1. Introduction

Breast cancer is the most common cancer among women worldwide, with an estimated one in eight women receiving a breast cancer diagnosis in her lifetime and 287,850 new cases and 43,250 resulting deaths expected in the US in 2022 [1]. Though advances in screening technology have contributed significantly to a steady decrease in mortality from breast cancer in recent decades, current clinical breast imaging methods suffer limitations due to their reliance on structural imaging [2]. While sensitive in detecting subtle anatomical abnormalities, they lack specificity when differentiating benign lesions from malignant tumors, leading to unnecessary biopsies and patient emotional burden [3]. In this context, diffuse optical tomography (DOT) offers an opportunity to complement clinical breast imaging with hemoglobin-based functional imaging information [4,5]. Driven by the need to support sustained growth, angiogenesis develops as a hallmark of invasive solid tumors once they grow beyond a few millimeters in size to provide cancer cells with oxygen and nutrients through vast but dysfunctional blood vessel networks [6–8]. Using multi-wavelength near-infrared systems with an array of source and detector locations, DOT can quantitatively image the distribution of oxy- and deoxy-hemoglobin concentrations within the breast tissue, providing valuable insights into tissue perfusion and metabolism. Various independent clinical studies have demonstrated that DOT-derived imaging markers, including total hemoglobin concentration (HbT), are effective not only in distinguishing benign and malignant lesions [9–14] but also in predicting ultimate neoadjuvant chemotherapy outcome based on early tumor response [15–17].

While offering a novel form of functional imaging contrast, DOT has limitations with respect to image resolution and tumor contrast recovery. Due to the diffuse nature of near-infrared light propagation in tissue, the inverse problem is underdetermined and ill-posed, requiring regularization that inevitably reduces the amount of tumor contrast recovered [18]. To address these limitations, structural priors based on established forms of breast imaging, e.g., x-ray mammography, magnetic resonance imaging (MRI), and breast ultrasound, are increasingly being integrated into the image reconstruction process, and their use has been shown to provide images with improved detail and more accurate quantification of tumor chromophores [19–21]. However, prior-guided image reconstruction requires accurate cross-modality registration, best achieved by employing multimodal imaging systems that maintain the patient’s breast in a fixed position while data from both modalities are acquired. This is not always possible due to logistical or instrumentation limitations. For example, from a regulatory standpoint, a standalone optical breast imaging device only needs to secure approval once, while multimodal combinations with mammography and MRI need to be recertified if either device changes, potentially requiring the cooperation of multiple unrelated commercial entities.

Our group has pursued the combination of DOT with digital breast tomosynthesis (DBT), the three-dimensional (3D) form of x-ray mammography, for the last two decades. We have previously demonstrated a simultaneous co-registered operation where glass fibers from the source and detector systems are coupled to plastic fibers for the final segment where the optical probes overlap the x-ray field, allowing for simultaneously co-registered DOT-DBT imaging [22]. However, the plastic fibers degraded more quickly than expected, likely due to frequent x-ray exposure, and had to be replaced by an all-glass fiber probe. Since glass fibers cast significant artifacts on mammographic images, separate compressions are needed – first with the optical probes mounted for DOT data acquisition, and then with the optical probes removed for DBT imaging, with patient repositioning in between. To allow prior-guided DOT reconstruction, we employed a contour-based post-acquisition image registration [23] to apply DBT anatomical structures and radiologic markings to the DOT image space. In this scenario, relatively small differences in positioning of the breast could result in inaccurate placement of lesion prior. While lesion priors have been shown to be robust against incorrect placement, showing no contrast enhancement when placed at a location misaligned with the true lesion region [19], the benefits of prior-guided reconstruction are not effectively realized without precise prior location for accurate quantitative imaging.

Previously, Fang has described a method [24] that leverages the robustness of prior-guided reconstruction against discordant priors to automatically recover the location of unknown tissue abnormalities by placing hypothesized tumor regions of interest (ROIs) at all possible locations of the breast on a search grid in prior-guided DOT reconstruction. Using custom-defined metric functions, Inácio et al. have demonstrated that this method can successfully identify small tumors at true lesion locations marked in simultaneously co-registered DBT images [25], and Xu et al. have further validated that this method can successfully localize and classify 93% of 29 malignant cases with over 10-fold enhancement in optical contrast, as well as identify 98% of 68 tumor-negative breasts, in a group analysis [26]. Moreover, Xu et al. have explored the utility of this method in determining the orientation and shape of unknown tumors using anisotropic priors in a simulation study [27].

Inspired by prior work and motivated by the need to establish a robust algorithm for accurate lesion prior placement to enable effective non-simultaneous multimodal DBT-DOT imaging when the optical data are acquired separately from the DBT scan, we present a multi-stage lesion scanning algorithm guided by HbT contrast. This method starts from using the contour-registration-based lesion location as guidance for initial placement of lesion prior, followed by systematic search steps both within the whole breast area in sparsely placed axial planes, then by a full breast thickness search in a focused scanning region, and concludes with additional 3D fine-tuning steps, with each step showing progressively improved tumor contrast. We demonstrate the application of this technique in three challenging patient cases where no clear optical contrast was initially shown at the expected tumor locations and demonstrate a significant improvement in the final reconstructed HbT images after utilizing the proposed lesion scanning approach. We believe this method can greatly improve the recovered tumor contrast in DOT systems where structural images are either absent or not natively co-registered with DOT images and could potentially improve the consistency of recovered tumor contrast in longitudinal imaging studies.

2. Methods

2.1 Imaging procedures

Our second-generation tomographic optical breast imaging (TOBI2) system has been described in detail previously [22]. Briefly, the system comprises 96 continuous-wave (CW) source locations evenly split between 690 nm and 830 nm wavelengths, 24 radio-frequency (RF) source locations emitting both 690 nm and 830 nm light, and 32 CW and 20 RF detectors, used for both wavelengths. The DOT source and detector plates, housing the 120 source glass fibers and 52 detector fiber bundles, are designed to be mounted on the compression paddle and x-ray detector cover, respectively, of a Hologic Selenia Dimensions DBT device to achieve mammography-like compressions. To acquire the accurate breast shape needed for the optical image reconstruction and contour-based image registration, three web cameras, shown in Fig. 1, were added to capture breast shapes from two top views and one side view.

 

Fig. 1. Optical probes of the TOBI2 imaging system mounted on a Hologic DBT device. Two top-view webcams are circled in green and cyan. The side-view webcam (not visible from this angle) is circled in purple, with an insert showing its positioning.

Download Full Size | PDF

During the study imaging session, with the optical probes attached to the DBT machine, the patient’s breast was compressed in the cranial-caudal (CC) view to half mammographic compression, and one minute of optical data was acquired. The optical acquisition is part of a multi-step optical imaging protocol that comprises three 1-min periods at half, full, and then again at half mammographic compression levels, respectively, as previously described in Ref. [22]. We chose to use data from only the first compression period in this line of investigation as tumor contrasts are typically higher under less compression and the tumor location in relative coordinates does not change at other compression periods. Upon completion of optical data acquisition, the three webcams mounted on the source probe took images of the top and side contours of the breast, from which tetrahedral breast meshes were generated for optical image reconstruction after contour tracing in the horizontal and vertical planes. The patient was then released from compression, the optical probe removed, and the patient repositioned for a separate DBT imaging at full mammographic compression in the same CC view as routine clinical mammograms. The mammographer, who also conducted the optical data acquisition, used her professional judgment to position the breast as consistently as possible between imaging modalities, but some shifts and inelastic deformations likely occurred. The breast lesion was later marked on the DBT image stack by the breast radiologist (M.S.) associated with the study.

All patients were enrolled under a pilot study that aims to use combined DOT and DBT imaging for the early detection of the pathological outcome of neoadjuvant chemotherapy (NCT03822312). Thus, all patients have invasive breast cancers that were imaged longitudinally, and only the baseline scans that were taken before the initiation of the treatment are presented here to demonstrate the effectiveness of the proposed lesion prior scanning method in accurately assessing tumor contrast. Written consents were obtained in accordance with the policies and guidelines of the Massachusetts General Hospital Institutional Review Board. A total of 20 patients were consented to the pilot study, of whom 4 were excluded from group analysis due to missing data. The 3 patient cases shown in the paper are selected among the most challenging ones in this 16-patient cohort.

2.2 Prior-guided optical image reconstruction

Optical image reconstructions were performed on a pair of tetrahedral meshes – a finer one for solving the optical forward problem and a coarser one for solving the inversion – generated using the MATLAB-based meshing toolbox “iso2mesh” [28] on the breast shape created from webcam images. These meshes were extended by 2 cm into the chest wall region to avoid boundary effect distortions in the main imaging area [29]. Raw optical measurements were first calibrated against a homogenous phantom with known optical properties and then fitted for bulk optical properties. Nonlinear, spectrally constrained inversion of the finite-element representation of the diffusion approximation using the Tikhonov-regulated Gauss-Newton approach was performed for ten iterations using the in-house software, i.e., Redbird, to reconstruct optical images shown in this paper [29,30].

In order to use lesion shape and location for prior-guided reconstructions, the DBT breast outline at each 1-mm slice was first registered to the DOT mesh contour at proportional depths by an affine transformation, as described in detail in our prior work [23]. The lesion markings on DBT slices were then transformed to the DOT imaging space through the contour-based registration and their outlines and 3D centroids were extracted. A two-composition structural prior comprising the probabilities of adipose and fibroglandular tissue at each tetrahedral node was generated based on the DBT image stack using a histogram-based segmentation method described previously [19,31]. A three-composition structural prior, including an additional lesion probability, was also generated (the lesion location and shape varied depending on the specific step of the overall algorithm, as described further below). When solving the inverse problem, two- or three-composition structural priors were used as soft constraints in our prior-guided reconstruction algorithm described previously [19,29]. It is worthwhile to mention that this initial cross-modality registration is not a prerequisite to use the lesion scanning approach detailed below. In a standalone DOT setup, image reconstruction guided only by the lesion prior can still be performed at the location identified by the scanning method to achieve a similar enhancement of tumor contrast.

2.3 Lesion scanning approach

To account for both in-plane translation and out-of-plane rotation between the separate DBT and DOT compressions, the lesion scanning approach proposed herein consisted of three steps. As shown in Table 1, the first step aimed to search for the approximate lesion location in the XY-plane and coarse location in the Z direction by generating a 5-mm step grid over the entire breast at three equally spaced Z depths (25, 50, and 75% of the overall breast thickness). After an appropriate XY centroid was identified based on locations where the tumor prior recovered significant HbT contrast, the second step primarily focused on finding the approximate lesion location in Z by going through the entire depth of the breast within a smaller 4 cm × 4 cm area around the approximate location in the XY-plane. The final step further fine-tuned the overall 3D localization by searching within a focused 2 cm × 2 cm × 2 cm cube with a finer 2.5-mm grid step in all three dimensions. This step could be repeated more than once if necessary. In the following subsections, a sample case of a breast cancer patient will be used to illustrate the lesion scanning approach in detail. The patient was diagnosed with a triple-negative (TN) invasive ductal carcinoma (IDC) measuring 31 mm × 25 mm × 23 mm in her left breast.

Table 1. Scanning Parameters at Each Step of the Lesion Scanning Approach

View Table

2.3.1 Step 1–whole-breast scanning

After the generation of the forward model tetrahedral mesh, new nodes were added to the mesh in a 5-mm X-Y grid at three depths corresponding to 25%, 50%, and 75% of the full mesh thickness, as shown by the red dots in Fig. 2, to facilitate specifying potential locations with tumor contrast. These “scanning nodes” were added at locations that excluded the extended chest wall region from 0 mm to -20 mm on the Y-axis, and the region within 1 cm of the surface nodes. At each of these grid positions, a spherical compositional lesion prior was generated using a Gaussian probability profile with a full-width half-maximum of 1 cm. Three-composition prior-guided reconstructions were run in parallel at each individual prior location on a 1,200-core shared computing cluster equipped with Intel Xeon Gold 6226R CPUs. For this sample case, which has a typical breast coverage and thickness, a total of 914 nodes were added for the whole-breast scanning step, and each reconstruction job took about 40 minutes to complete on a single CPU core, i.e., ∼ 600 single-threaded computing hours. On a single contemporary high-end workstation with 32 CPU cores, this corresponds to 18-20 total hours.

 

Fig. 2. Whole-breast grid at three depths used for the first lesion scanning step. (a) Side view that shows three scanning depths at 25%, 50%, and 75% of the full breast mesh thickness. (b) Top view that shows the 5-mm grid in the XY-plane overlayed on the original mesh. Gray dots: the original forward mesh nodes; Red dots: the additional lesion scanning nodes.

Download Full Size | PDF

After DOT image reconstructions at all scanning positions were completed, two metrics were calculated based on the recovered HbT images of each reconstruction:

The resulting metrics were plotted in a map corresponding to the prior location in the breast similarly as described in Ref. [32] and smoothed using a 3 × 3 square mean filtering kernel to avoid bias towards hot spots. The metrics maps for the sample patient using the above two metrics at three depths are shown in Fig. 3.

 

Fig. 3. Metrics maps for a breast cancer patient derived from the HbT images at each grid point of the whole-breast scanning step. The top (a-c) and bottom (d-f) rows show maps of the weighted HbT difference metric and the HbT ratio metric, respectively, at three depths. White outline: outermost breast contour of the forward mesh. Red line: transferred DBT lesion marking using the contour-based registration. Thick white line and cross: hand-drawn polygon around the high values on the metrics map and its center, where the lesion center would be shifted to.

Download Full Size | PDF

The most likely lesion location was then chosen manually based on the panel of six whole-breast metrics maps. Specifically, a region of relatively high metrics values was sought within the vicinity of the original DBT marking, accounting for a feasible level of changes in lesion location due to repositioning (both in XY and in depth, based on the DBT slice range where the lesion was visible). Once identified, a region of interest (ROI) was drawn and its centroid calculated, to which the original DBT lesion marking would be shifted. For example, as shown in Fig. 3(a) and 3(d), the region circled in a hand-drawn polygon was likely closer to where the true lesion was located based on the metrics maps at the depth of 17 mm. The newly determined lesion centroid was then used as the starting position for the subsequent lesion scanning steps.

2.3.2 Step 2–full-depth scanning

After shifting the original radiologic lesion marking to the centroid determined in Step 1, new scanning lesion priors were generated in 5-mm increments over a 4-cm grid area centered around this new lesion centroid in the X and Y directions, i.e., creating a 9 × 9 grid in the XY-plane, and in 5% thickness increments in the Z direction starting from 2.5% of the breast thickness and going up to 97.5% of the overall breast thickness (20 depths total). Instead of using the spherical Gaussian profile, the lesion priors used in Step 2, as well as subsequent steps, were generated using the true shape extracted from the DBT markings extended 5 mm in both Z+ and Z- directions centered at each scanning depth. Three-composition prior-guided reconstructions were again run at each of these new scanning locations. For the patient sample case, a total of 1,620 (81 in-plane grids × 20 depths) reconstruction jobs were done in parallel on a computing cluster, corresponding to 1,080 single-threaded computing hours. As in the previous step, upon completion of all jobs, a metrics map was created at each depth. However, for Step 2 and subsequent scanning steps, the mean absolute HbT values within the lesion region (${\overline {HbT} _{roi,3 - comp}}$) defined at each scanning location were used as metrics for determining the new lesion location. ${\overline {HbT} _{roi,3 - comp}}$ was calculated by averaging the HbT values on all lesion nodes defined in the prior used for each reconstruction, excluding those on the surface, within 1 cm from the top or bottom of the mesh, or within the extended chest wall region.

Figure 4 shows the absolute HbT metrics map at 20 depths in the sample case. At each depth, the average of metrics values at all XY-plane grid positions were calculated (shown in the text within each subplot) to guide the selection of the optimal depth. Typically, the optimal depth is chosen at the center of the z stack range showing elevated average metrics values. The lesion location in the XY-plane could also be adjusted based on the pattern of the metrics map. For example, in this patient case, clear high values (over 20 µM, or twice more than the rest) on the metrics maps showed up at multiple depths close to the bottom of the breast (spanning from z = 8.2 mm to 25.3 mm) with centers positioned somewhat towards the upper left. Based on the averaged HbT metrics at each depth, the selected lesion depth was 15.0 mm, which is centered in the z stack ranging from 8.2 mm to 25.3 mm in depth, and the XY location was shifted 1 step (5 mm) in each of the X and Y directions. Note that there was no valid metric value at the 1st depth as all lesion prior nodes fell within the regions excluded from metrics calculations.

 

Fig. 4. Metrics maps at various depths covering the full breast thickness showing the mean absolute HbT in the lesion region at each grid position. Depth and the averaged metric values across all grid positions at each depth are noted in the text within each subplot. Red cross at depth z = 15.0 mm: newly selected lesion center based on the full-depth scanning results.

Download Full Size | PDF

2.3.3 Step 3–3D refinement

To further refine the 3D lesion location, another round of scanning was done with a smaller grid step of 2.5 mm within a 2 × 2 × 2 cm³ cube, i.e., a 9 × 9 × 9 grid, centered around the lesion centroid determined in the previous full-depth scanning step. It takes about 486 single-threaded computing hours on our computing cluster CPUs to complete the 729 reconstructions for one iteration in Step 3. Typically, the central location in the highest contrast region at the depth of highest mean HbT was chosen as the final lesion centroid. It was expected that this would appear close to the center of the 9 × 9 grid on the XY-plane at the 5th depth, i.e., the central depth, assuming the location chosen in Step 2 is accurate. If it appeared near the edge of the scanning grid, this step could be repeated until the higher contrast region was better centered.

Figure 5 shows the 9 × 9 absolute HbT metrics map at 9 depths of the patient case. Based on the averaged HbT values of all grid positions at each depth (shown in the text within each subplot), the 6^th depth has the highest HbT contrast of the nine depths and the highest HbT value at this depth is centered. Thus, one iteration of the 3D refinement step was sufficient for this case, and the final lesion location was chosen in the position shown by a red cross in Fig. 5. Of note, for the vast majority of patient cases (15 out of 16) processed with the lesion scanning approach, the 3D refinement step was performed a maximum of two times (7 cases only needed one iteration), demonstrating the efficiency of the previous whole-breast and full-depth scanning steps in finding the proximity of the appropriate lesion location.

 

Fig. 5. Metrics maps of the refinement step covering a smaller 2 cm × 2 cm × 2 cm region showing the mean absolute HbT in the lesion region at each grid position. Depth and the averaged metric values across all grid positions at each depth are noted in the text within each subplot. Red cross: newly selected lesion center based on the refinement scanning maps.

Download Full Size | PDF

3. Results

By following the aforementioned lesion scanning method, we were able to significantly increase recovered HbT values in the lesion region by placing the lesion prior in a more accurate location. For the sample patient case used in the Methods section to explain the lesion scanning approach, i.e., the 67-year-old female with a 31 mm × 25 mm × 23 mm TN IDC in her left breast, reconstructed HbT images following each of the three lesion scanning steps are shown in Fig. 6. As shown in Fig. 6(b), the HbT image reconstructed without structural prior does not show any notable contrast in the expected lesion region. Moreover, placement of the lesion prior at the initial contour-registration-based location for a 3-composition prior-guided reconstruction results in a small negative HbT contrast, as is shown in Fig. 6(c), suggesting that the lesion prior was inaccurately located. Following an XY shift of ∼21 mm towards the medial chest wall (lower left in the orientation shown) based on the Step 1 whole-breast lesion scanning result, as shown in Fig. 6(d), the negative HbT contrast is no longer present, but the lesion HbT is still below that of the background. However, with the guidance of the Step 2 full-depth lesion scanning, the lesion prior is moved from 40.6 mm from the bottom surface of the breast to a significantly different depth of 15.0 mm (consistent with a rotation of the breast during re-positioning), as well as a further slight adjustment in XY, resulting in significantly improved lesion contrast as seen in Fig. 6(e). A small additional increase in HbT contrast is achieved upon completing the Step 3 2.5-mm 3D refinement step, as shown in Fig. 6(f). The final 3D lesion location of this case is shifted 4.3 mm in X, 19.5 mm in Y, and 22.1 mm in Z from the initial contour-registration-based location, with a final lesion HbT approximately 4 times higher than background HbT.

 

Fig. 6. Recovered HbT images of a 67-year-old breast cancer patient with a TN IDC in her left breast. (a) DBT image slice marked by radiologist. DOT reconstructions based on (b) no prior (unguided), and 3-composition priors with the location of the lesion prior determined by (c) the initial contour-based registration, (d) Step 1 whole-breast lesion scanning (white cross in Fig. 3(a)), (e) Step 2 full-depth lesion scanning (red cross in Fig. 4), and (f) Step 3 refinement (red cross in Fig. 5). Solid red line in (c-f): lesion prior location for each 3-composition prior-guided reconstruction. Dotted white line in (b, d-f): reference lesion location determined by the initial contour-based image registration between DBT and DOT. Depth, mean HbT within the lesion ROI and the background are noted in each subplot.

Download Full Size | PDF

Figure 7 shows a similar summary of results of the lesion scanning method at each step for a 68-year-old female with a 19 mm × 13 mm × 12 mm TN IDC in her right breast. As in as the first patient case, no clear contrast is observed in the HbT image using unguided optical image reconstruction as shown in Fig. 7(b). When the initial contour-registration-based lesion location is used in a 3-composition prior-guided reconstruction, an inverted HbT contrast is again observed in Fig. 7(c). Based on the Step 1 whole-breast lesion scanning result, the lesion location is shifted about 7.7 mm upward (laterally), resulting in a significant increase of the mean HbT value within the lesion ROI as shown in Fig. 7(d). Another notable boost in lesion HbT is further achieved, as shown in Fig. 7(e), when the lesion prior is placed at depth 12.9 mm instead of 21.7 mm and shifted 8.2 mm towards the chest wall based on the newly determined lesion location from the Step 2 full-depth lesion scanning. The fine adjustment scanning for the final 3D lesion location resulted in even higher lesion contrast, shown in Fig. 7(f). At the conclusion of all the three steps of lesion scanning procedure, for this second patient case, the final 3D lesion location is shifted 5.5 mm in X, 10.7 mm in Y, and 10.3 mm in Z from the initial contour-registration-based location with a final lesion HbT approximately 2.3 times higher than background HbT.

 

Fig. 7. Recovered HbT images of 68-year-old woman diagnosed of a 19 mm × 13 mm × 12 mm TN IDC in her right breast. (a) DBT image slice marked by radiologist. DOT reconstructions based on (b) no prior (unguided), and 3-composition priors with the location of the lesion prior determined by (c) the initial contour-based registration, (d) Step 1 whole-breast lesion scanning, (e) Step 2 full-depth lesion scanning, and (f) Step 3 refinement. Solid red line in (c-f): lesion priors placed for each 3-composition prior-guided reconstruction. Dotted white line in (b, d-f): reference lesion location determined by the initial contour-based image registration between DBT and DOT. Depth, mean HbT within the lesion ROI and the background are noted in each subplot.

Download Full Size | PDF

Figure 8 shows another panel of images for a third patient case of a 42-year-old female diagnosed with multi-focal TN IDC disease, with the primary tumor measuring 35 mm × 18 mm × 15 mm, in her left breast. As in as the other patient cases, no clear contrast is observed in the HbT image using unguided optical image reconstruction (Fig. 8(b)), but a low level of contrast is recovered in when using a lesion prior based on contour registration for reconstruction (Fig. 8(c)). Based on the whole-breast scanning metrics maps, the lesion is shifted 16.2 mm in the medial direction and 3.6 mm further away from the chest wall. As shown in Fig. 8(d), this change has actually resulted in a slight loss of contrast recovery, possibly due to the inaccuracy of using a single Gaussian spherical prior in this case with two separate lesion areas. After using the true lesion shape for prior generation in the Step 2 full-depth scanning, the loss of contrast is corrected by shifting 10.0 mm in X and 5.0 mm in Y back towards the original lesion location, as well as an 18.5 mm shift in Z. The resulting HbT image shown in Fig. 8(e) shows significantly enhanced contrast compared to the background. After Step 3 refinement, the lesion location is shifted 2.5 mm (1 grid step) closer to the chest wall, resulting a small further increase in HbT contrast recovery. The final location is shifted 6.2 mm in X, 4.1 mm in Y, and 18.5 mm in Z from the initial contour-registration-based location. Following all the lesion scanning steps, the recovered lesion HbT is approximately 4 times that of the background HbT.

 

Fig. 8. Recovered HbT images of a 42-year-old woman diagnosed with a multi-focal TN IDC breast cancer in her left breast. (a) DBT image slice marked by radiologist. DOT reconstructions based on (b) no prior (unguided), and 3-composition priors with the location of the lesion prior determined by (c) the initial contour-based registration, (d) Step 1 whole-breast lesion scanning, (e) Step 2 full-depth lesion scanning, and (f) Step 3 refinement. Solid red line in (c-f): lesion priors placed for each 3-composition prior-guided reconstruction. Dotted white line in (b, d-f): reference lesion location determined by the initial contour-based image registration between DBT and DOT. Depth, mean HbT within the lesion ROI and the background are noted in each subplot.

Download Full Size | PDF

4. Discussions and conclusions

Here we have presented a multi-step lesion scanning method capable of searching for an accurate 3D location to place a lesion prior for the 3-composition prior-guided DOT image reconstruction in cases where optical and structural imaging are not acquired simultaneously, as is the case of our DOT-DBT multimodal imaging study. For Step 1 (whole-breast scanning), two HbT-based metrics are used to help us identify an approximate XY location near the lesion marking registered to the DOT image space through a contour-based affine registration. From Fig. 3, it is evident that the overall patterns are quite similar when comparing the maps generated by the two different metrics. Both metrics have successfully identified the same enhanced region in the example case. This consistency has been observed in most patient images processed through the lesion scanning approach, reassuring our ROI selection. In some specific scenarios, each metric has its unique advantage. For example, in cases where 2-composition prior-guided reconstruction have already recovered notable amount of lesion contrast, the weighted difference metric may result in a less prominent contrast on these metrics maps compared to the ratio metrics maps. Since the ratio metric only extracts values from the 3-composition prior-guided reconstruction results, it is more likely to be biased by non-tumor regions that also have relatively higher HbT values than the bulk of the breast, especially the chest wall region as shown in the second row of Fig. 3. Moreover, in cases where the lesion size is small, the ratio metric is more likely to result in unstable hot spots on the metrics maps than the weighted difference metric. Provided with smoothed maps using both metrics, the selection of new ROI for downstream processing, though somewhat subjective, is well-informed.

The correctness of the selected ROI in the whole-breast scanning can be further validated in the next full-depth scanning step. As shown in Fig. 4, high HbT values centered within a well-defined shape can be seen in all depths starting as early as 4.8 mm to as deep as 32.1 mm. Moreover, the averaged HbT metrics (shown in the text within each subplot) at each depth have a clear peak among these depths, and the highest HbT within each of the 9 × 9 grid appears in the nearly identical location. These are strong indicators that the ROI selection is heading in the right direction. Similarly, the validation of ROI selection in the full-depth scanning can be confirmed in the final 3D refinement step with a fine 2.5-mm grid size. If previous steps have effectively pinpointed the correct tumor location, the highest HbT metric value is expected to be seen at or near the 5th scanning depth centered in the 9 × 9 grid. This is what we have found in the example case shown in Fig. 5.

The cases picked as examples to show in the Results section are particularly challenging, as no contrast was seen initially at the expected lesion location in either the standard unguided (subplots b in Figs. 6–8) or the prior-guided DOT image reconstruction (subplots c in Figs. 6–8). This is not necessarily representative of the majority of our DOT measurements – generally contrast has been seen in the vicinity (but rarely the exact location) of the co-registered tumor prior in 12 out of 16 patient cases enrolled in our pilot therapy monitoring study. However, we want to demonstrate that even in these challenging cases, following the application of our lesion scanning method, a marked improvement in the recovered tumor HbT contrast relative to the background can be achieved. Frequently, the most significant boost in contrast was seen following full-depth scanning, at which point the lesion location had been roughly found. Understandably, the final refinement step has usually achieved incremental improvements (approximately a 7% increase in Case 1 Fig. 6(f) and Case 3 in Fig. 8(f)) in tumor contrast. After completing all three lesion scanning steps, the absolute HbT values within the tumor are somewhat lower than those reported by other groups [14], likely due to the partial mammographic compression. Still, relative tumor HbT contrasts of 2.3 and higher have been recovered in all cases, consistent with literature values [5]. Though not shown, in the 12 cases where initial tumor contrast was present, better contrast recovery has also been achieved after implementing the lesion scanning steps (mean ± std = 3.6 ± 1.7), a significant increase from the original contrast of 1.4 ± 0.6.

Judging by the shifts from the initial contour-based registration locations seen in these cases, the primary contributor to the missed contrast is a large misalignment in the depth Z. In all three cases, a shift of over 10 mm has been observed, a displacement commensurate to the tumor size in Z. This means that the lesion prior could have been placed at a depth that entirely missed the true tumor, resulting in a loss of opportunity for contrast enhancement. The large differences in tumor depth could partially be attributed to the relative tumor location to the breast boundary. These three patients all have tumors that are fairly close to the periphery of the breast, thus being more prone to out-of-plane rotation between two separate positionings, especially in the first patient case shown in Fig. 6 where the tumor is located close to the nipple. Moreover, all three patients have large breasts, which allows for larger shifts in all dimensions compared to other less challenging cases. For example, in the first patient case, compared to the DBT positioning (measured 12.1 cm along the Y axis), the breast was intentionally inserted less (measured 9.5 cm along the Y axis) by the mammographer for optical imaging to ensure her lesion had adequate optode coverage. The differences in the amount of breast insertion are not accounted for by the contour-based image registration. As a result, this case has seen the largest shift in Y (close to 2 cm) compared to other cases. Regardless, the lesion scanning approach was effective in correcting repositioning mismatch to successfully regain tumor contrast even in these three difficult cases.

This lesion scanning approach does have some limitations. First, it is computationally expensive. To complete the three-step lesion scanning for a patient case, Step 1 requires ∼1,000 reconstructions, Step 2 requires up to 1,620 reconstructions (9 × 9 grid at 20 depths), and Step 3 requires up to 729 reconstructions (9 × 9 grid at 9 depths) and potentially two iterations; that is, a total of ∼4,078 reconstruction jobs are needed, which requires a substantial amount of computing resources and time both to run the reconstructions and to pool results to calculate the metrics maps. We have leveraged a shared 1,200 CPU core computing cluster, funded by the Massachusetts Life Sciences Council, to parallelize the reconstructions to significantly shorten the time to several hours to complete all reconstruction jobs at each step. It is also worth to mention that for most cases that show tumor contrast initially it is feasible to reduce the scanning region and steps to accelerate processing. Second, human intervention is needed to determine the ROI at each step. This requires not only expert knowledge in DOT imaging to avoid potential bias by image artifacts but also robust reasoning skills to judge what amount of shift in lesion prior can be reasonably expected. For example, out-of-plane rotations can cause a large displacement in the Z direction near the outer contour of the breast, but less so towards the center, and lesions near the chest wall are unlikely to show much displacement in X. These judgement calls, though cumbersome, may be necessary at this stage, especially in the challenging cases shown in this work where initial tumor contrast is absent while multiple confounding regions are present. In the future, however, developing an objective and robust automated scanning pipeline is of interest to translate the lesion scanning approach in practice. Moreover, while our method can effectively correct inaccurate tumor locations in our multimodal setup, it cannot address the inaccuracy in two-composition adipose/fibroglandular structural prior created based on separately acquired DBT images. However, limited quantification errors in background tissue do not negate the benefit of tumor contrast enhancement. As a matter of fact, cross-modality image registration is not required to take advantage of the lesion scanning approach described herein. Standalone DOT system can still reconstruct images guided only by the lesion prior precisely localized by our method using a homogenous background assumption. Finally, our grid size of 5 mm for the first two steps could limit its effectiveness in small lesions. Patient cases shown in this work were enrolled under a therapy monitoring study that had an inclusion criteria of tumor size larger than 10 mm. While a 5-mm scanning grid works well in these cases at the baseline scan, in subsequent follow-up scans of patients who responded well to their therapy, unstable patterns sometimes were seen in the metrics maps, which indicates the need of finer scanning grid. Moreover, for smaller lesions, more than three whole-breast scanning planes may be needed in Step 1 to ensure a commensurate gap in z planes with lesion size.

In summary, we believe that application of the proposed lesion scanning method can be useful in improving the quantification accuracy in optical breast images by searching for the tumor contrast within the vicinity of radiologic markings. While simultaneous acquisition of optical and anatomical data is ideal for the purpose of prior-guided image reconstruction, our approach provides an alternative that enables separately acquired DOT data to benefit the same contrast enhancement as natively co-registered multimodal systems. We also expect this approach to greatly improve the consistency in the recovered tumor contrasts in longitudinal therapy monitoring studies that require repositioning of patients’ breasts in every follow-up scan.