination picture in Figure 4.24, which is sensitive to large blood vessels
on the surface near the detection plane. This effect is also apparent in the THC images, but to a
much lesser degree. Therefore, slices near the detection plane (i.e. within 1 cm) were excluded in
calculations of average values of THC, StO2 and µ′s.
StO2 images in Figure 4.28 were relatively homogeneous and do not show contrast in the tumor
region (i.e. not exceeding 2 × σp). Note, however, the overall StO2 value decreased significantly
after the 5th chemotherapy cycle and remained constant thereafter (i.e. from 81.2 ± 1.4% to 59.9
± 0.6% to 61.1 ± 1.0%).
MRI images projected in sagittal and axial views are shown in Figure 4.29 for the pre-chemotherapy
point, after AC therapy (week 12) and after Taxotere therapy (week 23). The intensity ranges were
134
5
15After 4th Chemotherapy µs’ at 786 nm (cm−1)
5
15After 5th Chemotherapy µs’ at 786 nm (cm−1)
5
15After 7th Chemotherapy µs’ at 786 nm (cm−1)
sourceplane(head)
detectorplane(feet)
Figure 4.27: Three-dimensional reconstructed images of µ′s at 786 nm. Image slices from sourceto detection plane are presented at 1 cm intervals in caudal-cranial view, from left to right. DOTimages corresponding to after 4th, 5th and 7th chemotherapy were arranged from top to bottom.
fixed among images at different time points. The intensity of the DCE-MRI image is higher in the
tumor due to increased tumor vascularity and gadolinium contrast uptake. Before chemotherapy,
the tumor is clearly seen around 12 o’clock to 1 o’clock (upper quadrant in sagittal view, near center
in axial view). After completion of chemotherapy cycles, the size and intensity of the enhancing re-
gion decreased significantly. Upon completion of chemotherapy, MRI demonstrated an amorphous
5 - 6 cm non-enhancing soft tissue region with only a scattered punctate form of enhancement. The
majority of the visible mass at this time point was poorly enhancing and was deemed to represent
fibrosis. The tumor position and volume measured by DCE-MRI is summarized in Table 4.4.
Histology revealed extensive fibrosis with focal areas of inflammation, but with few vessels.
However, viable tumor cells, morphologically identical the original core biopsy specimen (obtained
prior to chemotherapy) were identified. They were diffusely scattered throughout the fibrotic region
Figure 4.28: Three-dimensional reconstructed blood oxygen saturation images. Image slices fromsource to detection plane are presented at 1 cm intervals in caudal-cranial view, from left to right.DOT images corresponding to after 4th, 5th and 7th chemotherapy were arranged from top tobottom.
as individual cells and small groups. No macroscopical viable tumor mass was identified.
In both MRI and DOT, the tumor was found in 12 o’clock, many centimeters away from the
nipple. In an attempt to assess the correlation between MRI and DOT images, the simple scaling
transformation scheme described earlier was performed with the assumption that nipple and tumor
center are good common reference points. The DOT axial image slice corresponding to rescaled
tumor center, rtumorDOT (with respect to nipple position), is shown along with the corresponding MRI
axial image slice in Figure 4.30(a). The tumor center found in MRI (X) and DOT (square) agrees
well with employment of this transformation. To show the extent of the tumor in detail, the axial
DOT images with smaller y intervals (0.4 cm) are shown in Figure 4.30(b). Higher THC contrast
in DOT are distributed in three-dimensions approximately within the tumor margins defined from
the scaled MRI data.
136
Sagittal view Axial view
Pre-chemotherapy
After completion ofAC cycles
x
zz
y
7.5 cm
9 cm
7.5 cm
18 cm
After completion ofTaxotere cycles
Figure 4.29: Dynamic-contrast enhanced MRI images of left breast. From top to bottom, rep-resentative sagittal and axial slices along highest tumor contrast are shown for each time point :pre-chemotherapy, after completion of AC cycles, and after completion of Taxotere cycles.
The tumor volume was defined independently for MRI and DOT. MRI tumor volume was de-
termined by a radiologist, whereas DOT tumor volume was determined by thresholding from the
distribution of DOT parameters. Figure 4.31(a) shows the decrease of tumor volume with progres-
sion of chemotherapy. MRI results show the tumor size decreased significantly after completion
of AC cycles. There is a general trend of decreasing tumor volume in MRI and DOT respectively.
The degree of change is different between MRI and DOT, but direct comparison is not possible
because the measurement time-point mismatch. The coinciding time-point is just after the 4th
137
YMRI
ZMRI
rMRItumor
YDOT
ZDOT
rDOTtumor
x
Z MRI
MRIX
DOTX
ZDOT
max
max
max
max max
max
max
max
(a)
15
30After 4th Chemotherapy cycle Total Hemoglobin Concentration ( µM )
∆ y = 0.4 cm Sourceplane (y=0)
y=3.2 cm
(b)
Figure 4.30: Correlation of tumor position in three-dimensions. (a) Tumor center position fromMRI image (a X in the axial slice) is transformed by a linear scaling scheme to the DOT coordinates(a square in the corresponding DOT axial slice). The scaled tumor center (a square) lies within ahigh total hemoglobin concentration region. (b) The scaled tumor boundaries around the tumorcenter is superimposed in three-dimensional DOT images. Axial images shown with smaller yintervals (0.4 cm) show the extent of the tumor in detail.
chemotherapy cycle (completion of AC), where tumor volume measured by DOT is greater than
tumor volume measured by MRI. We note that DOT tumor volume can be overestimated due to
the ill-posed nature of DOT, which introduces blurring [62]. In DOT, there was a significant tumor
volume decrease in each chemotherapy cycle accompanied by rTHC decrease in Figure 4.31(b).
Significant decreases of rTHC were observed after the 5th chemotherapy cycle, which is the first
Taxotere chemotherapy. The rTHC after 5th and 7th chemotherapy cycles are equivalent, but
138
accompanied with significant tumor volume decrease, implying tumor neo-vasculature regression
with chemotherapy.
0 3 6 9 12 15 18 21 240
5
10
15
20
25
30
35V
olum
e (c
m3 )
Weeks after 1st Chemotherapy Cycle
Tumor Volume
MRIDOT
1st 2nd 3rd 4th 5th 6th 7th 8th chemotherapy
DOT 0.21 ± 0.10 cm3
AC cycles T cycles
(a)
0 3 6 9 12 15 18 21 241
1.1
1.2
1.3
1.4
Weeks after 1st Chemotherapy Cycle
THCtumor
/THCnormal
DOT
rTH
C
(b)
Figure 4.31: (a) Decrease of tumor volume quantified by DOT and MRI, (b) Change of rTHC inthe tumor volume. Significant decrease in rTHC is noted before and after the 5th chemotherapy.
139
4.5.4 Discussion
We have demonstrated three-dimensional DOT images of total hemoglobin concentration (THC),
tissue blood oxygenation (StO2), and scattering coefficient are useful for localizing, quantifying
and tracking breast cancer based on vascularity during neoadjuvant chemotherapy. THC and scat-
tering showed localized contrast between the cancer and the normal region. The THC increase in
tumor is expected due to angiogenesis accompanying tumor growth [255]. The enhancement of
scattering might be expected based on changes in nuclear size and on the increase in concentration
of organelles such as mitochondria due to the high metabolism in cancer cells [255]. Generally,
StO2 might have been expected to be lower in the tumor region due to high oxygen demand, how-
ever this effect was not apparent.
The localized changes of these physiological parameters over time clearly demonstrate the
dynamic imaging capability of the DOT method. The DOT measurements were carried out at time
points just after completion of AC (Adriamycin+Cytoxan) cycles and then just after the first and
third Taxotere cycles. The anti-vascular and apoptotic effects of taxane [104] are consistent with
the significant decrease in tumor vasculature as measured in the tumor THC contrast, the tumor
volume decrease, and the intensity decrease from DCE-MRI. Interestingly, the variations in average
THC appear to be qualitatively correlated with measurements of patient hematocrit over the same
time period, which varied from 38% to 33% to 36%. The trends observed by DOT and MRI are
consistent with pathologic findings about the effectiveness of chemotherapy. The carcinoma cells
in the post-chemotherapy stage were finely and diffusely dispersed in fibrous connective tissue
which represented the bulk of the residual mass. This remaining viable tumor was detected by MRI
as focal enhancements and by DOT as small but positive THC contrast. Even though there was
no significant spatial contrast in StO2 for this subject, a significant decrease of average StO2 was
140
observed over time; this type of information might also be useful for chemotherapy monitoring,
since tumor response can depend on the oxygenation of the tumor and surrounding tissues [265,
266].
Jakubowski et al [140] reported THC and H2O as major contrast factors, but did not report the
behavior of scattering. In our case, water concentration was estimated by the frequency-domain
measurements and was then fixed for continuous-wave image reconstruction. In addition to fix-
ing the water and lipid concentration, we have tried fitting for the water concentration, but the
reconstructed water concentration did not change much from its initial value. This may be because
our measurements used only 4 wavelengths (< 900nm), all of which were outside of the water-
sensitive range. As for the scattering contrast, it is possible that some portion of this signal may
arise from absorption-scatter image cross-talk, since the wavelengths used to discern chromophores
and scattering were not optimized [58]. The scattering artifact near the detection plane resembled
the blood vessel in transillumination, and thus raises concerns about scattering-absorption cross-
talk. However, our 3D simulations with noise in the same geometry find less than 20% cross-talk
between THC and scattering, whereas the observed scattering contrast (averaged over tumor vol-
ume) amounts for more than a 60% increase over the normal tissues. We therefore suspect that the
observed scattering contrast may indeed originate, at least partially, from tumor physiology.
To improve quantification of tumor optical contrast and validate the DOT method in clinical
settings, several improvements are underway. Laser sources at optimal wavelengths for separation
of chromophore contributions including water and scattering should further improve quantification
accuracy. Currently, the imperfect tumor region correlation between MRI and DOT data arises
because the linear approximations we have made for rescaling image do not account for the shift
141
of tumor due to compression and deformation of the breast. Better modeling of the elastic defor-
mation of breast in different compression geometries is under investigation. For better correlation
between DOT and MRI, it would be ideal to synchronize measurement time and employ the same
compression schemes. In this present study, it was not possible to synchronize DOT and MRI
due to logistics of two separate research protocols and the constraints of patient availability. The
improved coordination of DOT and MRI is now being performed for neoadjuvant chemotherapy
patients. The measurement of additional pathologic parameters such as microvessel density or nu-
clei size will be invaluable in correlating with tumor vasculature and scattering factors and thus
confirming DOT findings; these parameters are not routinely measured in the clinic. Finally, with
respect to monitoring, more frequent DOT measurements should improve the therapeutic value of
this technique.
4.5.5 Conclusion
We have demonstrated 3D diffuse optical tomography for monitoring physiological tumor re-
sponses during neoadjuvant chemotherapy in a single patient with a locally advanced breast cancer.
We have also compared our results to dynamic contrast enhanced magnetic resonance imaging and
pathologic analysis of the same patient. Three-dimensional reconstructed total hemoglobin con-
centration and scattering images successfully localized tumors and quantified the tumor volume
decrease and the THC contrast decrease over the course of chemotherapy. These physiological
parameters, measurable by DOT, may help to improve our understanding of chemotherapy mech-
anisms, and hold potential to play a role in assessment of treatment response.
142
4.6 Optical measurement of Blood flow in breast cancer
Until now, blood flow has been a quantity that is inaccessible to deep tissue optical methods. Our
laboratory has pioneered the development of diffuse correlation spectroscopy (DCS) for measure-
ment of blood flow in deep tissues (for a recent review see Durduran [80]) carrying out extensive
validation and demonstrating its utility. In the following, we explore the blood flow contrast of
breast tumors following an approach to DOS advanced by Tromberg’s group [139,140] which uses
hand-held spectroscopic probes. They have demonstrated its utility in measuring both the static
tumor contrast as well as changes in contrast due to chemo-therapy, hormonal status and age. Such
hand-held probes are attractive since they allow development of portable instruments and therefore,
frequent measurements with little patient disconfort. Our approach in this work is similar and uses
a hand-held probe that is scanned over the normal tissue and the palpable tumor regions, except we
measure blood flow.
Blood flow in breast cancer is an important quantity to monitor which also provides a novel
contrast over methods that measure essentially tumor morphology. Some amount of differentiation
capability was demonstrated between malignant and benign tumors [170] as well as the ability to
monitor various therapies [71, 79, 172]. Therefore, we expect that diffuse correlation spectroscopy
to be of value since the instruments are inexpensive, portable and the signals are robust. It is further
known that the metabolic changes can preceed dimensionally measurable changes [31] which have
been accessible to traditional imaging or clinical palpation methods. DCS can also be readily
combined with DOS measuring oxygenation and therefore, calculating the oxygen metabolism of
the tumors making this parameter more accessible.
There have been previous attempts to measure blood flow in the breast using PET [17,283,284],
color and power doppler ultrasound [59,149,170] and MRI [71]. PET studies were limited in extent
143
but have shown that blood flow tends to increase in malignant tumors. Ultrasound techniques
on the other hand were used from late 80s to end of 90s and have been inconclusive as to the
clincial utility of the technique. PET was limited because of its cost and availability, whereas
ultrasound techniques had poor signal-to-noise and low contrast. Ultrasound techniques are also
biased towards large vessels and therefore to issues such as arterio-venous shunting. MRI studies
require research instrumentation and are limited by their signal-to-noise.
In this investigation, we have recruited three subjects with palpable tumors, two subjects with
mammographically identified calcification and two healthy subjects. The measurements were car-
ried at Hospital of University of Pennsylvania and were approved by the Internal Review Board.
Briefly, subjects were asked to lay back in the supine position, thus flattening the breast and in-
creasing tumor accesibility. An experienced researcher marked the position and the extent of the
tumor on a transparency paper with a grid for the record of tumor position. Then she used the
hand-held probe shown in Figure 4.32 to scan in horizontal and vertical directions in 2 cm incre-
ments across the tumor. Two scan directions were used to check the repeability of the signal and
ensure that variations were not due to changes in the probe pressure. In the case of healthy volun-
teers, an arbitrary region was drawn as the tumor site and a measurement was obtained by scanning
across that region. This measurement provided information about heterogeneity of blood flow in
breast tissue. Average optical properties necessary for analysis were obtained from separate DOS
measurements of same patients as explained elsewhere [49].
Details of the instrument are described elsewhere [80]. Briefly, a long coherence laser (Crysta
Laser, NV) operating in continous wave mode at 785nm is coupled to a 1×4 optical switch (Di-
con, CA) and used to serially switch between four source positions. Four fast, photon-counting
avalanche photodiodes (Perkin Elmer, Canada) coupled to four single mode fibers were used to
144
2.5 cm
Hand-heldProbe
DCS Dets.Sources
Tumor
Figure 4.32: Hand-held probe with four source-detector pairs is scanned horizontally and verticallyin 2 cm increments spanning the estimated tumor region as well as the surrounding healthy tissue.
detect the intensity fluctuations of the surface speckles. The TTL output is fed to a four-channel
custom build correlator board (Correlator.Com, NJ) and the resulting intensity auto-correlation
functions are recorded by a computer. A complete set of data is acquired every 6 seconds and five
such sets are acquired at each position. For this study, we disregard the crossed source-detector
pairs and record the position of each of the four source-detector positions directly across from each
other (separation of 2.5 cm) from each scanned position. The recorded correlation functions are
then fit to a solution of the diffuse photon correlation correlation equation [80] to obtain an index
proportional to the blood flow. The results are normalized to the mean value of the measurements
of the healthy tissue and the standard deviation is reported as the error bar. We, therefore, report
the averaged relative blood flow (% rBF) at each position.
Figure 4.33 shows four correlation curves from two patients. When blood flow increases, the
temporal auto-correlation function decays more rapidly. It is evident that the blood flow is larger
in the tumor region (compare dark and light curves) in both cases. Figure 4.34 shows horizontal
and vertical profiles from one malignant tumor and one healthy breast. There is very little variation
observable in the healthy breast, where as the blood flow increased in both directions over the
145
−6 −5 −4 −3 −21
1.1
1.2
1.3
1.4
log(τ) (sec)
g 2 (τ)
TumorNormal
(a)
−6 −5 −4 −3 −21
1.1
1.2
1.3
1.4
log(τ) (sec)
g 2 (τ)
TumorNormal
(b)
Figure 4.33: Temporal auto-correlation curves measured in tumor (dark) and healthy (light) tissuefrom two patients. Faster decay corresponds to increased blood flow.
tumor indicating that the observed contrast is due to the tumor and not because of the natural
heterogeneity of the breast.
In order to quantify the blood flow change in the tumors, we have used the estimated tumor
outline and tabulated the mean (± standard deviation) rBF in that region. Table 4.5 shows the dis-
tribution of the values for all subjects. Three groups are visible; (1) there is very little heterogeneity
in the healthy breast (2.7 % variation), (2) the blood flow of malignant tumors is increased to 230
% of healthy tissue, whereas, (3) there is only a moderate increase in benign tumors (to 153 %).
Although, the power of the statistics of this study is not enough to conclusively claim differentia-
tion, we note that these results are in qualitative agreement with previous Doppler ultrasound and
PET results [17, 59, 149, 170, 283, 284] where ∼ 470-550 % increases in blood flow were reported
in malignant tumors with smaller contrast in benign cases. In studies with larger populations, blood
flow indices were used to differentiate upto nine different types of breast diseases [170].
These findings clearly demonstrate that we are able to optically detect robust changes in blood
flow in palpable tumors. Futher studies with more source-detector pairs are now being undertaken
to analyze the potential partial volume effects that may influence our results. We note that these
146
−6 −4 −2 0 2 4 650
100
150
200
250
300
350
Position
rBF
HealthyPatient
(a)
−6 −4 −2 0 2 4 650
100
150
200
250
300
350
Position
rBF
HealthyPatient
(b)
Figure 4.34: Relative blood flow (rBF) scans from one patient with a malignant tumor and a healthyvolunteer are shown for both (a) horizontal and (b) vertical scans.
Table 4.5: Tabulation of relative blood flow (rBF) measured at the estimated tumor regions fromall subjects grouped as healthy, benign and malignant diseases.
palpable tumors are relatively superficial and previous optical studies [139, 140] have shown that
source detector separations around 2.5 cm can probe them in a repeatable manner. Additionally,
if we assume that the palpable region corresponds to roughly the same depth from the skin, the
partial volume effects are further divided out by normalizing to the healthy tissue blood flow. In
the future, we will acquire data with a hybrid instrument [80] in order to measure the oxygenation
and total hemoglobin concentrations changes simultanously and estimate the changes in oxygen
metabolism of the tumors. The instruments are built on small clinical carts and the study time is
relatively short (∼ 10 minutes). Therefore, it is feasible to acquire data at each patient visit and in
the triage area. We anticipate these methods will be clinically useful for therapy monitoring, dose-
adjustment and potentially for assesing the efficacy of the therapy from the first week, therefore,
avoiding unnecessary discomfort to the patients.
4.7 Summary and Future Outlook
Using DOT/DOS technique, we have quantified in vivo healthy breast properties and demonstrated
the detection and quantification of breast tumor contrast against the surrounding nondiseased tis-
sue. Also, treatment monitoring capability of three-dimensional DOT was demonstrated for a case
148
of locally advanced breast cancer going through neoadjuvant chemotherapy. Blood flow measure-
ments of breast shed light as an additional optical parameter to aid in classification of breast cancer.
The theoretical and experimental techniques involving DOT/DOS are constantly evolving to
better quantify the breast tumor properties. For our current CCD-based instrument, our reconstruc-
tion is largely based on CW component of data, where we need to consider crosstalk between ab-
sorption and scattering. Our current wavelengths are not optimal for separating this crosstalk [57].
However, the simulation using current wavelength only yielded 20 - 30 % crosstalk between THC
and scattering, which implies our scattering contrast is not wholly due to the crosstalk. Care-
ful tissue phantom studies using in vitro hemoglobin will help identify this issue, but the origin
of scattering contrast in in vivo measurement still is obscure. It is shown that microvessel den-
sity correlates well with total hemoglobin concentration measured by DOT [222]. Retrospective
histopathology correlation study with emphasis on the microvessle density and other parameter
highlighting the source of scattering would be beneficial linking the microscopic histology and
macroscopic DOT measurement.
Another technical difficulty we have faced during clinical measurements is the current and
intrinsic limitation of parallel plate design to accommodate lesions near axillary tail and chestwall.
Also, near the chestwall, the side of the breast is exposed to the boundary of air and the matching
fluid, which needs careful modeling. We are currently in the process of adding optimal wavelengths
and optimizing the table design to enable more breast tissue to be inside our imaging volume
without being exposed to the air boundary. In the extreme case of breast cancer near axillary tail
and chestwall, we are utilizing the portable DOS/DCS instrument.
In some in vivo measurements, the reconstructed images could be contaminated with source
and detector artifacts. These may be coming from source detector coupling mismatch between
149
the breast and the matching fluid. In addition to the use of source detector coupling fit, various
approaches are devised to reduce these effects which may degrade the quantification of breast
cancer. The quantification of breast cancer could be substantially improved using a priori spatial
information from other imaging modalities [33, 34, 165, 294]. In our laboratory, initial phantom
measurements incorporating the scanning ultrasound has shown good quantification capability.
Incorporation of ultrasound to accommodate clinical measurement geometry is anticipated to give
additional capability to our current DOT scheme.
We are currently in the beginning stage of new breast cancer imaging technology where we
are validating our results against the gold standard or the existing imaging modalities such as X-
ray mammography, ultrasound and/or MRI. In section 4.5, the correlation study between MRI
and DOT on a locally advanced breast cancer showed reasonable agreement even with use of
simple scaling approach. More elaborate modeling considering the elastic deformation of breast
is needed, which may require the hardware modification and careful coregistration scheme using
markers. Of course, the measurements of MRI and DOT in the same table would be ideal for
coregistering purpose. With careful design, CCD-based DOT system could be also incorporated
into MRI environment.
We have briefly mentioned the construction of the optical index based on the relative tumor
contrast of optical properties for carcinoma cases. Additional measurement such as blood flow
could be added into the index. Or the metabolism derived from the blood flow and saturation can
be added as another parameter. This concept of optical index could lead to identification of optical
properties which discern the benign and malign tumors. More in vivo data on various kinds of
tumor is needed to explore the possibility.
150
Neoadjuvant chemotherapy setting has emerged as an ideal platform to demonstrate the moni-
toring capability of DOT. With recently established collaboration, more coordinated measurements
of DOT with MRI are being performed in our laboratory. However, more frequent DOT measure-
ments are desirable since the significant physiological changes relating to treatment efficacy may
happen within a week [140]. Then the treatment efficacy could be assessed in the similar ap-
proach taken by Yu et al [292], Wang et al [273] and Sunar et al [250]. We have recently acquired
DOT measurements with Indocyanine Green (ICG) injection for some neoadjuvant chemotherapy
patients. In addition of enhancement of optical contrast, the kinetics of ICG could reveal the infor-
mation on the blood flow characteristics in the breast tissue. With current instrument, we can only
follow the kinetics on fixed source position. More rapid instrument enabling switching between
multiple source position during 5 minutes interval would be ideal for ICG kinetics monitoring
purpose. Comparing the blood flow information extracted from ICG kinetics and the blood flow
information from the Diffuse correlation spectroscopy may be interesting.
151
Chapter 5
Dynamic Diffuse Optical Spectroscopy
on Fetal Brain in utero
5.1 Introduction
The ultimate goal of fetal brain project is to develop an accurate non-invasive trans-abdominal
monitoring device for the fetal brain oxygenation state, while the fetus is still in the uterus. The
clinical motivation and brief history regarding the fetal oximetry project is described in the Chapter
1, Section 1.3. In the following sections, the feasibility of in utero fetal brain oxygenation moni-
toring is demonstrated using a pregnant ewe model with induced fetal hypoxia. Then the prospect
of translating this technology to clinical setting is discussed in the following section based on the
preliminary clinical experience.
152
5.2 Trans-abdominal Near Infrared Oximetry of Hypoxic Stress in
Fetal Sheep Brain in utero
5.2.1 Introduction
The goal of this study was to demonstrate the feasibility of trans-abdominal NIR spectroscopy
for detecting and quantifying fetal hypoxia in utero in a pregnant ewe model (n = 5). We have
built a multi-wavelength NIR frequency-domain instrument with the capability to perform NIR
photon diffusion measurements through tissue over a wide range of source-detector separations.
We also developed a two-layer numerical diffusion model (for the maternal and fetal layers) to
quantify fetal cerebral blood saturation in utero. Good agreement was found between fetal blood
saturation determined by the trans-abdominal NIR method, and arterial and venous fetal blood
saturation quantified from fetal blood samples using a hemoximeter (gold standard). We conclude
trans-abdominal NIR oximetry has the capability to quantify different degrees of hypoxia in the
fetal brain in utero.
5.2.2 Materials and Methods
5.2.2.1 Animal Protocol
Five pregnant ewes (132-144 days gestation) were evaluated in this study. The animals were han-
dled according to the National Institutes of Health guidelines of the Institutional Animal Care and
Use Committee. The protocol consisted of the following steps: (1) anesthesia and catheteriza-
tion of the pregnant ewe, (2) catheterization of the fetus, (3) catheterization of the ewe for aortic
occlusion, and (4) NIR photon diffusion measurements and fetal blood sampling during baseline,
hypoxia and recovery.
153
Transabdominal probe
amniotic sac
maternalskin
detector fiber
source fiber
d : distance between probe and brain
aortic occlusion through catheterwith balloon
probe
Figure 5.1: Fetal hypoxia ewe model
First, the pregnant ewe was anesthetized with halothane [187]. The carotid artery was catheter-
ized for maternal arterial blood sampling and blood pressure monitoring. Second, the uterus was
exposed by a mid-line abdominal incision, and a small hysterotomy was performed to expose the
fetus for catheterization. The left brachial artery was catheterized for fetal arterial blood sampling
and the right brachial artery was catheterized for fetal blood pressure monitoring. The jugular vein
was catheterized for fetal venous blood sampling. The fetal body was then placed back in the
uterus. The uterus was tied around the fetal neck (purse-string method) to expose the fetal head
for the trans-abdominal NIR measurements. The exposed head was placed directly underneath the
maternal skin and secured by suturing its ears to skin. The purse-string approach was employed
to minimize the effect of the uterus in this pilot investigation. Third, a catheter with an inflatable
balloon was inserted through the femoral artery of the pregnant ewe for aortic occlusion. Aortic
occlusion through the femoral artery of the pregnant ewe is expected to directly reduce uterine
154
blood flow to the fetus, but have a minimal effect on maternal oxygenation. Finally, the probe was
placed on the maternal abdomen directly above the fetal head as shown in Figure 5.1. The details
of the instrument are described in Chapter 3, Section 3.6.3. The NIR measurements commenced
and were performed continuously during the entire baseline-hypoxia-recovery cycle. Fetal arterial
and venous blood samples were drawn during the baseline NIR measurement. Then the balloon
was inflated to block uterine blood flow and induce fetal hypoxia. The balloon was inflated until
the fetal blood pressure dropped rapidly. The inflation was maintained for 209± 38 seconds. Then
the balloon was deflated and the fetus was allowed to recover. Blood samples were drawn from
fetus every 30 seconds during hypoxia and once after recovery. Maternal arterial blood samples
were sampled and checked periodically to ensure the maternal arterial saturation was not perturbed
by aortic occlusion.
The thickness of the maternal layer was measured with a caliper to be 4.0 ± 0.4 mm. The
thickness of the fetal skull was obtained postmortem and it was 5.0 ± 0.5 mm. Maternal arte-
Table 5.1: The values of fixed parameters in data analysis. A was determined by b and µ′786nms .
Since CHb and CHbO2of fetal brain are the major parameters of interest and make the largest
contribution to the signal variation, we fixed all other parameters. Table 5.1 shows the fixed pa-
rameters for each layer of the two-layer model. In the top layer, d was fixed based on thickness
measurements of the maternal skin, fetal skin and fetal skull postmortem. The background µbga (λ),
the scattering properties A and b, the baseline THC of both layers, and the baseline StO2 of the top
layer were assumed based on values reported in the literature [179, 254, 286]. The StO2 values of
the bottom layer were obtained from the hemoximeter measurements during the normoxic baseline
157
in each cycle. Specifically, the baseline StO2 value for the fetus (bottom layer) was determined
using a compartmental model [63] where StO2 is made up of 43% SaO2 and 57% SvO2. Since
the aortic occlusion protocol is expected to perturb fetal hemodynamics only, the top layer THC
and StO2 were assumed to be the same as the baseline throughout the cycle. Once all the fixed
parameters were defined for the calculation of Φc, the MATLAB function, fminsearch, utilizing
the Nelder-Mead Simplex method (iterative method) was used for χ2 minimization in order to ex-
tract fetal brain CHb and CHbO2. In order to test the effect of the fixed parameters on the error
propagation in the two-layer diffusion algorithm, each parameter was varied within a range (≈ ±
25%) reported in the literature. An error analysis indicated that the two-layer diffusion algorithm
was relatively insensitive to variations in the majority of fixed parameters. Variation in these fixed
parameters resulted in at most 2% variation in the fetal blood saturation values. The assumptions
relating the arterial and venous contribution to the fetal baseline StO2 were limiting parameters
in our calculations. Varying the arterial/venous contribution from 43/57% [63] to 30/70% [124]
resulted in 0.5-5% variation in calculated fetal StO2, depending on the level of hypoxia. This
variation was the basis for the fetal StO2 error bars.
5.2.3 Results
Normalized measurements of amplitude (Am/Abm) and phase-shift (θm − θbm) vs. time at a source-
detector separation of 4.0 cm during a normoxia-hypoxia-recovery cycle is shown in Figure 5.2
at wavelengths 690, 786 and 830 nm. The amplitude and phase were normalized to the baseline
amplitude and phase. The effects of hypoxia are particularly evident in the amplitude vs. time
trace. The decrease in amplitude at 675 nm and increase at 830 nm indicates a relative increase in
the deoxy-hemoglobin concentration and a relative decrease in the oxy-hemoglobin concentration,
158
0 200 400 600 800 1000 1200 14000.4
0.6
0.8
1
1.2
1.4
Time (sec)
Am
/ A
mb
(a)
600 800 1000 1200 1400 1600 1800 2000−0.04
−0.02
0
0.02
0.04
0.06(b)
Time (sec)
θ m −
θmb
(ra
d) 675 nm786 nm830 nm
Inflation Deflation
Figure 5.2: Normalized measurements of amplitude (Am/Abm) and the phase-shift (θm − θbm) vs.
time measured at a source-detector separation of 4.0 cm during a normoxia-hypoxia-recovery cycleat wavelengths of 675, 786, and 830nm. The amplitude and phase were normalized to the baselineamplitude and phase, which were averaged from the first 100 data points in the time trace.
which is the expected trend for hypoxia. The amplitude at 786 nm, which is close to the isos-
bestic point of hemoglobin, exhibits an intermediate response. Time traces at other source-detector
separations showed similar trends, but with different magnitudes. No obvious physiological inter-
pretation could be drawn from the phase data.
In Figure 5.3 two examples of the fetal blood saturation obtained from NIR trans-abdominal
spectroscopy and from fetal blood samples quantified with the hemoximeter are presented. There
is a decrease in blood saturation with inflation and an increase in blood saturation with deflation of
the balloon. The blood saturation determined from the two-layer model fit shows good agreement
with the hemoximeter results. Herein the two-layer diffusion model fits of blood saturation will be
denoted as fetal tissue blood saturation (StO2).
159
0 200 400 600 800 1000 1200 14000
20
40
60
80
100
time (sec)
Blo
od
sat
ura
tio
n (
%)
Two−layer fitArterialVenous
Inflation Deflation
(a)
100 200 300 400 5000
20
40
60
80
100
time (sec)
Blo
od
sat
ura
tio
n (
%)
(b)
Inflation Deflation
Figure 5.3: Two examples of the fetal blood saturation obtained from NIR trans-abdominal spec-troscopy and from fetal blood samples quantified with the hemoximeter. ‘Two-layer fit’ is theblood saturation quantified by NIRS using two-layer model fit, ‘Arterial’ is arterial fetal blood and‘Venous’ is venous fetal blood measured by the hemoximeter.
The correlation between fetal blood saturation measured by trans-abdominal NIR spectroscopy
(NIRS) and hemoximeter was also examined. NIR blood saturations StO2 were selected from
points in the time traces where fetal arterial and blood samples were withdrawn. Blood saturation
values were calculated using the previously described compartmental model [48, 63, 106] from
the fetal arterial and venous blood saturations (measured from the fetal blood samples using the
hemoximeter). These values served as the gold standard.
A linear relationship between the NIR and hemoximeter fetal blood saturation over a wide
range of blood saturation values is observed in Figure 5.4, with a correlation coefficient R equal to
0.76 (p < 0.01).
The blood saturation data were obtained from normoxic, hypoxic and intermediate data points
(n = 47), collected from seven hypoxic cycles (n = 7). Notice that the variance in blood saturation
increases as the fetus progresses from a normoxic to a hypoxic state. Furthermore, there is a
poorer correlation between the NIRS and hemoximeter measured blood saturation in the lower
160
0 20 40 60 80 1000
20
40
60
80
100
Blood saturation by Hemoximeter(%)
Blo
od
sat
ura
tio
n b
y N
IRS
(%)
Figure 5.4: Linear relationship between the fetal blood saturation measured by NIR instrument andhemoximeter over a wide range of blood saturation values with a correlation coefficient R equal to0.76 (p < 0.01).
range of blood saturations. This may be because the blood saturation of the fetus in this range was
transient, so it was difficult to perfectly synchronize the NIR measurements and the blood sample
withdrawals.
The baseline state (before inflation of the balloon) and stable hypoxic state (while the balloon is
in the fully inflated state) was considered for further assessment of comparison. The difference be-
tween the baseline state and stable hypoxic state for the NIRS (∆StO2) and hemoximeter (∆ShO2)
blood saturations are compared in Figure 5.5. The data is shown for two groups of hypoxia. Mod-
erate hypoxic cycles (n = 3) are grouped with ∆StO2 = 30 ± 7% and severe hypoxic cycles (n =
4) are grouped with ∆StO2 = 60 ± 5%. A paired t-test analysis shows ∆StO2 and ∆ShO2 are
statistically similar within the moderate and the severe group (p < 0.05 for null hypothesis). An
unpaired t-test shows that the difference between moderate and severe ∆ShO2 is significant (p <
161
0
10
20
30
40
50
60
70
SO
2(bas
elin
e) −
SO
2(hyp
oxi
a) (
%) ∆ShO2 (Hemoximeter)
∆StO2 (NIRS)
Moderate Hypoxic Group (N=3)
Severe Hypoxic Group (N=4)
Figure 5.5: Difference in blood saturation between the baseline state and stable hypoxic state forthe NIRS (∆StO2) and hemoximeter (∆ShO2) measured blood saturations. The data is shown fortwo groups of hypoxia.
0.05). This is true for ∆StO2 as well.
5.2.4 Discussion
This study demonstrates for the first time, the feasibility of using trans-abdominal NIR spec-
troscopy for detecting and quantifying fetal hypoxia in utero. The pregnant ewe model is a
widely used model for studying fetal physiology and was ideally suited for this proof-of-principle
study. Perturbations of fetal blood saturation were performed in a controlled manner and fetal
arterial and venous blood saturations determined from fetal blood samples served as a reliable
gold standard to which the NIR measured blood saturations could be compared. The multi-
wavelength, multi-separation NIR frequency domain instrument coupled with the numerical two-
layer diffusion model proved capable of retrieving fetal blood saturation in utero, accurately and
162
non-invasively. The frequency-domain technique is more effective than the previous CW tech-
niques [227, 228, 270]; the former is better able to decouple absorption from scattering, which
is important for quantifying oxy- and deoxy-hemoglobin concentrations in tissue. The two-layer
numerical diffusion model also represents a significant improvement over the widely used homo-
geneous model; the former model was clearly able to deconvolve fetal from maternal blood satu-
rations, rather than volume-averaging them. When the homogeneous model was used on this fetal
data under the same condition, it underestimated the change in blood saturation. Multi-wavelength
NIR photon diffusion measurements enabled the use of a priori spectral information to reduce
number of unknowns (absorbers) in the two-layer diffusion model. Finally, NIR reflectance mea-
surements at multiple source-detector separations probe different tissue depths, thus optimizing the
deconvolution of fetal from maternal signals using the two-layer diffusion model.
In this study, several modifications were made to the pregnant ewe model to simplify the fetal
hypoxia protocol. First, hypoxia in the fetus was indirectly induced through aortic occlusion of
the maternal femoral artery. Alternative approaches for inducing fetal hypoxia are to lower the
maternal blood saturation by lowering the fraction of inspired oxygen (FiO2) or through umbilical
cord occlusion. These alternative approaches were evaluated in preliminary studies. The problem
with lowering maternal FiO2 was that both maternal and fetal blood saturations were affected
and therefore the perturbation was not unique to the fetus. With the umbilical cord occlusion
approach, fetal morbidity and mortality were significant. Another modification made in this study
was the uterine layer was removed from the field during trans-abdominal NIR spectroscopy of the
fetus. The protocol was simplified in this manner to first establish the accuracy of quantifying fetal
cerebral blood saturation, for the case in which the overlying layers are not affected by the hypoxic
perturbation. The success of this pilot investigation sets the precedent for future animal model
163
studies. In future studies, the complexity of the additional uterine layer and different perturbation
approaches will be investigated. Additionally, a larger number of animals will be studied to obtain
more statistically significant results.
5.3 Outlook towards translation to human
We now discuss the future outlook of the fetal brain project with focus on the traslation to human
case based on the experience gained from the preliminary C-section clinical data.
5.3.1 Preliminary Clinical Data
Preliminary clinical data were collected using a simple dual wavelength, dual light source frequency-
domain spectroscopy with one detector channel (Figure 5.6). Two 750 nm laser diodes (Sharp,
LT031MD and LT030MD) and two 780 nm laser diodes (Sharp, LT024MD) were amplitude mod-
ulated by a 70 MHz local oscillator (Wilmanco, VSA-70+13dBm). LT031MD was used for larger
source detector separation because of its higher power. The laser diodes were time-shared at a
frequency of 1.25 Hz. Two pairs of a 750 nm and a 780 nm laser diodes were coupled to an optical
fiber (diameter = 1mm, N.A. = 0.37) by a laser diode power combiner (Oz optics) respectively.
The laser outputs were balanced to give similar level by adjusting the attenuator embedded in the
power combiner. The laser outputs after an optical fiber were about 1 mW. The optical fibers
were mounted on a probe at two different distance from a detector. A photomultiplier tube (PMT,
Hamamatsu, H5783-01) was used for the detection. The detected signal is amplified by a series
of amplifiers (Mini-circuits, ZFL-2000, MAN-1LN) and a filter (Mini-circuits, PLP-90). A homo-
dyne scheme based on a I&Q demodulator (Mini-circuits, MIQY-70D) with lowpass filters was
used to detect the amplitude and phase of the diffuse photon density wave (DPDW) [288].
The coupling coefficients from detectors and light sources were corrected by reference mea-
surements on a solid homogeneous tissue phantom (made of lesin) with known optical properties
accompanied for each clinical measurement.
A clinical protocol was designed to include post-partum measurements on a newborn neonate
to correlate with the pre-partum transabdominal measurements on a pregnant woman (Figure 5.7).
An elective Cesarean section was chosen particularly for this purpose since the time between the
transabdominal measurements and delivery is more controllable than in a labor. When available,
the cord gas was analyzed for oxygen saturation. The protocol was approved by the Internal Review
Board at the University of Pennsylvania.
When a patient consented to participate in the study, the location of fetal head was found by
hand-held ultrasound tranducer which was in regular use in the hospital. A pre-partum measure-
ment was done on the patient’s abdomen right after anesthesia was administered to her and before
the Cesarean section operation began. When the neonate was delievered, a post-partum measure-
ment on neonate’s forehead was performed. For rare occasions, we had access to the fetal forehead
165
right before the umbilical cord was cut by using sterilized field over the probe.
In the pre-partum measurement, a soft black rubber probe with two optical fibers for light
sources placed at 3 cm and 7 cm from a PMT respectively was used. TO-8 PMT was mounted
directly on the probe and was operated at high voltage of 1000 V to maximize the signal to noise
ratio at separation of 7 cm. The probe was placed directly over the fetal head after it was located
by ultrasound transducer. The ultrasound also provided information on the distance between the
maternal abdomen and fetal skull. In the post-partum measurement, a soft black rubber probe with
one optical fiber placed at 3 cm from an optical bundle coupled to a PMT was used on a neonate
forehead.
(a) (b)
(c) (d)
Figure 5.7: Illustration of clinical in utero measurements. (a) Transabdominal measurement on thematernal abdomen, (b) Inside view of transabdominal measurement, (c) Direct fetal brain measure-ments before detechment of the umbilical cord during Cesarean section, (d) Direct neonate brainmeasurements after birth
It was feasible to use the instrument to follow this protocol in the Cesarean section environment
166
without causing much extra delay for the operation. However, the improvement in the probe design
would be much anticipated for the future study. Especially PMT was difficult to use in the operation
environment due to its sensitivity to ambient stray light. Fast shutter mechanism to protect PMT
from being exposed to the ambient light directly should be implemented as well as the general
isolation of stray light.
The optical properties and blood oxygenation obtained from single-spectral semi-infinite ho-
mogeneous medium fit (as described in Section 2.3) and Beer-Lambert law are summarized in the
following. (A) is from the transabdominal measurements at source detector separation of 3 cm and
(B) is from that of 7 cm. (C) is from the neonatal forehead measurement after the umbilical cord
was detached. Typically, measurement (C) was carried out 17 ± 6 minutes after the birth.
Table 5.2: Summary of µa from clinical in utero data
N , number of subjects measured is different in each measurements. Dual source configuration
measurements were not available from the beginning of the study ((A) and (B)). For some cases
where the neonate’s state needs care, we did not have access to the neonate measurements in (C).
Since the fetal head has higher absorption coefficient as shown in (C), µa at separation 7 cm (B) is
expected to be higher than that of at 3 cm (A).
The averaged µas agree with those found in the literature. Fishkin et al [99] reported the normal
abdomen µa = 0.0626 ± 0.003 cm−1 and µ′s = 9.11 ± 0.15 cm−1 for source-detector separation
of 2.7 cm at 811 nm which is congruent with values found in (a). Gratton et al [112] reports human
forehead µa = 0.16 cm−1 and µ′s = 7.3 cm−1 at 715 nm. Matcher et al [179] reports human
167
forehead µa = 0.16 ± 0.01 cm−1 and µ′s = 9.4 ± 0.7 cm−1 at 800 nm for 14 adult subjects. The
averaged absoprtion coefficients at separation 3 and 7 cm (Table 5.2 (A) and (B)) do not exhibit
the difference. For individual patients, when the fetal head depth was less than 2.3 cm, µa at 7 cm
was greater than µa at 3 cm as expected. However, µa contrast between the separations were quite
small. µ′s at 7 cm is consistantly lower than µ′s at 3 cm. Whether this is truly physiological result
or this is due to µa and µ′s crosstalk or coupling coefficient offset, it is rather ambiguous at this
point since there were only marginal number of measurements to satisfy the number of unknowns
and the coupling coefficients were assumed to be the same in the solid phantom and the tissue.
The effects of coupling coefficient have recently gained considerable interest. The magnitude of
coupling coefficient would be such that not to affect homogeneous fit too much but would be large
enough to skew the sensitivity to embedded objects or layers [25].
Neonate blood saturation is lower than expected. This has been observed in the animal stud-
ies on piglets and lambs [51] and also observed by Hueber et al [133]. They attributed this to
wavelength dependent effective background absorption coefficient whose origin could be model
mismatch.
For three cases, measurements on the fetal forehead right before the umbilical cord was de-
tached were available. In these cases, the blood saturation from the fetus correlated linearly in
positive direction with maternal blood saturation whereas the neonate saturation (C) was not. How-
ever, fetal blood saturation was much lower than maternal blood saturation in this case. Also, for
the case blood saturation was monitored within a few minutes after umbilical cord detachment,
blood saturation of neonate increased with time which was also monitored with pulse oximeter on
neonate’s finger. Same phenomena has been reported by Isobe et al [136]. From these observa-
tions, one can draw a hypothesis that the fetal blood saturation may be lower and increases after the
168
umbilical cord is detached. This suggests the difficulties in the original study design in correlating
the neonate blood saturation to transabdominal blood saturation unless the neonate is accessible
for measurements within minutes after birth.
Figure 5.8 summarizes above observation by showing a bar graph of µa and µ′s at 750 and
780 nm, THC and StO2 of transabdominal measurement (N=15) at separation 7 cm, fetal head
measurement (N=3), and neonatal head measurement (N=10). Transabdominal THC is much lower
than that of fetal head and neonate head. µ′s does not seem to be significantly different from each
group. However, StO2 of fetal head is extremely low.
Low neonate blood saturation indicates the need of incorporation of more wavelength to the
instrument. Lack of contrast in absorption coefficients at short and long separations indicates either
the measurements at source detector separation of 7 cm is not sensitive enough to the fetal signal or
the over-simplified homogeneous model of in-utero system is underestimating the effect of fetus.
Even though the comparison between averaged µa at source detector separation 3 cm and 7
cm did not yield much difference, this can be easily attributed to model mismatch (i.e. not using
two-layer model). The clinical data still reveal the encouraging trend in Figure 5.9, where the µa
at source detector separation of 7 cm decreases with increase of the fetal brain depth measured by
the ultrasound. Since the absorption of fetal brain is higher than that of maternal layer, as the fetal
brain depth is smaller, the expected absorption from homogeneous fit would be higher due to more
signal contribution from fetal brain.
5.3.2 Clinical translation outlook
In order to translate the technology developed in Section 5.2 to clinical settings, further improve-
ments must be considered. Implementation of the two-layer diffusion model required information
169
0
0.05
0.1
0.15
0.2
0.25
µ a (cm
−1 )
750nm786nm
Transabdominal Fetal head Neonate head 0
2
4
6
8
10
12
µ s′ (cm
−1 )
750nm786nm
Transabdominal Fetal head Neonate head
0
0.05
0.1
0.15
0.2
0.25
0.3
Tot
al h
emog
lobi
n co
ncen
trat
ion
(mM
)
Transabdominal Fetal head Neonate head 0
20
40
60
80
100
Blo
od o
xyge
n sa
tura
tion
(%)
Transabdominal Fetal head Neonate head
Figure 5.8: Summary of physiological parameters (µa, µ′s, THC, and StO2) at prior (transabdomi-nal, N=15), during (fetal head, N=3) and after birth (neonate head, N=10).
about several variables: top layer thickness which served as a priori spatial information, and fetal
baseline blood saturation from the hemoximeter measurements. Other parameters were assumed
according to the literature values, and their variation resulted in minimal influence on the fetal
blood saturation calculation. In clinical studies, the top layer thickness can be measured by ul-
trasound. Other available parameters such as maternal arterial blood saturation can be utilized to
further constrain the variables. However, the baseline fetal blood saturation is not available in the
clinical environment. If the baseline fetal blood saturation had been overestimated, the algorithm
would have underestimated the severity of hypoxia in the fetus and vice versa. Absolute quan-
tification of baseline fetal blood saturation is possible with measurements at more source-detector
170
1.5 2 2.5 3 3.5 40
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
fetal head depth (cm)µ a (
cm−1
) at
750
nm
Figure 5.9: µa measured at source detector separation of 7 cm using transabdominal DOS vs fetalbrain depth measured by ultrasound
separations and more wavelengths [58] (with extensive calibration [25]). The other major difficulty
arises because the distance from the maternal abdomen to the fetal brain is 2-4 cm [227] in humans.
For the clinical situation, the two-layer approximation is especially crucial since the optical
properties of maternal layer and fetal head are distinctively different as well as the sheer dimension.
To explore the degree of influence in human case, two-layer system consisting of maternal layer
(mostly fat) and fetal head as shown in Figure 5.10 is simulated with optical properties assigned
from clinical study.
Figure 5.10: Approximation of in utero system to two-layer system
171
Most of the signal gets diluted due to the presence of meternal fat layer, and it is a function
of the layer thickness. This is illustrated in Figure 5.11, where the two-layer data from the optical
properties similar to clinical in utero system are generated using finite difference method. Then
two-layer data at source detector separation of 3 and 7 cm were fitted to the semi-infinite homoge-
neous solution (Equation 2.7). Also, fit using two-layer solution is presented. The homogeneous
fit underestimates the optical properties, resulting in somewhat in-between values between the fe-
tal head and maternal layer properties. The degree of degradation is more pronounced for the
shorter source detector separation as expected. Given the underestimation of optical properties by
homogeneous solution, it is clear that one needs to fully utilize the two-layer approximation.
2 2.5 3 3.5 4 4.50.06
0.08
0.1
0.12
0.14
0.16
0.18
Thickness of the top layer (cm)
µ a (cm
−1 ) two−layer fit
semi−infinite fit at 7cmsemi−infinite fit at 3cmexpected value
(a)
2 2.5 3 3.5 4 4.56
7
8
9
10
11
12
Thickness of the top layer (cm)
µ s′ (cm
−1 )
(b)
Figure 5.11: Effect of homogeneous fit on two-layer system. (a) Fitted absorption coefficient usinghomogeneous solution and two-layer solution from the simulated data based on a two-layer model,(b) Fitted reduced scattering coefficient.
This requirement of two-layer in turn needs optimization of the instrumentation towards mea-
suring at larger source detector separations than used in the animal study. The method to determine
172
the optimal source detector separation and the choice of detectors is illustrated in Appendix (Sec-
tion 5.4).
Even with the optimized instrument, the validation of this technique in the clinical setting is
still problematic. Additional animal study may be necessary before taking to the clinic. Especially
the animal model should be designed to reflect the similar thickness of maternal layer as the hu-
man case. Also, the effect of uterine layer should be assessed as well as exploring different type
of oxygen perturbation to the fetus. In the clinical C-section setting, correlating neonatal blood
oxygen saturation with umbilical cord is not effective since the neonate’s oxygen level rises rather
rapidly after birth. The best correlation would be through measurement on fetal head right be-
fore the umbilical cord is detached. For this measurement, better probe design which can readily
accommodate sterile requirement of operation is needed. The access to the hypoxic fetus would
be necessary in the end. Ideal setting for validation of this technique may lie in the fetal surgery
environment where the fetus is taken out of womb for surgical intervention and put back.
5.4 APPENDIX: Instrument optimization for Human Case
This appendix section is included to illustrate the methodology of estimating the optimal source
detector separation by forward-model based SNR calculation.
5.4.1 Two-layer forward model
The two-layer model used is a three dimensional rectangular block (21 cm × 14 cm × 7 cm)
consisting of two different layers of different optical properties: a top layer simulating the maternal
layer and a bottom layer simulating the fetal head. Typically µa1 = 0.06 cm−1 and µ′s1 = 8.0 cm−1
for the top layer and µa2 = 0.15 cm−1 and µ′s2 = 6.0 cm −1 for the bottom layer was used. The
173
0 2 4 6 8 10 12 140
20
40
60
80
100
120
source detector separation (cm)O
vera
ll si
gnal
/ no
ise
2.0 cm2.5 cm3.0 cm
Figure 5.12: Noise model is incorporated into the two-layer signal to give overall signal to noise.This reflects the signal to noise decrease with increase of source detector separations. Layer thick-ness of 2 cm (solid line), 2.5 cm (dashed line), 3 cm (dotted line) are shown.
µa1 and µ′s1 were from the average clinical data from measurements at 3 cm separation (Table 5.2
(B),(C)). The boundaries were set as absorbing (n = 1.0) in all sides except one side (maternal
abdomen) with air to tissue index mismatched boundary (n = 1.4). After the optical properties of
each layer and the thickness of the top layer was assigned for each voxel, finite difference method
was used to solve the diffusion equation for this geometry [129].
5.4.2 Incorporation of noise to two-layer model
A noise model described in Chapter 3, Section 3.2.2.6 is utilized to generate the noise which
depends on the signal level. In the following the case where Psource = 20 mW, floss = 0.05, Adet
= 0.1 cm2 was considered to convert the fluence from forward model shown in Figure 5.12. The
overall signal to noise starts decreasing after source detector separation of 8 cm in both cases. By
increasing the factors Psource and ∆A, and decreasing floss shift the curve in Figure 5.12 to right
side of horizontal axis.
174
0 2 4 6 8 10 12 140
20
40
60
80
100
source detector separation (cm)
Rel
ativ
e am
plitu
de (
%)
2.0 cm2.5 cm3.0 cm
(a)
0 2 4 6 8 10 12 140
5
10
15
20
25
30
source detector separation (cm)
Pha
se d
iffer
ence
(de
gree
) 2.0 cm2.5 cm3.0 cm
(b)
Figure 5.13: Fractional (a) amplitude difference in percentage and (b) phase difference in degreeare obtained between the amplitude from two-layer system and amplitude from homogneous sys-tem with top layer optical properties at different source detector separations. It is a measure ofsensitivity to the bottom layer optical properites. Layer thickness of 2 cm (solid line), 2.5 cm(dashed line), 3 cm (dotted line) are shown.
5.4.3 Sensitivity to fetal signal
The sensitivity to the signal from the bottom layer is obtained by treating the contribution of fetal
signal on the overall signal is a perturbation caused due to its presence as contrasted to the case
where the whole region is homogeneous with optical properties of maternal layer. In the Figure
5.13, the fractional amplitude difference between two-layer case at top layer thickness of 2 - 4
cm and the homogeneous case of µa1 = 0.06 cm−1 and µ′s1 = 8.0 cm−1 are plotted against the
source detector separations. It shows that the fractional differences increases with increase of
separations. The fractional differences decrease with increase of thickness. As expected before,
the sensitivity to fetal signal increases with source detector separations and decrease with maternal
layer thickness.
175
0 2 4 6 8 10 12 140
100
200
300
400
500
source detector separation (cm)D
iffer
entia
l sig
nal /
noi
se
2.0 cm2.5 cm3.0 cm
Figure 5.14: Differential signal to noise corresponding to fetal signal is plotted. Optimal sourcedetector separation exist at the maximum differential signal to noise. Layer thickness of 2 cm (solidline), 2.5 cm (dashed line), 3 cm (dotted line) are shown.
5.4.4 Optimal Source Detector Separation
Decrease of overall signal (Figure 5.12) and increase of sensitivity (Figure 5.13) with increase of
source detector separation competes with each other in detection of fetal signal. Multiplying the
overall signal and sensitivity gives the differential signal over noise which is the SNR from fetus.
In Figure 5.14, this effective SNR from fetus peaks at certain optimal source detector separation
where the effects of overall signal and sensitivity are balanced. The optimal separations increase
with increase of layer thickness. The optimal separations range around 8 - 10 cm. However,
the overall differential signal to noise ratio decreases with increase of layer thickness. Therefore
for thickness of 4 cm, even if the optimal separation is selected the measurement will give poor
information on fetus since the effective SNR is so low.
5.4.5 Detectability of fetal signal in two-layer model: Inverse Problem
Ultimately, the detectability of fetal signal relies on the ability of the model based inversion algo-
rithm to extract fetal information accurately. Therefore, two-layer diffusion model based inversion
176
1 1.5 2 2.5 3 3.5 40.04
0.06
0.08
0.1
0.12
0.14
0.16
Thickness of the top layer (cm)
µ a (cm
−1 )
(a)
1 1.5 2 2.5 3 3.5 44
5
6
7
8
9
10
Thickness of the top layer (cm)
µ s′ (cm
−1 )
homogeneous fit at 3cmhomogeneous fit at 7cmtwo layer fit of top layertwo layer fit of bottom layer
(b)
Figure 5.15: Comparison between the homogeneous model fit and the two-layer model fit. In theforward data, amplitude noise based on (a) and random phase noise of 0.1o were added. Then it isfitted to semi-infinite homogeneous analytic solution at source detector separation of 3 cm (cross)and 7 cm (filled triangle). The optical properties of top layer (open circle) and bottom layer (filledcircle) are give by fitting to a two-layer numerical solution.
algorithm was developed using the finite difference method based diffusion equation solver as for-
ward model. In the clinical situation, we can take advantage of the fact that the thickness of the
maternal layer can be measured with the ultrasound technique. In order to retrieve the optical prop-
erties (i.e. µa1, µ′s1, µa2 and µ′s2), Nelder-Mead Simplex algorithm was used to update the optical
properties of each layers until χ2 = Σ(ΦmΦ0m− Φc
Φ0c) where Φm is measured sample fluence, Φ0m is
measured baseline fluence, Φc is calculated sample fluence, Φ0c is calculated baseline fluence, is
minimized.
The detectability of fetal signal, that is the ability of the inverse algorithm to retrieve accu-
rate fetal information in the presence of realistic noise factors is demonstrated for noise-model
incorporated simulated data and compared with fitting result from homogeneous model.
Random amplitude noise based on the realistic noise model was added on the forward data
177
generated for two-layer and homogeneous case using finite difference method for two-layer mea-
surement and baseline measurement. This simulates the experimental situation where the signal
deteriorates with increase of source detector separation. However, constant random noise of 0.1o
was added to the phase. Differential signal between two-layer and homogeneous baseline was then
used as inputs for the inverse procedures. Crosses and filled triangles in Figure 5.15 are given by
fitting to homogeneous analytic solution for semi-infinite geometry at 3 cm and 7 cm respectively.
Open circles and filled circles are top layer and bottom layer optical properties obtained by fitting
to two-layer model based numerical solution at two source detector separations. Typically, the
short separation is fixed at 3 cm and longer source separation is determined by optimal source de-
tector separations found in Figure 5.14. The semi-infinite fit is stable with respect to the amount of
noise added. However, the fitted values approaches to the values of top layer as thickness of layer
increases. Even at thickness of 1 cm, optical properties do not reach those of bottom layer. Also,
the contrast between 3 and 7 cm result is inherently small which is consistent with the findings
in the clinical data analysis. The two-layer model fit, on the other hand, is sensitive to the level
of noise in the data. In this case, if the layer thickness exceeds 3 cm, the fit becomes unreliable.
However, the overall value reflects much better on the values of bottom layer than homogeneous
fit. The effective signal to noise should be at least higher than 10 to give accuracy up to 10 % in
absorption coefficient estimation of bottom layer. Estimation of top layer optical properties were
always much better (less than 5 %) and it seems to be less affected by the layer thickness.
5.4.6 Instrument Requirements
The methodology to predict the performance of the instrumentation for given system specification
described above can be applied to any sets of specification. It is shown for the given example that it
178
is feasible to detect fetal signal for thickness of 2 - 3 cm for NEP of system around 10−13W/√Hz
and thickness of 2 - 4 cm for NEP of system around 10−16W/√Hz at source detector separations
ranging 8 - 11 cm for modulated light source 20 mW.
So far, only a special case of two-layer was considered. However, the same methodology can
be applied for various optical properties and thickness variation. The optical properties and layer
thickness variation gives a range of signal. From a series of simulation of varying parameters, top
layer optical properties are found to be the main factors. Variation of µa1 = 0.02 - 0.12 cm−1 at
µ′s1= 8 cm−1 causes variation of amplitude signal in order of 107 across the optode separation of
3 - 10 cm. However, if the optode separation is chosen, the variation range is about 103 which can
be easily achieved experimentally. Variation of µ′s1 = 5 - 12 cm−1 at µa1 = 0.04 cm−1 is similar in
magnitude for fixed optode separation. The variation induced by the bottom layer optical properties
is at most order of 102 when the optical properties of top layer and thickness is fixed.
Therefore, when selecting the detector, NEP of the detector should be as small as possible,
detector area should be large and the dynamic range should be as large as possible. PMT is more
desirable to APD in terms of low NEP. However, dynamic range of PMT is smaller compared with
APD. When choosing the detectors, a detector with high gain and high anode sensitivity is favored
in terms of signal amplification. The chains of amplifiers can be selected such that the signal is
amplified not to be limited by A/D board resolution. Efforts to reduce the overall NEP should be
carried out for the detection system. Phase noise was not discussed in detail in this investigation.
However, phase noise contributes in raising the overall NEP. Heterodyne scheme may work better
than homodyne scheme in terms of phase noise. The comparison of two schemes are for future
investigation. Increasing modulated light source power increases the sensitivity to the bottom layer
detection. If the optode separations were fixed, modest dynamic range 103 − 104 per detection
179
position to give effective signal to noise much better than 10 is required to give reliable fit to two-
layer model. Coupling coefficients were not considered in much detail. However, their magnitude
may be large enough to obscure the sensitivity to fetal signal. Source detector combination should
implement scheme to fit for coupling coefficients reliably. For this at least 2 sources and 2 detectors
are needed. Increasing wavelength of light is expected to improve the fidelity in blood volume and
oxygenation evaluation.
180
Chapter 6
Summary
In this thesis, the motivation, theoretical background, experimental techniques and clinical results
on in vivo non-invasive breast cancer imaging and fetal brain oxygen monitoring using diffuse op-
tical tomography and spectroscopy were presented. For the breast cancer imaging application, the
quantification of tumor contrast based on total hemoglobin concentration and scattering through
three-dimensional DOT reconstruction is demonstrated as well as the monitoring capability fol-
lowing neoadjuvant chemotherapy. Also, a pilot study of blood flow measurement on breast cancer
cases is presented. The breast cancer imaging project is at an exploratory stage of finding the opti-
cal cancer contrast, constantly improving techniques for better quantification. The future direction
lies in the (1) improvement of the instrument focusing on multiple source detector positions, multi-
ple wavelength and speed, (2) incorporation of various techniques into reconstruction algorithm to
use a priori information and to reduce imaging artifacts, (3) active exploration of additional optical
parameters such as blood flow as well as establishing links with histopathology, (4) coregistration
with other imaging modality to complement each imaging technique, (5) threrapy monitoring for
181
treatment efficacy, and (6) active exploration of optical contrast agent usage which may make tran-
sition to cancer-specific targetting contrast agent.
For the fetal brain oxygen monitoring application, accurate quantification of fetal hypoxia using
two-layer model based diffuse optical spectroscopy was demonstrated using a pregnant ewe model
for aortic occlusion induced fetal hypoxia. The translation of this technique to the clinical setting
involves (1) construction of an optimized instrument for deep tissue detection, (2) a validation study
on the animal model with thickness of maternal layer similar to the human case, and (3) a carefully
designed clinical validation study (either in Cesarean-section or fetal surgical environment).
182
Glossary
—— SYMBOLS ——
ε Extinction coefficient.
λ Wavelength of Light usually in nano-meters (nm).