Chapter 9

STRATEGIES FOR ABSOLUTE CALIBRATION

OF NEAR INFRARED TOMOGRAPHIC TISSUE
IMAGING
Troy O. McBride, Brian W. Pogue, UlfL. Osterberg, & Keith D. Paulsen
Thayer School of Engineering. 8000 Cummings Hall. Dartmouth College. Hanover. New
Hampshire 03755

Abstract:

Quantitative near infrared (NIR) imaging of tissue requires the use of a diffusion
model-based reconstruction algorithm, which solves for the absorption and
scattering coefficients of a tissue volume by matching transmission
measurements of light to the predictive diffusion equation solution. Calibration
problems as well as other practical considerations arise for an imaging system
when using a model-based method for a real system. For example, systematic
noise in the data acquisition hardware and source/detector fibers must be
removed to prevent spurious results in the reconstructed image. Practical
considerations for a NIR diffuse tomographic imaging system include: (I)
calibration with a homogeneous phantom, (2) use of a homogeneous fitting
algorithm to arrive at an initial optical property estimate for image
reconstruction of a heterogeneous medium. and (3) correction for fluctuations in
source strength and initial phase offset during data acquisition. These practical
considerations, which rely on an accurate homogeneous fitting algorithm are
described. They have allowed demonstration of a prototype imaging system
that has the ability to quantitatively reconstruct heterogeneous images of
hemoglobin concentrations within a highly scattering medium with no a priori
information.

Key words: blood, calibration, hemoglobin, photon migration, reconstruction, tomography

Oxygen Transport to Tissue XXIV, edited by

Dunn and Swartz, Kluwer AcademiclPlenum Publishers, 2003

85

86

1.

McBride et al.

INTRODUCTION

Tissue is highly scattering at visible and near-infrared (NIR.)

wavelengths, thus a simplistic back-projection image is highly blurred and
difficult to interpret as the effects of scattering and absorption can not be
distinguished. However, by using model-based imaging and computational
methods in combination with frequency domain measurements, moderate
resolution, quantitatively accurate images of absorption and scattering can be
obtained within highly scattering media, such as tissue [1,2]. This method
may lead to a potentially powerful medical imaging modality for noninvasive hemoglobin tomography.
Practical calibration issues arise when using a model-based image
reconstruction approach which may not be present in modalities such as xray tomography due to the necessity of fitting the data to a partial differential
equation solution. In addition, while many system calibrations can be
ignored in qualitative imaging, more care is required for quantitative
imaging of hemoglobin concentrations on an absolute scale. Issues
encountered in the development of the NIR imaging system addressed in this
paper include: (1) elimination of system-based offsets, (2) choice of
heterogeneous starting value, and (3) offsets due to long-term drift. These
issues are addressed through a calibration protocol which involves a
homogeneous phantom and the use of a precise homogeneous fitting
algorithm that is resistant to measurement noise.
A 4: 1 increase in hemoglobin concentration [3] and a 1.4 - 4.4 times
lower oxygen pressure [4] have been observed in breast cancer. By using
NIR. frequency domain measurements which have been properly calibrated
at multiple wavelengths, quantitative absolute images of hemoglobin related
parameters can be obtained [5]. Once calibrated imaging is implemented in
practice it should be possible to determine whether NIR. hemoglobin
concentration and oxygen saturation information is useful in the diagnosis of
breast tumors.

2.

METHODS

2.1

Imaging system

The frequency-domain near-infrared imaging system consists of three

main components: (1) data acquisition system, (2) image reconstruction

87

Strategies For NIR Calibration

Data
Acquisition
At a single
wavelength .
acquire 256
measurements

'Inverse' Image
Reconstruction
Reconstruct into
quantitative image
of absotption and
scattering

Tissue Functional
Parameters
Combine (I east squares regression)
absotption images at multiple wavelengths
to generate images ofhemoglohin
concentration and oxygen saturation.

Repeat for all

wavelengths

Figure 1. Schematic of near-infrared tomographic imaging system

algorithm, and (3) spectroscopic determination of functional properties. A

schematic of the imaging process is shown in Figure 1.

2.1.1

Data acquisition system

The data acquisition system (shown in Figure 2) uses an amplitudemodulated (100 MHz) wavelength-tunable (700-850 nm) TiSapphire laser as
the light source. Sixteen source and sixteen detector optical fibers are
arranged in a circular geometry to analyze a single tomographic plane of the
measured tissue or tissue-simulating phantom. The detector for the system is
a photomultiplier tube with built-in heterodyning circuitry. The heterodyne
signal (1 kHz) amplitude and phase shift are read into a computer using a
commercial-grade data acquisition board. The source and detector are
multiplexed to acquire the 256 data points using linear translation stages.
[5,6]
(b)

(a)

....----., loo.oonrnz

hIOO.1tll MHz -

--I

Figure 2. (a) Diagram and (b) photograph of the data acquisition system.

88
2.1.2

McBride et al.
Image reconstruction

A finite element based solution to the frequency-dependent diffusion

equation is used to calculate a fit to the measured projection data of
amplitude and phase shift. Absorption and scattering coefficient images are
generated using a Newton iterative scheme [1,7] with update vectors
determined from a full matrix inversion during each iteration. A schematic
of the image reconstruction algorithm is illustrated in Figure 3.
Measured Phase and Amplitude
data, ~.aS!Jred (h.t.ro) , at 256
points

....

Solve forward diffusion equation using latest

estimate of absorption and scattering at each
finite element node to determine calculated
phase and amplitude at 256 points.

IOriginal estimate of
optical properties

....

.....

IReturn
New I
Estimate

Compare calculated and measured

phase and amplitude data at 256 points.
Error below
tolerance

Error above
tolerance

+r

FiniS"ied

Update vector based on full matrix

inversion generates new absorption
and scattering at each me&! node

Figure 3. Schematic of the finite element based reconstruction algorithm based on the nonlinear estimation problem of inferring optical property maps using the frequency-domain
diffusion equation.

2.1.3

Determination of functional properties

Multiple wavelength images of absorption coefficient can be used to

determine maps of hemoglobin related parameters by incorporating the
previously measured absorption spectra of pure oxygenated hemoglobin, deoxygenated hemoglobin [8], and other tissue chromophores such as lipids [9]
and water [10]. Because the image reconstruction algorithm recovers
quantitative values for the absorption coefficient, a least squares fit can be
used to determine the metabolic chromophore concentrations from that data
as shown schematically in Figure 4.

89

Strategies For NIR Calibration

Known mol ar absorption (a) at each
wavelength (A) from measured values for
the unknowns (e.g. HbR and Hb02)
Hb - R =
a 1l,1i2,~n

[mm~mMJ

Hb - 01 =
a1l,1i2,b

[mm.1mM]

Measured absorption (W at "n"

wavelengths at each point in the image from
finite element reconstruction algorithm

Ji",.tI,.t2,An

= [~m

Calculate concentrations (in. this case [Hb

02] and [HbR]) of unknowns for each
point using least squares fit.

1i.,AI = [Hb - 02l a':-02 +[Hb - R] .a':-R

li a,A2 =[Hb-02la~-02 +[Hb-R].a~-R

1i.,hI = [Hb- 02laJ!b-02 +[Hb- R]a:!'-R

Figure 4. Schematic of methodology for determining hemoglobin related parameters from
multiple wavelength images of absorption coefficient.

2.2

Practical considerations

Calibration and other practical considerations arise when using a modelbased image reconstruction method with data acquired from a real imaging
system. In the NIR diffuse tomography context these include: (1) systembased offsets, (2) initial optical property estimate for a heterogeneous
medium, and (3) long term fluctuations of source strength and initial phase
offset.

2.2.1

System-based offsets

Systematic offset in the measured data occurs at various source and

detector locations due to optical fiber differences, multiplexing imprecision,

90

McBride et al.

and other systematic inconsistencies. Obviously, these offsets need to be

removed from the' data prior to image reconstruction in order to prevent the
appearance of artifacts in the resultant tissue property profiles. The method
adopted here measures a homogeneous phantom and subtracts differences
between measured and calculated values:
:alibrated(heterO)

=~eaSUred(heterO) - (~easured(hOmO) -

:alculated(hOmO)

(1)

where ~easured is the measured amplitude and phase at each i of the 256
measurement locations for the calibration phantom (homo) and the actual
heterogeneous object (hetero), while ;alculated is the corresponding
calculation for the homogeneous (homo) case. This requires either exact
knowledge of the scattering and absorption coefficients for the measured
homogeneous phantom or a homogeneous fitting algorithm which
determines these properties for the measured data on the homogeneous
calibration phantom. The homogeneous fitting algorithm (described in
Section 2.3) is an important component of the practical reconstruction
algorithm and is, therefore, also used in the calibration procedure.
The calibration procedure for eliminating systematic offset at different
source and detector locations involves: (1) measurement of a homogeneous
phantom (each day and after equipment changes), (2) determination of the
scattering and absorption coefficient for the phantom (from the measured
data using the homogeneous fitting algorithm), and (3) use of the difference
between calculated and measured values as the calibration factor.

2.2.2

Heterogeneous starting values

The image reconstruction algorithm starts from an estimated set of

optical properties and then calculates an update vector based on "l (the
squared difference between the calculated and measured values). If the
initially estimated set of optical properties is far from the actual optical
properties of the heterogeneous object being imaged, this inaccurate starting
point can slow convergence and even lead to erroneous answers. In order to
determine the initial optical property estimate for patient data or a
heterogeneous phantom, the measurements are averaged for each of the
sixteen sources and the homogeneous fitting algorithm is used to determine a
"homogeneous" estimate of the properties that are consistent with a diffusion
equation solution which matches the averaged data. In this manner, the
initial estimate is determined solely from the heterogeneous phantom or
patient measurements, yet, is close enough to the unknown actual
heterogeneous parameter distribution that the algorithm will converge.

91

Strategies For NIR Calibration

2.2.3

Long term drift

Differences between the source strength and initial phase at the time of
the measurement of the homogeneous calibration phantom and the actual
source strength and initial phase observed at the time of the heterogeneous
measurement can lead to an overall offset (mainly due to long term drift)
between the calibrated measured and calculated data. This difference in
overall offset needs to be removed in order for the reconstruction algorithm,
which fits to the absolute amplitude and phase data, to be effective. To
account for long term drift, the offset for both the homogeneous phantom
measurements and the actual heterogeneous measurement are calculated
based on the homogeneous fitting algorithm. (The homogeneous fitting
algorithm responds to the slope of the data and is therefore independent of
initial source strength and phase shift.) The offset is estimated as the
average difference between the measured and calculated data.
N

L (tP~easUred(hOmO)

tPojfset(homo)

- tP:alculated(hOmO)

=. . . : , i = . . : . l - - - - - N - - - - - -

(2)

L (tP~easUred(heterO)

ffojfset(hetero)

tPojJset(net)

where

- tP:alculated(heterO)

=. . . : , i = : : . : l - - - - - N - - - - - - -

= tPojfset(heterO) -

tPojfset(nel)

tPojJset(homO)

(3)
(4)

is the long term drift correction for initial phase and

source strength.

2.3

Homogeneous fitting algorithm

An important part of practical imaging is the homogeneous fitting

algorithm. The homogeneous fitting algorithm is critical for both calibration
of the system and for providing the initial optical property estimate for the
image reconstruction process. Assuming a homogeneous medium, only one
source location is needed to determine the absorption and scattering
coefficient of the material from measurements around its periphery. The
absorption and scattering coefficients which result in the best fit to the

92

McBride et al.

measured data can be determined by using a Newton-Raphson iterative

scheme applied to the finite element forward solution of the diffusion
equation for the relevant geometry. To reduce the effect of noise on the
fitting algorithm, the data acquired is averaged together for all sources based
on detector distance from the source location. The Newton-Raphson
solution is simplified by reducing the fit to two parameters: slope of the
phase with respect to distance from the source location and slope of the log
of intensity times distance with respect to distance. These two parameters
were chosen because they are nearly constant and can be obtained from the
data through linear regression. They are constant for the analytic infinite
medium diffusion equation solution [11]. This method is insensitive to noise
due to the large amount of averaging which occurs (256 measurements are
used to find two parameters).
Figure 5 contrasts two versions of the homogeneous fitting algorithm.
Method A performs a Newton-Raphson iterative scheme based on an
estimate of the first derivative for the averaged data (number of detectors
minus one (15) parameters), while Method B uses the slope of the phase
with respect to distance from the source location and slope of the log of
intensity times distance with respect to distance determined from linear
regression (2 parameters).

IMeasured data --

ZS6 (16 source x 16

detector) JMllnts for phase and Intensity

IAverage
data to I
I
source x detectors

Iteratively solve for optical

16

Method A: NewtonRaphson Iterative method

(minimization of I') based
On approximation or nrst
derivative of the measured
data

properties by comparing
wtth nnlte element fOl"ward
solution of dl/fuslon equation
for drcular geometry and a
homogeneous medium

Method B: Newton-Raphson
Iterative method based on slope of
de
the phase shirt versus distance (7,)
and slope oC In(dlstance intensity)
d(ln(dAC
versus distance (-dr- )

z' =I; Qin(I"'1 l-ln(I,'lHn(I':1l-ln(I,mlY

X' =[d(In~I"_d)) d(in~I'"'I""~'))J

+ cro."1 - 11,' ]- [0.:1 _ l1,m Y

+[ dfl'"'''_dr-I
dr
dr

P'1

where N =16 detectors, In(J')am (f are the

caJculaJed In(Iniensity) and phase shift from
finite element solution, am In(i") and Ir are
the measured In(Intensity) andphllJie.

dr

dr

[d(ln(>-i)

del

where all slope calculations -----;;:-. b

are perfonned with a linear regression to
the measured daJa or the calculated daJa
from the finite element solution

Figure 5. Schematic describing two homogeneous fitting algorithms.

93

Strategies For NIR Calibration

3.

RESULTS

3.1

Practical considerations

3.1.1

Homogeneous calibration

Figure 6 shows systematic offset from a slight difference in alignment of

two optical fibers with the laser source fiber during multiplexing. This offset
is consistent over a period of weeks and only changes during system
modifications/maintenance such as re-alignment of the laser. The resultant
difference data (Fig.6.c.) between the measured (Fig.6.a.) and the calculated
data (Fig.6.b.) is used as a calibration factor in the reconstruction algorithm.

(c)

(b)

(a)
6,00

6.00

5,00

5,00

5,00

' .00

' ,00

' ,00

3.00

3,00

6,00

~2 .00

~ 1 .00

0.00

3.00

~2 , OO

1.00

E0.00

'~2.00
~

~' ,oo

-0,00

1.00

1.00

1 ,00

2,00

2 ,00

2,00

3,00

3,00 -

- - - _J

-3 .00

Figure 6. Plot of (a) In(Intensity) data at 16 detector sites showing offset due to alignment
differences, (b) calculated data from finite element solution, and (c) difference between (a)
and (b). Difference data (c) is used as the 'calibration' factor.

3.1.2

Heterogeneous starting value

Simulations were performed to demonstrate the effect of different

starting values during reconstruction of a heterogeneous object. Images for
three different initial optical property estimates ( 0%, 10%, and 50% error
from the actual average optical properties) are compared in Figure 7 after
one iteration. The algorithm converges to the correct image (2: I absorbing
object with Gaussian profile) during the first iteration for an initial estimate
which is equal to the average optical properties of the heterogeneous test
object (Fig.7.a,d). When the initial estimate is significantly removed from
the actual average optical properties, the algorithm takes longer to converge.
This slowdown is evident in the other images in Figure 7.

94

McBride et al.

"r: "r-;
(a)

(b)

.
"
".

.,

',. .[7\ ' '.' [1\ "' ''[ 1

::,," " I

;1

.1

~iI

,
II

(d)

II

:!J

I.

,.

(t)

(e)

lin" .

tlIHt 4 UO

Cl .uuHIO

(lOUHCI

tlO,

II iIP "" .

tI OlljlUI

O.UtI Hlll

(J

n llO " lilll

II !lIl J

U U(l3110

II 4lD '0.
II UIIU .

,.,

10

,II

~ I)

)0

.H'. , / : :

111

I'll

)11

4U

J tI

OOl$C1

(J

OO}OO

I()

OtHSCl

OIHUCJ

10

i ll

;lO

40

JO

Figure 7. Image after first iteration of reconstruction of absorbing object with Gaussian
profile with (a) initial estimate equal to average optical properties, (b) \0% error in initial
estimate, and (c) 50% error in initial estimate. The vertical axis units are in mm t (d), (e),
and (t) are profile plots along a vertical line through the center of images (a), (b), and (c)
respectively.

3.1.3

()ffset

Offset between the natural log of intensity profile of a calculated and

measured phantom due to long term drift in the Ti:Sapphire laser intensity is
shown in Figure 8. The offset is accounted for by comparison of the changes
in offset between the homogeneous calibration phantom and the measured
heterogeneous object. In the extreme case shown in Figure 8, the offset in
intensity is almost two natural log units.
10.00
8.00
6.00
4.00 .

~ 2.00

'iii

!c

0.00

~ -2 . 00

-4 .00 .
-6.00
8.00
10.00

I.....
"*"

calibrated Measured d1
Calculated Data

_ .. J

Figure 8. Plot demonstrating measured data with overall offset due to long term drift of laser
intensity. The average difference is used to correct this offset.

95

Strategies For NIR Calibration

3.2

Testing of homogeneous fitting algorithm

Using repeated measurements of the same phantom, the results of the

homogeneous fitting algorithm (Method B in Figure 5) have been shown to
have an average deviation of 0.5% even in the presence of 5% measurement
noise. This compares very favorably with our previous approach (Method A
in Figure 5), which did not take full advantage of the data averaging, where
5% measurement noise translated to 5% average deviation in the
homogeneous fit. This decrease in deviation is demonstrated in Figure 9
which shows data for a phantom study involving triplicate measurement of
objects with increasing amounts of blood. The absorption coefficient
increases while the scattering coefficient remains constant. The study is
performed at three wavelengths and the slope of the line of increasing
absorption with added blood is compared with expected values of the molar
extinction coefficient from Wray et al. [8] in Table 1.
(a)
0.8
0.7
_ 06

0.5
1 0.4

'"
~

0.3
0.2
0. 1
0.0

(b)

t~'n

J
a

(c)

(d)

0.0 0 7 - , - - - - - - - - ---,

0.007 -

_ 0.006

0.006

.,

- -_ _-----1

~
~

.,

1 0.005

i o.005 '

..

~ 0.004

~
"'"
.,
IIJO

0.003 1 -_ _ _ _ _ __

ml bloo d I liter O. soh Intrallpld

ml blood I liter 0.5% Intrallpld

~ 0.004

0.2
0. 1 I
0.0 --,-~~-__._ _-___I

':" 0.6 t

-r-- - - - - - - - - - - - ,

0.8
0.7

mI blood 1111 0.5% Intrallpld

--'

'

0.003 - - - . - -

. 2

.~----

--;
5

ml blood I liter 0.5% Intrallpld

Figure 9. Homogeneous fits to measured data for the two methods described in Figure 5: (a)
Method A scattering coefficient data, (b) Method 8 scattering coefficient data, (c) Method A
absorption coefficient data, (d) Method B absorption coefficient data. Each concentration was
measured three times at each of three wavelengths.

96

McBride et al.

Molar Extinction Coefficient (mm1 mM1)

Wavelength Wrall et al. data
Method B
750 nm
3.10E-04
2.49E-04
800 nm
4.50E-04
5.06E-04
830 nm
5.30E-04
4.79E-04

Percent difference
-19.68%
12.44%
-9.62%

Table 1. Comparison of molar extinction coefficient for oxygenated hemoglobin measured

by Wray et al. [8] with results from use of homogeneous fitting algorithm (Method B.) based
on data from phantoms increasing hemoglobin concentration (Figure 9.d).

3.3

Modified imaging algorithm

The modified image reconstruction algorithm is shown schematically in

Figure 10. This modified algorithm uses the three practical imaging
considerations discussed in this paper. These modifications are denoted in
the figure by the thicker border lines.
Once a day MellllUfe
known hcmogenoous

phantom;- .:.....,.d (homo)

Mea!llred Phase and Amplilude

data, '-..ur.d (Iw..",) , 81256

p.xnts for objec:t ofinterest

"

Avet'l8e over 16 sources and obtain

estimate for qtical properties by Iterative
Least Squares Fit to data as!llming

bomogeoeous Iis!II

Use calilnled data:?ca.bralod(Iw'"",) - ....nvod(Iw,."') - (....surwd(ho'lfID) - caJaJaIod(ho'fOO - .ffi."..')

--+

Solve forward diffusion e"",lion using latest

eslimllle of absorplion and scaltering at each
finite .lemeDl node to determin. calculllled
phase and 8IIlplilude at 256 points.

J Original estimate of
I qtical properties

IEstimate
RdumNew I

Compare calculilled and measured

pbase and amplilude data at 256 points.
Error below
tolerance

ItolOIllDCO
Error above

~Fini!hed

Update vector baaed (]II full mabix

inveni (]II generatea new absorpti(]ll
and scattering III each melb node

Figure 10. Schematic of revised image reconstruction algorithm (compare with Figure 3)
which includes calibration and other practical imaging considerations added in bold boxes.

3.4

Phantom imaging results

The final calibrated system was used to measure a tissue-simulating

phantom. A representative image is shown in Figure 11.

97

Strategies For NIR Calibration

(a)

(c)

(b)

-0.015

I
I

-0.010
-0.005
-0.000

(d)
-60.000
-50.000

- 100.000
-80.000

-40.000

-60.000

-30.000
-20.000

__. . _ (D__

-10.000
-0.000

-40.000
-20.000
-0.000

60.0
40.0
20.0

E
J

0.0-

Figure 11. Representative phantom images from the calibrated imaging system. The 90 mm
diameter phantom consists of a mixture of 0.5% Intralipid and 20 microMolar hemoglobin
concentration in water. The hemoglobin in the phantom is fully oxygenated except for a 24
mm diameter object to the right of center which was de-oxygenated by bubbling with
Nitrogen gas. Absorption coefficient images (scale is rom-I) of the phantom are shown (a) at
720 nm, (b) at 750 nm, and (c) 800 nm. These three images were combined to form images of
(d) hemoglobin concentration (units are microMolar) and (e) hemoglobin oxygen saturation
(in percent oxygenation). (t) and (g) are horizontal profile plots through the center of (d) and
(e) respectively.

4.

DISCUSSION

4.1

Practical considerations

Practical considerations for a NIR diffuse imaging system include: (1)

system calibration with a homogeneous phantom, (2) use of a homogeneous
fitting algorithm to arrive at an initial optical property estimate for image
reconstruction of a heterogeneous medium, and (3) correction for long term

98

McBride et al.

drift by subtraction of source strength and initial phase offsets. These

practical considerations ar~ necessary when using data from a real imaging
system with a model-based reconstruction scheme in order to reduce datamodel mismatch errors associated with system imperfections which are not
idealized in the model. These modifications allow for the realization of a
prototype imaging system that has the potential to quantitatively reconstruct
heterogeneous images of hemoglobin-related parameters of highly scattering
tissue in vivo.

4.2

Testing of homogeneous fitting algorithm

An important component of our practical imaging system is a robust,

stable homogeneous fitting algorithm. One method which is insensitive to
noise due to the large amount of data averaging which is possible (256
measurements are used to find two parameters) is a Newton-Raphson
minimization based on two parameters: slope of the phase with respect to
distance from the source location and slope of the log of intensity times
distance with respect to distance. This method produces only 0.5% deviation
in determined optical properties in the presence of over 5% measurement
noise in repeated phantom measurements.

5.

CONCLUSION

Quantitatively accurate images of hemoglobin related parameters

recovered on an absolute scale have been demonstrated with no a priori
information. Production of these quantitative images of hemoglobin
concentration and oxygen saturation in a tissue-simulating phantom were
critically dependent on a prototype imaging system implementation which
includes the practical considerations of source/detector calibration, an
accurate first estimate of the optical property distribution, and a correction
for long term source/sensor drift. Once properly calibrated, this nearinfrared tomographic imaging system has the potential to provide in vivo
images of hemoglobin related parameters which may make it useful in breast
cancer imaging and therapy monitoring.

ACKNOWLEDGEMENTS
This work has been supported by the NIH through grants #ROICA69544 and
POICA80139 awarded by the National Cancer Institute. Authors gratefully acknowledge
previous development work by Huabei Jiang and David Rinehart.

Strategies For NIR Calibration

99

REFERENCES
I.

Jiang H, Paulsen KD, Osterberg UL, Pogue BW, Patterson MS. Optical image
reconstruction using frequency-domain data: simulations and experiments. J Opt Soc Am
A 1996;13:253-266.
2. Arridge SR, Schweiger M. Image reconstruction in optical tomography. Phil Trans R.
Soc Lond B 1997;352:717-726.
3. Profio AF, Navarro GA. Scientific basis of breast diaphanography. Med Phys 1989;
16:60-65.
4. Vaupel P, Kallinowski F, Okunieff P. Blood flow, oxygen and nutrient supply, and
metabolic microenvironment of human tumors: a review. Cancer Research 1989;49:
6449-6465.
5. McBride TO, Pogue BW, Gerety ED, Poplack SB, Osterberg UL, Paulsen KD.
Spectroscopic diffuse optical tomography for quantitatively assessing hemoglobin
concentration and oxygenation in breast tissue. Appl Opt 1999;38:5480-5490.
6. Pogue BW, Testorf M, McBride T, Osterberg U, Paulsen K. Instrumentation and design
of a frequency-domain diffuse optical tomography imager for breast cancer detection.
Optics Express 1997;1:391-403.
7. Pogue BW, McBride TO, Osterberg UL, Paulsen KD. Spatially variant regularization
improves diffuse optical tomography. Appl Opt 1999;38:2950-2961.
8. Wray S, Cope M, Delpy DT, Wyatt JS, Reynolds EOR. Characterization of the near
infrared absorption spectra of cytochrome aa3 and haemoglobin for the non-invasive
monitoring of cerebral oxygenation. Biochim Biophys Acta 1988;933: 184-192.
9. Quaresima V, Matcher SJ, Ferrari M. Identification and quantification of intrinsic optical
contrast for near-infrared mammography. Photochem Photobiol 1998;67:4-14.
10. Hale GM, Querry MR, Optical constants of water in the 200-nm to 200- Ilm wavelength
region. Appl Opt 1973;12:555-563.
11. Pogue BW, Patterson MS, Frequency-domain optical absorption spectroscopy of finite
element tissue volumes using diffusion theory. Phys Med Bio 1994;39:1157-1180