2.1.

MRI Imaging Studies

A 3T Philips MR scanner was used to acquire T1-weighted and DCE-MR scans. 2-D T1-weighted Spin Echo (TR/TE = 900/10, flip angle = 90 deg) images were obtained for all subjects. For the subject with abnormality, regions of interest, determined by DCE-MR, were recorded by injecting a bolus of contrast agent (Magnavist) intravenously and taking a series of 3-D T1-W volume images (TR/TE = 10/6, flip angle 20 deg) after each minute, beginning 40 s postinjection.

2.2.

NIR Imaging Studies

The NIR instrumentation is shown in Fig. 1 and used 16 fibers to sequentially transmit light from each of multiple laser sources and deliver the measured signal to a bank of photomultiplier tube detectors. Light from six individual diode lasers (wavelength ranges of 660–850 nm) was sequentially delivered through the fibers to the breast with an intensity modulation frequency of 100 MHz. The amplitude and phase shift of the transmitted light were recorded at multiple locations around the periphery of the breast. These frequency domain measurements allow separation of absorption and scatter processes during image reconstruction. The fiber-optic cables were nonmagnetic and integrated with a MR breast coil, allowing NIR data acquisition while the scanner was in operation. The patient was positioned prone inside the MRI (3T Philips Achieva, X-series) with the breast pendant into the fiber-holder interface. Although much of our initial work positioned the fibers in a circular geometry,²⁹ a version of the system that conforms to the standard parallel-plate biopsy geometry²⁷ is also available to facilitate clinical workflow. This change in geometry does not significantly alter the NIR imaging performance.

Fig. 1

(a) Biopsy plate slab interface with slots for fibers, (b) circular breast interface containing optical fibers shown positioned in the MR breast coil, and (c) integrated mobile cart containing the NIR data collection instrumentation.

2.3.

Image Segmentation and Breast Surface Rendering

The MR images were segmented into adipose and fibroglandular (FG) tissues using the T1-weighted MR acquisitions based on thresholding and region-growing operations applied to the gray-scale intensities. The image segmentation was accomplished using a commercial software package (Mimics^TM, Materialise Inc.). An example of this process is shown in Fig. 2. In the subject with cancer, the corresponding DCE-MRI was used to segment the lesion. The segmented tissues were surface rendered using the software and exported to apply diffusion light modeling for image reconstruction.

Fig. 2

(a) Sample MR coronal image of a healthy subject, (b) thresholding results in a binary mask delineating the FG tissue from the rest of the breast shown in purple, and (c) binary mask delineating the breast from the background. Surface renderings for the breast outer contour and FG tissue from 3-D segmentation are shown in (d). This geometry was directly used in the BEM toolbox for NIR image reconstruction.

2.4.

BEM-Based Multispectral Image Reconstruction

The diffusion approximation to radiative transport theory has been successfully used in the past to model light propagation in highly scattering media.^{30, 31, 32} The BEM was applied to the diffusion approximation under the assumption that the tissue being imaged contains homogeneous regions whose boundaries are known a priori (from MRI).^{28, 33} It reduces the dimensionality of the problem, requiring only surface rather than volume discretization in three dimension, which is significantly easier and more reliable to generate. Using this numerical model to solve the diffusion equation, unknown NIRS parameters can be reconstructed from boundary measurement of light amplitude and phase. This inverse problem was solved iteratively using a modified Newton's method and Levenberg–Marquardt regularization.³⁴

By applying a BEM model for the diffusion equation, average region-based IG-NIRS estimates of HbT, oxygen saturation, water, and scatter were obtained for each of the constituent tissues in the breast. Briefly, the acquired boundary measurements of light amplitude and phase underwent a calibration procedure to compensate for system offsets due to source–detector fiber transmission, alignment characteristics, and errors in discretization or model–data mismatch.³⁵ This calibration procedure, also called homogeneous fitting, used a two-step procedure to estimate the slopes of the log of light intensity and phase with distance r using (i) analytical solution to an infinite medium and (ii) Newton–Raphson fitting with a homogeneous BEM model. Postcalibration, the clinical image reconstructions were accomplished using a multispectral direct chromophore and scatter recovery algorithm that has been demonstrated to reduce cross-coupling between parameters and provide the most accurate optical estimates of tissue composition.^{36, 37} In this reconstruction, concentrations of oxyhemoglobin (HbO), deoxyhemoglobin (Hb), water, and scatter were obtained, and extended indices such as [HbT] = [Hb] + [HbO], and oxygen saturation S_tO₂ = [HbO]/[HbT] in percent were calculated. The starting values for the reconstruction were obtained from the homogeneous fitting procedure, except for water whose starting value was set at 100%. This recovered the most accurate water content in simulations and experiments. The reconstruction constrained water content to be within the range 0–100%.

2.5.

Patients

The clinical enrollment consisted of seven patients of whom six were healthy subjects with no known abnormality, and one was diagnosed with infiltrating ductal carcinoma (IDC). All subjects provided informed consent and were imaged according to study protocols approved by the Institutional Review Board at Dartmouth. NIR measurements were acquired simultaneously with MRI acquisition in all subjects with no additional increase in imaging time. Up to six wavelengths were applied during the NIR exam. The MRI of the subjects was used to provide anatomical tissue structure for optical image reconstruction.

The subject with cancer was a 36-year-old female volunteer diagnosed with a 3-cm, intermediate-grade, triple-negative (for estrogen, progesterone, and HER2/neu expression), node-positive IDC with three satellite malignancies in her left breast. She underwent neoadjuvant chemotherapy, and her chemotherapy regimen consisted of six cycles of docetaxol/adriamycin/cylcophosphamide. The subject had a complete pathological response subsequent to chemotherapy as confirmed by histology. Multiple imaging sessions of T1-weighted and DCE-MRI and NIR measurements were performed over the course of treatment. The first imaging session occurred 1 day prior to her first cycle of chemotherapy (cycle 1) in the circular geometry. Imaging sessions thereafter were performed within 48 h of chemotherapy for cycles 2 and 4 and postchemotherapy, prior to surgery. Out of these four imaging sessions, the plane of the NIR fibers in session 1 was offset from the lesion sampled in sessions 2 and 3 by 5 mm. The plane of fibers was consistent in sessions 2 and 3, and optical data from these visits (labeled 1 and 2 in results) were used to reconstruct the IG-NIRS estimates for background and tumor tissues. In session 4, the tumor was no longer visible in the MRI and hence is not shown here.

3.1.

Experimental Validation in a Breast-Shaped Phantom

A gelatin phantom was created using a breast mold fabricated from a segmented breast MR volume. Pig blood was added to the phantom during preparation to cause absorption due to hemoglobin and titanium dioxide (TiO₂) was added to introduce scattering.³⁸ By measuring the hematocrit of the blood and converting to HbT, the mixture represented a 14-μM concentration. A cylindrical tube of 30 mm diam filled with a 26-μM solution was inserted into the gelatin to create an inclusion that presented a total hemoglobin contrast of ∼2 : 1. A photograph of the phantom is shown in Fig. 3. A homogeneous cylindrical phantom was also made from the same gelatin mixture and utilized for data calibration. The breast phantom with inclusion was imaged optically using six wavelengths, and MR images of the phantom were acquired with fiducial markers to denote fiber locations; a cross section is shown in Fig. 3. The MRI was segmented to create 3-D surfaces of the breast and inclusion as shown in Fig. 3. These surfaces were used along with optical frequency domain measurements to reconstruct the NIRS parameters for the background and the inclusion. The recovered values in the inclusion are compared to the expected values in Fig. 3. The results show that the error in recovered HbT was <3%, illustrating that quantitative characterization is possible with BEM based on experimental measurements. The error in oxygen saturation was 6.3%.

Fig. 3

(a) Breast phantom imaged in the MRI-NIR system is shown with a representative MR slice in (b) containing fiducial markers indicating the positioning of the fibers (in red) and (c) segmented background and inclusion surfaces. These surfaces were used to reconstruct the NIRS parameters of the phantom shown in (d) with comparison to expected values. Water was reconstructed with 100% accuracy in this case, because it was the maximum allowable by the reconstruction.

3.2.

Healthy Breast Tissue Characterization of Adipose and FG Tissue Types

For healthy breast characterization, the MRI of each subject was segmented into its adipose and FG tissue structures, and then surface rendered using Mimics^TM. These surfaces along with a representative MR image for these subjects are shown in Fig. 4. They were input as the geometry for IG-NIRS image reconstruction. All NIR parameters were updated during the iterative reconstruction procedure. Using the BEM spectral reconstruction with starting values obtained from calibration and the measured data, IG-NIRS estimates were recovered for adipose and FG tissues of the breast. The reconstructed results are shown in Fig. 4. The data show that the average HbT in FG tissue is higher than that of adipose tissue. A similar trend was observed in water content with FG tissue containing higher water, when compared to adipose tissue. Oxygen saturation was slightly increased in FG (72% ± 17%) compared to adipose (66% ± 12%) tissue. Statistical analysis using paired t-test showed a p-value of 0.062 for difference in HbT, p-value of 0.34 for oxygen saturation and p-value of 0.054 for difference in water. The p-values for HbT and water are close to being statistically significant (p-value = 0.05); however due to the limited sample size and variation in number of wavelengths used, it is difficult to draw conclusions from this population. In the individual subjects, every subject showed higher HbT and water in FG as compared to adipose tissue. Scattering estimates were also reconstructed, but an insufficient number of wavelengths were available to determine their accuracy.

Fig. 4

(a) Representative MRI coronal slices are shown for healthy subjects 1–3 and (b) corresponding surfaces for adipose and fibroglandular (FG) tissues obtained from MR image segmentation. (c) Same as (a) for healthy subjects 4–6 imaged in the slab-geometry and (d) corresponding surfaces. Reconstructed estimates of adipose and FG are shown in (e) the bar graph with standard deviation for these six subjects.

The reconstructions converged in four to six iterations, and convergence was deemed when the change in projection error (least-squares difference between measured and model data) was <0.5% between successive iterations. The normalized projection error did not drop below 0.5 for clinical data, probably because of increased noise in the measurements and the greater approximation in the data-model match in vivo. The experimental projection error in phantoms dropped to 0.25 in comparison.

3.3.

Imaging Subject with Cancer

The T1-weighted MRI from each session was used to segment the adipose and FG tissues and the corresponding DCE-MRI was used to segment the lesions in a semi-automatic way with guidance from radiologist. The segmented surfaces are displayed in Fig. 5 along with the imaging plane denoted by the fiber locations (in red).

Fig. 5

Surface renderings of adipose, FG, and lesions segmented from T1-weighted MRI and DCE-MRI of subject 7 undergoing NACT are shown for (a) visit 1 corresponding to cycle 2 and (b) visit 2 corresponding to cycle 4. The red markers indicate optical fiber locations. (c) Reconstructed IG-NIRS values are displayed for HbT in adipose, FG, and tumor tissues for both visits. The subject experienced a pathologic complete response.

Following data calibration, the IG-NIRS estimates were reconstructed for adipose, FG, and tumor tissue and are presented in Fig. 5 for HbT. The results show a reduction in the tumor hemoglobin from visit 1 to visit 2. The tumor HbT was higher than the HbT of background tissues (adipose and FG) as well as higher than the averages from the healthy subjects. A reduction in the adipose and FG HbT was also observed in visit 2 relative to visit 1, indicating that both the background and tumor vasculature have undergone change during treatment. Oxygen saturation was found to decrease in all tissues, and no noticeable trend was observed in water content.