Reducing image artifact in diffuse

optical tomography by iterative

perturbation correction based on
multiwavelength measurements

K. M. Shihab Uddin
Quing Zhu

K. M. Shihab Uddin, Quing Zhu, “Reducing image artifact in diffuse optical tomography by iterative
perturbation correction based on multiwavelength measurements,” J. Biomed. Opt. 24(5),
056005 (2019), doi: 10.1117/1.JBO.24.5.056005.
Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Journal of Biomedical Optics 24(5), 056005 (May 2019)

Reducing image artifact in diffuse optical tomography

by iterative perturbation correction based on
multiwavelength measurements
K. M. Shihab Uddin and Quing Zhu*
Washington University in St Louis, Biomedical Engineering Department, St. Louis, Missouri, United States

Abstract. Ultrasound (US) guided diffuse optical tomography has demonstrated great potential for breast
cancer diagnosis, treatment monitoring, and chemotherapy response prediction. Optical measurements of
four different wavelengths are used to reconstruct unknown optical absorption maps, which are then used
to calculate the hemoglobin concentration distribution of the US visible lesion. Reconstructed absorption
maps are prone to image artifacts from outliers in measurement data from tissue heterogeneity, bad coupling
between tissue and light guides, and motion by patient or operator. We propose an automated iterative pertur-
bation correction algorithm to reduce image artifacts based on the structural similarity index (SSIM) of absorption
maps of four optical wavelengths. The initial image is estimated from the truncated pseudoinverse solution.
The SSIM was calculated for each wavelength to assess its similarity with other wavelengths. An absorption
map is repeatedly reconstructed and projected back into measurement space to quantify projection error.
Outlier measurements with highest projection errors are iteratively removed until all wavelength images are
structurally similar with SSIM values greater than a threshold. Clinical data demonstrate statistically significant
improvement in image artifact reduction. © The Authors. Published by SPIE under a Creative Commons Attribution 4.0 Unported License.
Distribution or reproduction of this work in whole or in part requires full attribution of the original publication, including its DOI. [DOI: 10.1117/1.JBO.24.5
.056005]
Keywords: diffuse optical tomography; image reconstruction techniques; breast cancer diagnosis; image artifact reduction.
Paper 180668R received Dec. 14, 2018; accepted for publication Apr. 19, 2019; published online May 22, 2019.

1 Introduction artifact removal.20 To improve the imaging quality, projection

Breast cancer is the second leading cause of death for women in error-based adaptive regularization techniques have been
employed, which outperform standard Tikhonov regularization.21
the United States.1 Diffuse optical tomography (DOT) and dif-
However, these approaches do not compensate for wavelength-
fuse optical spectroscopy are functional imaging modalities
dependent problems. For example, the 740-nm wavelength is
being explored for breast cancer diagnosis and treatment
prone to measurement errors from tissue heterogeneity caused
prediction.2–9 DOT provides quantitative estimations of oxygen-
by dark skin and skin pigment. As another example, the 830-nm
ated and deoxygenated hemoglobin concentrations calculated
wavelength has lower signal to noise ratio at longer source and
from reconstructed optical absorption maps of breast lesions.
detector distances due to the reduced sensitivity of photomulti-
Hemoglobin concentrations are directly related to tumor angio- plier tube (PMT) detector beyond 800 nm.
genesis, a hallmark of cancer. Since diffused light suffers from In our US-guided DOT approach to assessing breast cancer,
poor lesion localization and inaccurate target quantification due perturbation, which is the normalized difference between the
to intense scattering, image reconstruction is often guided by lesion breast and the contralateral normal breast (reference)
another high-resolution imaging modality, such as ultrasound measurements, is used for mapping lesion absorption at each
(US),6 magnetic resonance imaging,8,9 or x-ray,10 which provide wavelength. The total hemoglobin map is computed from the
structural information about the tumor. While DOT guided absorption maps of four optical wavelengths. The tissue hetero-
by other modalities can significantly reduce the number of geneity of the reference measurements contributes to outliers in
voxels with unknown optical properties for more accurate the perturbation measurements. In a recent investigation by our
reconstruction,7 image artifacts caused by outlier measurements group, Vavadi et al.22 introduced a statistical method based on
due to wavelength-dependent tissue heterogeneity, bad coupling the semi-infinite tissue model to automatically remove outliers
between tissue and sources and detectors, and patient motion or from contralateral normal breast measurements. However, this
operator hand motion can distort the reconstructed images. method cannot be used for perturbation measurements because
Several experimental and modeling approaches have been lesion measurements are expected to be more heterogeneous
developed for system calibration of optical source strengths, than the reference measurements. To separate the measurement
detection channel gains and phase shifts,11–13 source and detec- errors from lesion heterogeneity, more information from multi-
tor (optodes) position errors, and coupling errors between skin ple wavelength measurements can be incorporated in the prepro-
and optodes.12,14–17 For correcting motion artifacts, different cessing before image reconstruction. Recently, Althobaiti et al.23
algorithms have been proposed, including cubic spline interpo- introduced an approach for data filtering based on multiple
lation,18 adaptive Kalman filtering,19 and wavelet-based motion wavelength measurements collected at the lesion site. The
method combines data collected from multiple sets of lesion
measurements to detect and correct outliers caused by wave-
*Address all correspondence to Quing Zhu, E-mail: [email protected] length-dependent measurement errors in the perturbation.

Journal of Biomedical Optics 056005-1 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Uddin and Zhu: Reducing image artifact in diffuse optical tomography. . .

However, this approach requires that two to three wavelength given as Ur ðiÞ ¼ Ar ðiÞejφr ðiÞ , and the lesion measurement is
perturbation datasets must be correlated, and then the rest of U l ðiÞ ¼ Al ðiÞejφl ðiÞ , where i ¼ 1; 2; : : : ; m, and m is the total
the wavelength-dependent distortion can be compensated for. number of source–detector pairs or the total number of measure-
In this paper, we propose an iterative perturbation correction ments. Perturbation, U sc ðiÞ, is defined as the normalized differ-
algorithm by using structural similarity index (SSIM) as an ence between the reference and target measurements:
image quality assessment criterion. The initial estimate of the
Al ðiÞejφl ðiÞ − Ar ðiÞejφr ðiÞ
EQ-TARGET;temp:intralink-;e001;326;697

absorption map is obtained from the truncated pseudoinverse

solution. In subsequent iterations, the average SSIM for each U sc ðiÞ ¼
Ar ðiÞejφr ðiÞ
wavelength and errors between the measurement and the pro-

jected data are computed and outliers are removed from the mea- Al ðiÞ
¼ cosðφl ðiÞ − φr ðiÞÞ − 1
surements based on the errors. This procedure is iterated until Ar ðiÞ
the SSIM reaches a preset threshold. We demonstrate the effec-

A ðiÞ
tiveness of this approach in phantoms and clinical data. þj l sinðφl ðiÞ − φr ðiÞÞ : (1)
Ar ðiÞ
2 Methods
The first term is the real part of the perturbation, and the sec-
2.1 US-guided DOT System, Data Acquisition, and ond term is the imaginary part of the perturbation. A typical 2-D
Calibration representation of perturbation data for a phantom is shown in
Fig. 1, with real perturbation on the x axis and imaginary per-
US-guided DOT system is a frequency domain imaging system turbation along the y axis. The unit circle represents the
consisting of a hand-held DOT imaging probe with an US trans- expected boundary that perturbation data should lie within.
ducer located in the middle.24 Four laser diodes with wave- From simulations of different target contrasts and locations in
lengths 740, 780, 808, and 830 nm are sequentially switched depths, it was shown that maximum phase difference for
by a 4 × 1 and a 1 × 9 optical switch to deliver light modulated any source–detector pair should not exceed 90 deg, even in
at 140 MHz to each of the nine source positions on the probe.
extreme cases.22 Since 0 ≤ cosðφl ðiÞ − φr ðiÞÞ ≤ 1 for
Fourteen parallel PMTs detect reflected light via light guides
− π2 ≤ φl ðiÞ − φr ðiÞ ≤ π2 for all i, real perturbation should be
from the tissue. A custom A/D board samples detected signals
greater than −1. For a high-contrast target, Al ≪ Ar , so pertur-
from all channels and stores data in a PC. Using data collected
bation is more skewed toward the negative real axis. For a low-
from a homogenous intralipid solution, the system is calibrated
contrast target, perturbation may be small and clustered around
to compensate for system gains and phase delays. Multiple data-
the origin, or distributed more toward the positive real axis, or
sets acquired from the contralateral normal breast are used to
distributed evenly across both positive and negative sides around
compute a robust reference.22 The selected reference is consid-
ered as a homogeneous reference. The fitted optical absorption the origin. Note that the imaginary part of the perturbation can
(μa0 ) and reduced scattering (μs00 ) are used to calculate a weight be positive or negative because −1 ≤ sin½φl ðiÞ − φr ðiÞ ≤ 1
matrix, W, for image reconstruction using a dual mesh scheme. for − π2 ≤ φl ðiÞ − φr ðiÞ ≤ π2.
Figure 2 shows one set of perturbation data of a highly
absorbing malignant breast lesion and a low absorbing benign
2.2 Data Preprocessing
breast lesion. As evident from the figure, clinical data is more
DOT data acquisition is performed on both a lesion breast and a scattered because of tissue heterogeneity, bad coupling of the
contralateral normal breast, referred to as the reference breast. tissue and source and detector fibers, and patient movement
Amplitude and phase measurements are extracted from the or operator hand motions.
detected radio frequency signal using the Hilbert transform. Multiple perturbation datasets are compiled together, and
For the i’th source–detector pair, reference measurement is a multivariate Gaussian is fitted, and data points are removed

Fig. 1 Phantom perturbation data. (a) Data measured from a high-contrast phantom target imbedded in
intralipid solution. (b) Data measured from a low-contrast phantom target imbedded in intralipid solution.

Journal of Biomedical Optics 056005-2 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Uddin and Zhu: Reducing image artifact in diffuse optical tomography. . .

Fig. 2 Clinical perturbation data. (a) A malignant breast lesion and (b) a benign breast lesion.

measurement’s sensitivity to the absorption and scattering

changes. In the end, the reconstruction problem can be formu-
lated as a regularized optimization problem:


λ
fðxÞ ¼ arg minX kUsc − WXk2 þ kðX − X 0 Þk2 :
EQ-TARGET;temp:intralink-;e003;326;500 (3)
2

A two-step imaging reconstruction was proposed earlier to

solve this regularized optimization problem after obtaining an
initial image estimate, X 0 , by truncated pseudoinverse in the
first step and refining the solution using the regularized conju-
gate gradient (CG) algorithm.25 The implementation of CG is
adapted from Ref. 26.

2.4 Image Quality Assessment

In DOT reconstruction, quantitative assessment of imaging qual-
Fig. 3 Data preprocessing and iterative perturbation correction
ity poses a challenge. Previously, image distortion and inconsis-
algorithms. tent images for different wavelengths were visually inspected,
and perturbation was manually corrected by an experienced
operator. Such manual data processing is operator dependent
if there are any with a Mahalanobis distance greater than the and time consuming. In this manuscript, we propose to use
threshold computed from the inverse chi-square distribution the SSIM to quantitatively evaluate imaging quality by taking
with a cumulative probability of 99%. Based on chest wall all four wavelength images into account. The SSIM measure is
matching of the reference and lesion breast, a single measure- a function of the image’s luminance, contrast, and structure.27,28
ment dataset is selected from multiple measurements. The struc- The SSIM between two images X and Y is defined as in Ref. 27:
tural similarity-based perturbation correction algorithm depicted
in Fig. 3 is applied to perturbation data to obtain corrected per- SSIMðX; YÞ ¼ ½lðX; YÞα · ½cðX; YÞβ · ½sðX; YÞγ ;
EQ-TARGET;temp:intralink-;e004;326;249 (4)
turbation and artifact-free images.
where lðX; YÞ, cðX; YÞ, and sðX; YÞ are the luminance, contrast,
and structure similarity, respectively, and α > 0, β > 0, γ > 0
2.3 DOT Image Reconstruction are three parameters used to adjust relative importance of the
The DOT inverse problem is typically linearized by Born three components of the similarity measure. The luminance,
approximation. By digitizing the imaging space into N voxels, contrast, and structure of an image are computed from mean,
the resulting integral equations are formulated as follows: standard deviation, and normalized images28 as follows:

lðX; YÞ ¼ ð2μX μY þ C1 Þ∕ðμ2X þ μ2Y þ C1 Þ;

EQ-TARGET;temp:intralink-;e005;326;151

½Usc M×1 ¼ ½WM×N ½δμa N×1 ¼ WX;

EQ-TARGET;temp:intralink-;e002;63;152 (2)
cðX; YÞ ¼ ð2σ X σ Y þ C2 Þ∕ðσ 2X þ σ 2Y þ C2 Þ;
where Usc is the measured scattered photon density wave, M is
the number of measurements, and δμa denotes the unknown sðX; YÞ ¼ ðσ XY þ C3 Þ∕ðσ X σ Y þ C3 Þ: (5)
differences in the absorption coefficients of each voxel. The
weight matrix, W, describes the distribution of the diffused Here, μX , μY , σ X , σ Y , andσ XY are the means of pixel values of
wave in the homogenous medium and characterizes the image X and image Y, the standard deviation of image X and

Journal of Biomedical Optics 056005-3 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Uddin and Zhu: Reducing image artifact in diffuse optical tomography. . .

image Y, and the covariance of image X and Y, respectively. C1 , the absorption map for wavelength λi using regularized CG.
C2 , and C3 are constants. SSIMðλi Þ is recomputed and compared with the threshold.
For each wavelength, λi ∈ f740; 780; 808; 830 nmg, the This process is repeated until the lowest SSIMðλi Þis greater
other three wavelength images are used as references to compute or equal to the threshold. This iterative correction procedure
SSIMs for three image pairs. An average of the three SSIMs is is performed for each wavelength until the SSIMðλi Þvalues
the quantitative image quality index, SSIMðλi Þ, used to evaluate for all four wavelengths are above the threshold.
the reconstructed image quality of wavelength λi as given Note that there is a wavelength-dependent variation in
below: absorption values of different wavelengths for different oxygen
nwavelength
1 X conditions. However, the same lesion should have similar
SSIMðλi Þ ¼ SSIMðimagei ; imagej Þ: absorption distributions over the narrow wavelength window
nwavelength − 1
EQ-TARGET;temp:intralink-;e006;63;664

j¼1;j≠i of 730–830 nm. Thus, we use SSIM to characterize the similar-

(6) ity between wavelengths. The algorithm does not enforce a per-
fect or 100% image similarity but somewhat similar not totally
different as determined by this preset threshold.
2.5 Iterative Perturbation Correction
Iterative perturbation correction is performed based on
3 Results
SSIMðλi Þ for each wavelength. The wavelength with the mini-
mum SSIMðλi Þ is corrected first. The initial estimate is from the 3.1 Phantom Experiments
truncated pseudoinverse. If SSIMðλi Þ is lower than a preset
threshold (0.9), perturbation from λi wavelength is corrected In phantom experiments, intralipid solution was used to
based on the original perturbation and projected perturbation. simulate a homogeneous background medium. The experi-
The reconstructed image, δμa0 , for λi is projected into measure- ment was repeated for solid spherical balls with different con-
ment space by multiplying the weight matrix, W, to obtain pro- trasts and sizes simulating different types of tumors in
jected data: intralipid solution in different depths. The average image sim-
ilarity index was computed for all phantom absorption map
½Uprojected  ¼ ½W½δμa0 : (7)
images. The reconstructed absorption map for a ball of 2-
EQ-TARGET;temp:intralink-;e007;63;493

Based on the Euclidean distance of original perturbation cm diameter at 2-cm depth is shown in Fig. 4. The average
data, U sc , and projected data, Uprojected , projection error, Eproj , structural similarity indices for four wavelengths —740,
is calculated as follows: 780, 808, and 830 nm—are 0.98, 0.97,0.99, and 0.96,
respectively.
Eproj ¼ kU projected − Usc k2 :
EQ-TARGET;temp:intralink-;e008;63;433 (8) Pairwise SSIMs (mean  standard deviation) are presented
in Table 1. Large similarity indices indicate strong structural
The data point with maximum projection error is removed similarity among different wavelengths, which is visually appar-
from U sc . Modified perturbation is again used to reconstruct ent in Fig. 4.

Fig. 4 Reconstructed image similarity for phantom data: (a) US image and (b) reconstructed absorption
maps (two layers at z ¼ 1.5 cm and z ¼ 2 cm) for all four wavelengths. Each 2-D layer is 8 cm × 8 cm.
Average SSIMs are 0.98, 0.97, 0.99, and 0.96 for 740,780, 808, and 830 nm, respectively.

Table 1 SSIM (mean ± standard deviation) for phantom data.

740 nm 780 nm 808 nm 830 nm

740 nm — 0.976  0.004 0.988  0.003 0.942  0.015

780 nm 0.976  0.004 — 0.985  0.006 0.943  0.023

808 nm 0.988  0.003 0.985  0.006 — 0.947  0.020

830 nm 0.942  0.015 0.943  0.023 0.947  0.020 —

Average 0.954  0.019 0.969  0.008 0.957  0.018 0.974  0.009

Journal of Biomedical Optics 056005-4 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Uddin and Zhu: Reducing image artifact in diffuse optical tomography. . .

3.2 Clinical Study coefficients change to 0.1582, 0.1470, 0.1345, and

0.1484 cm−1 , respectively.
A clinical study was approved by the local Institutional Review Iterative changes in the perturbation and absorption map for
Board and was compliant with the Health Insurance Portability this case at 830 nm are shown in Fig. 6. For iteration 0, we see
and Accountability Act. Informed consent was given by each the original dataset and the absorption map: the map is similar to
patient. Data used in this study have been deidentified. A total that in Fig. 5(b). In successive iterations, we removed perturba-
of 40 patients were studied including 13 malignant and 27 tion points denoted by red dots. We continue to remove pertur-
benign lesions, based on biopsy results. All patients were cat- bations until absorption map is structurally similar to the maps
egorized into two categories: patients with image artifact present of other wavelengths.
in one or more wavelength absorption maps (17 patients) and An example of malignant breast cancer with mixed ductal
patients with no image artifact (23 patients). This categorization and lobular features is shown in Fig. 7. Figure 7(a) shows an
was based on a preset cutoff value of 0.9 for structural similarity US image with a lesion marked by a white ellipse. Figure 7(b)
among the four wavelengths reconstructed absorption map shows reconstructed absorption maps for the four wavelengths.
images. An example of benign fibroadenoma is shown in Each wavelength absorption map shows two layers at depths,
Fig. 5. Figure 5(a) shows the US image with the lesion marked z ¼ 1.5 and 2 cm. Mean image similarity indices for the four
by a white ellipse. Figure 5(b) shows reconstructed absorption wavelengths 740, 780, 808, and 830 nm are 0.83, 0.82, 0.77,
maps for four wavelengths. Each wavelength absorption map and 0.79, and reconstructed maximum absorption coefficients
has one 2-D layer at depth, z ¼ 1 cm. The mean SSIMs for are 0.254, 0.237, 0.070, and 0.054 cm−1 , respectively. Image
the four wavelengths 740, 780, 808, and 830 nm are 0.87, artifact is present in 808- and 830-nm absorption maps.
0.91, 0.87, and 0.82. The reconstructed maximum absorption Figure 7(c) shows reconstructed absorption maps after perturba-
coefficients are 0.2463, 0.2069, 0.1326, and 0.3316cm−1 , tion correction. Mean SSIMs for four wavelengths improve to
respectively. Image SSIM indicates that there is an image artifact 0.94, 0.94, 0.93, and 0.91 while reconstructed maximum
at wavelength 830 nm, and visual inspection confirms this. absorption coefficients changed to 0.254, 0.237, 0.228, and
Figure 5(c) shows reconstructed absorption maps after perturba- 0.175 cm−1 , respectively.
tion correction. The mean SSIMs for the four wavelengths Figure 8 shows iterative changes of perturbation and absorp-
change to 0.95, 0.97, 0.97, and 0.96 while maximum absorption tion maps for the malignant case at 808 nm. In iteration 0, we

Fig. 5 Image artifact reduction for a benign case: (a) US image, 1 cm lesion depth, (b) absorption maps
for original data before perturbation correction, and (c) absorption maps after perturbation correction.

Fig. 6 Iterative changes in absorption map and perturbation filtering for 830 nm for the benign case. Red
dots denote removed data points.

Journal of Biomedical Optics 056005-5 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Uddin and Zhu: Reducing image artifact in diffuse optical tomography. . .

Fig. 7 Image artifact reduction for a malignant case: (a) US image, 1.5- and 2-cm lesion depths,
(b) absorption maps for original data before perturbation correction, and (c) absorption maps after per-
turbation correction.

Fig. 8 Iterative changes of absorption map and perturbation filtered at 808 nm for the malignant case.
Red dots denote removed data points.

have original dataset and absorption map similar to that in Student’s t-test on images with no artifacts shows no significant
Fig. 7(b). In successive iterations, we removed perturbation change in terms of structural similarity (p-value 0.52), which is
points denoted by red dots. We continue to remove perturbations expected.
until the absorption map is structurally similar to maps of other
wavelengths. 4 Discussion and Summary
Perturbation correction statistically improves the SSIM In summary, an iterative perturbation correction algorithm based
among different wavelengths, as depicted in Fig. 9. A two-tailed on image similarity is introduced and its performance in image
paired t-test was done for images with artifacts both before and artifact reduction is demonstrated using clinical data. This algo-
after perturbation correction, and the SSIM is statistically higher rithm follows two simple assumptions. First, absorption map
after perturbation correction, with a p-value less than 0.001. images for all four wavelengths are assumed to be structurally

Journal of Biomedical Optics 056005-6 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Uddin and Zhu: Reducing image artifact in diffuse optical tomography. . .

When each wavelength data is used for reconstructing each

absorption map and then all absorption maps are used to com-
pute HbO2 and Hb, there is no cross-coupling of HbO2 and Hb.
In our earlier studies, we have attempted to simultaneously
reconstruct both absorption and scattering distributions of breast
lesions.30 We had success in phantom data but these algorithms
are not robust for patient data when they applied to a large
patient database. In the simultaneous reconstruction, the distri-
bution of lesion diffusion coefficient, DðrÞ ¼ 1∕½3 × μs 0 ðrÞ, is
reconstructed with the distributin of lesion absorption, μa ðrÞ,
where r ¼ ðx; y; zÞ. However, DðrÞ is one order of magnitude
smaller than μa ðrÞ and cannot be reconstructed reliably for
all patient data. Additionally, simultaneously reconstruction
of both absorption and scattering distributions doubles the
unknowns in the image reconstruction. Since μa ðrÞ is directly
related to tumor angiogenesis and we have focused on this
important parameter in our algorithm development
In summary, the proposed iterative artifact reduction algo-
Fig. 9 Comparison of SSIMs of reconstructed images before pertur- rithm significantly reduces the effect of wavelength-dependent
bation correction (blue box) and after perturbation correction (red measurement errors in DOT perturbation, which helps to
box). achieve a more accurate reconstruction of the optical properties
of breast lesions. This automated method also helps to minimize
both the user interface and the time for data preprocessing. The
similar. Since we are imaging the same tissue region with average time for an experienced user to manually perform data
closely spaced wavelengths, the image structure should be sim- preprocessing for one patient’s data is from 15 to 30 min. The
ilar, even though the local absorption coefficients might differ automated method could reduce this time to less than a minute
due to wavelength-dependent absorption variations. Second, and facilitate the clinical translation of US-guided DOT technol-
image artifacts in all four wavelengths are assumed to be dis- ogy. Although the method is demonstrated using US-guided
similar. Data acquisition is done sequentially, from one DOT data, it is applicable to any DOT data preprocessing
wavelength after another, in a few seconds. Motion or experi- obtained with multiple wavelengths.
mental errors can affect one or more wavelengths, but these
effects are random and are unlikely to generate structurally sim-
ilar artifacts at all four wavelengths. Additionally, certain tissue Disclosures
heterogeneity caused artifacts may not be present at all wave- No potential conflicts of interests to disclose.
lengths, for example, 740 nm is very sensitive to dark skin pig-
ment than other wavelengths. The total hemoglobin distribution,
which is calculated by linear weighting of multiwavelength Acknowledgments
absorption maps based on extinction coefficients, is signifi- The authors acknowledge and are grateful for funding support
cantly improved due to artifacts reduction. However, the average for this work from NIH (Grant Nos. RO1EB002136 and
maximum total hemoglobin levels which we have used to clas- RO1228047). The authors also thank James Ballard for proof
sify malignant versus benign lesions remain statistically the reading of the manuscript.
same as compared with no perturbation correction. This is
because one or two absorption distributions are often signifi-
cantly improved on artifacts, but the maximum total hemoglobin References
level may not change much. 1. National Breast Cancer Foundation Inc., 2018, https://www
Other approaches, such as the weighted least-square (WLS) .nationalbreastcancer.org/.
approach,29 can compensate measurement differences between 2. Q. Zhu et al., “Benign versus malignant breast masses: optical differ-
entiation with US-guided optical imaging reconstruction,” Radiology
different wavelengths (or source/detector channels). Such a
237(1), 57–66 (2005).
method depends on the accurate modeling of the system 3. R. Choe et al., “Differentiation of benign and malignant breast tumors
noise. However, the noise in the DOT measurements includes by in-vivo three-dimensional parallel-plate diffuse optical tomography,”
coherent and incoherent components as well as random motion J. Biomed. Opt. 14(2), 024020 (2009).
produced noise. We have evaluated the WLS approach to com- 4. B. J. Tromberg et al., “Imaging in breast cancer: diffuse optics in breast
pensate measurement differences between different wavelengths cancer: detecting tumors in pre-menopausal women and monitoring
neoadjuvant chemotherapy,” Breast Cancer Res. 7(6), 279–285 (2005).
by downweighting the noisy data instead of completely remov-
5. S. P. Poplack et al., “Electromagnetic breast imaging: results of a pilot
ing them. However, we have found that WLS may not be suit- study in women with abnormal mammograms,” Radiology 243(2), 350–
able for the measurement noise we experienced in the patient 359 (2007).
data. 6. C. Xu et al., “Ultrasound-guided diffuse optical tomography for predict-
In the past, we have performed simultaneously the ing and monitoring neoadjuvant chemotherapy of breast cancers: recent
reconstruction of oxygenated hemoglobin, HbO2 , and deoxy- progress,” Ultrason. Imaging 38(1), 5–18 (2016).
7. A. Mostafa et al., “Diffuse optical tomography using semiautomated
genated hemoglobin, Hb. However, because the HbO2 is
coregistered ultrasound measurements,” J. Biomed. Opt. 22(12),
about twice as large as Hb, the Hb is over-reconstructed and 121610 (2017).
HbO2 is under-reconstructed due to the weight distribution in 8. V. Ntziachristos et al., “MRI-guided diffuse optical spectroscopy of
weight matrix for Hb and HbO2 in simultaneous reconstruction. malignant and benign breast lesions,” Neoplasia 4(4), 347–354 (2002).

Journal of Biomedical Optics 056005-7 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Uddin and Zhu: Reducing image artifact in diffuse optical tomography. . .

9. M. A. Mastanduno et al., “Sensitivity of MRI-guided near-infrared spec- 23. M. Althobaiti, H. Vavadi, and Q. Zhu, “An automated preprocessing
troscopy clinical breast exam data and its impact on diagnostic perfor- method for diffuse optical tomography to improve breast cancer diag-
mance,” Biomed. Opt. Express 5(9), 3103–3115 (2014). nosis,” Technol. Cancer Res. Treat. 17, 153303381880279 (2018).
10. Q. Fang et al., “Combined optical and X-ray tomosynthesis breast 24. H. Vavadi et al., “Compact ultrasound-guided diffuse optical tomogra-
imaging,” Radiology 258(1), 89–97 (2011). phy system for breast cancer imaging,” J. Biomed. Opt. 24(2), 021203
11. D. A. Boas, T. Gaudette, and S. R. Arridge, “Simultaneous imaging and (2018).
optode calibration with diffuse optical tomography,” Opt. Express 8(5), 25. K. M. S. Uddin et al., “Two step imaging reconstruction using truncated
263–270 (2001). pseudoinverse as a preliminary estimate in ultrasound guided diffuse
12. J. J. Stott et al., “Optode positional calibration in diffuse optical tomog- optical tomography,” Biomed. Opt. Express 8(12), 5437–5449 (2017).
raphy,” Appl. Opt. 42(16), 3154–3162 (2003). 26. W.H. Press et al., Numerical Recipes in C (2nd ed.): The Art of Scientific
13. N.G. Chen et al., “Simultaneous near-infrared diffusive light and ultra- Computing, Cambridge University Press, New York (1992).
sound imaging,” Appl. Opt. 40(34), 6367–6380 (2001). 27. Z. Wang et al., “Image quality assessment: from error visibility to struc-
14. J. P. Culver et al., “Volumetric diffuse optical tomography of brain activ- tural similarity,” IEEE Trans. Image Process. 13(4), 600–612 (2004).
ity,” Opt. Lett. 28(21), 2061–2063 (2003). 28. G. P. Renieblas et al., “Structural similarity index family for image qual-
15. M. Schweiger et al., “Image reconstruction in optical tomography in ity assessment in radiological images,” J. Med. Imaging 4(3), 035501
the presence of coupling errors,” Appl. Opt. 46(14), 2743–2756 (2017).
(2007). 29. S. M. Kay, Fundamentals of Statistical Signal Processing, Prentice-Hall
16. M. Mozumder et al., “Compensation of optode sensitivity and position Inc., Upper Saddle River, New Jersey (1993).
errors in diffuse optical tomography using the approximation error 30. B. Tavakoli and Q. Zhu, “Two-step reconstruction method using global
approach,” Biomed. Opt. Express 4(10), 2015–2031 (2013). optimization and conjugate gradient for ultrasound-guided diffuse opti-
17. R. Fukuzawa et al., “Reduction of image artifacts induced by change in cal tomography,” J. Biomed. Opt. 18(1), 016006 (2013).
the optode coupling in time-resolved diffuse optical tomography,”
J. Biomed. Opt. 16(11), 116022 (2011). K. M. Shihab Uddin is a PhD candidate in the Biomedical
18. F. Scholkmann et al., “How to detect and reduce movement artifacts in Engineering Department at Washington University in St. Louis. He
near-infrared imaging using moving standard deviation and spline inter- received his bachelor’s degree in electrical and electronics engineer-
polation,” Physiol. Meas. 31(5), 649–662 (2010). ing from the Bangladesh University of Engineering and Technology in
19. M. Izzetoglu et al., “Motion artifact cancellation in NIR spectroscopy Bangladesh. His research is focused on data analysis, nonlinear opti-
using discrete Kalman filtering,” Biomed. Eng. Online 9(1), 16 mization based reconstruction, and classification of breast lesions
(2010). based on ultrasound-guided diffuse optical tomography.
20. B. Molavi and G. A. Dumont. “Wavelet-based motion artifact removal
Quing Zhu is a professor of the Department of Biomedical
for functional near-infrared spectroscopy,” Physiol. Meas. 33(2), 259–
Engineering and Radiology at Washington University in St Louis.
270 (2012). She is a pioneer of combining ultrasound and near infrared imaging
21. H. Niu et al., “Improving image quality of diffuse optical tomography modalities for diagnosis and treatment assessment of breast cancers.
with a projection-error-based adaptive regularization method,” Opt. She has been named a fellow of the Optical Society of America,
Express 16(17), 12423–12434 (2008). and a fellow of SPIE. Her research interests are cancer detection,
22. H. Vavadi and Q. Zhu, “Automated data selection method to improve diagnosis, and treatment assessment, using ultrasound-guided dif-
robustness of diffuse optical tomography for breast cancer imaging,” fuse optical tomography, photoacoustic imaging, and optical coherent
Biomed. Opt. Express 7(10), 4007–4020 (2016). tomography.

Journal of Biomedical Optics 056005-8 May 2019 • Vol. 24(5)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Biomedical-Optics on 09 May 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use