High-speed Intraoperative Assessment of Breast Tumor Margins by Multimodal Ultrasound and Photoacoustic Tomography

Rui Li, Lu Lan, Yan Xia, Pu Wang, Linda K. Han, Gary L. Dunnington, Samilia Obeng-Gyasi, George E. Sandusky, Jennifer A. Medley, Susan T. Crook, and Ji-Xin Cheng* 1 Weldon School of Biomedical Engineering, Purdue University, 206 S Martin Jischke Dr., West Lafayette, Indiana, 47907, USA. 2 School of Biological Science and Medical Engineering, Beihang University, Beijing, China, 100083 3 Beijing Advanced Innovation Center for Biomedical Engineering, Beihang University, Beijing, China, 102402 4 Vibronix, Inc., 1281 Win Hentschel Blvd., West Lafayette, Indiana, 47906, USA. 5 Indiana University Health Melvin and Bren Simon Cancer Center, 1030 W. Michigan St., Breast Center, Indianapolis, Indiana 46202, USA. 6 Department of Pathology & Laboratory Medicine, Indiana University School of Medicine, 350 West 11th Street, Indianapolis, Indiana, 46202, USA. 7 Department of Radiology and Imaging Sciences, Indiana University School of Medicine, 550 N. University Blvd, Indianapolis, Indiana 46202, USA 8 Purdue University Center for Cancer Research, 201 S. University Street, West Lafayette, Indiana, 47906, USA. 9 Photonics Center, Boston University, 8 Saint Mary’s Street, Boston, MA 02215, USA


Accepted Article
This article is protected by copyright. All rights reserved.

Engineering an intraoperative multimodal ultrasound and PAT system
Using overtone absorption of lipid as contrast, we built a multimodal ultrasound and PAT system capable of fast and automatic detection of ultrasound/photoacoustic signals (Figure 1) with the goal of distinguishing breast cancer from non-cancerous tissue within 2 mm surface in 10 minutes. The system employed a custom-built, all-solid-state Raman laser as the excitation source, since the output wavelength lies in the optical window to solely visualize lipid in the breast [42,43]. The 10 Hz, 10 ns pulse trains with the pulse energy of 100 mJ at 1197 nm wavelength were delivered to the excised breast tissue via a fiber bundle to effectively and sufficiently excite the lipid in the breast tissue to generate photoacoustic signals ( Figure S2). The generated photoacoustic signals were acquired by a customized 18 MHz high-frequency ultrasound array with 128 elements and 50% bandwidth. Meanwhile, the same transducer array emits and receives ultrasound signals, which were then processed by a high-frequency ultrasound imaging system. The laser and the ultrasound system were connected to a computer to synchronize the trigger and visualize the registered ultrasound/photoacoustic images. In order to meet the need of fast intraoperative margin assessment, an automatic tissue scanner ( Figure S3) was designed and built based on a series of clinical testing of specimens from a total of 36 patients. In the final version, we achieved twodimensional scanning of 10 x 10 cm 2 in 2 minutes, which covered the size of the majority of excised breast tumor tissues in lumpectomies [22]. To adapt to the tissue surface irregularity during the scanning, a collinear design was applied to fabricate the imaging probe ( Figure S4), which was comprised of the ultrasound transducer, a fiber bundle, a pair of cylindrical lenses, and two glass slides. Light illumination from the imaging probe was consistent with the transport simulations ( Figure S5) using ray-optics simulation toolbox from Matlab, demonstrating that the light from the end of the probe was collimated. With the collinear design, a fresh excised breast tumor tissue was

Accepted Article
This article is protected by copyright. All rights reserved.
imaged, showing 6 mm photoacoustic imaging depth and 13.4 mm adaptation, which was two times better than the traditional bifurcation design ( Figure S6).

Development of an imaging protocol
A total of 30 female patients (Table S3) were enrolled in a study to determine the sensitivity and specificity of the multimodal ultrasound and PAT system. Patients requiring lumpectomy for diagnosis of breast cancer, including ductal carcinoma in situ (DCIS) and invasive ductal carcinoma (IDC) diagnosed by pre-operative imaging and core biopsies, were recruited at Indiana University Health Simon Cancer Center. All the patients signed and gave the informed consent for this study one day before the surgery per protocols approved by the Institutional Review Boards at Indiana University Health. Patients' names and other HIPPA identifiers were removed from all sections of the study.
In lumpectomy operation, after a tumor mass was resected, it was first oriented with a long suture marking the lateral margin and a short suture marking the superior margin. Then, it was placed on a sample tray, and transparent ultrasonic gel was applied on its surface for signal coupling (Figure 2A).
The tray with gel coated specimen was inserted into the tissue scanner. The cover of the tissue scanner was then closed, and the gel contacted with a plastic film of a water reservoir. In the meantime, distilled water was automatically poured into the reservoir when the cover was closed. A series of 1,000 two-dimensional (2D) ultrasound/photoacoustic images ( Figure 2B, C) were taken over a 10 by 10 cm 2 area within 2 minutes, and presented in three-dimensional (3D) layout ( Figure   2D, E). Upon completion of the first imaging, the tissue was taken out from the scanner, flipped 180 degrees, and the aforementioned procedures were repeated to acquire the margin information of the other surface. With the automatic-scanning design, we minimized the whole procedure time to < 10 min. Finally, the tissue was cleaned and delivered to the histology room for further standard

Accepted Article
This article is protected by copyright. All rights reserved. histopathology analysis. The researchers involved in the data acquisition were completely blinded to the final histology results.

Image analysis & evaluation protocol
All the ultrasound/photoacoustic images were acquired and processed with the same standard, and displayed in the same brightness level and contrast scale, as recommended by board-certified radiologists with fellowship training in breast imaging. Data collected from the first enrolled 10 patients was used as a training set to establish the evaluation criteria. The representative images from the training data set including different features were shown in  [23]. Therefore, for IDC, if it was not all covered by adipose tissue (no photoacoustic signals on the surface), it was considered a positive margin ( Figure 4A-C), which was consistent with the current invasive breast margins guideline of margins on ink. For DCIS, if it was not all covered by adipose tissue with at least 2 mm thickness or micro-calcifications were found within 2 mm surface, it was considered a positive margin ( Figure 4G-I). The margin assessment was made using the full set of ultrasound/photoacoustic images for each tissue. As long as at least one image frame showed abnormal, the tissue was deemed as positive in margin assessment.

Accepted Article
This article is protected by copyright. All rights reserved.

Performance based on subjective reading
A blinded subjective reader study was performed to evaluate the statistical performance of the intraoperative multimodal ultrasound and PAT system. Data from the remaining 20 patients were used to calculate the sensitivity and specificity ( Table 1) associated with the margin interpretation of the breast tumor tissues in the tissue/patient level. Three readers including an imaging researcher, a breast surgeon, and a board-certified radiologist were recruited for this study. The readers were firstly given a training set of sample ultrasound/photoacoustic images showing tumor mass, microcalcifications, adipose tissue, and breast cyst to familiarize with the images acquired by the imaging system. Then, each reader was instructed to visualize the images of the 20 breast tissues and score the tissue on a scale of 1 to 4 as following: i) a score of 1 means that the reader is confident that the margin is negative for cancer; ii) a score of 2 means that the reader thinks that the margin is likely negative, but there is some uncertainty; iii) a score of 3 means that the reader thinks that the margin is likely positive, but there is some uncertainty; iv) a score of 4 means that the reader is confident that the margin is positive for cancer. After reading, the tissue margin was declared as negative only if given a score of 1 and positive if given a score of 2, 3, and 4, which represented a conservative clinical scenario. During the reading, each reader at first was instructed to evaluate the tissue margins by visualizing only ultrasound images of each tissue. Then, each reader was instructed to evaluate the tissue margins by visualizing ultrasound plus photoacoustic images of each tissue to judge whether the photoacoustic images would improve the reading results or not. Table 1

Performance based on machine learning
Subjective interpretation of the ultrasound/photoacoustic images by either breast surgeons or radiologists is time-intensive (10 to 15 min/tissue, 3 to 5 s/frame) and reader-biased. Consequently, a system which is able to rapidly deliver the consistent and accurate margin evaluation will be preferred during the breast-conserving surgeries. We employed a machine learning-based algorithm called deep convolutional neural network (CNN) for breast tumor margin assessment because it needed relatively less pre-processing and could achieve end-to-end supervised learning [46,47]. In this study, GoogLeNet Inception v3 CNN architecture was applied and adapted to the margin assessment through transfer learning. Based on the inception v3 architecture pre-trained on the ImageNet dataset, the original leaf nodes were firstly replaced with our own two-node (positive against negative). Then, the optimal threshold value for ultrasound model and photoacoustic model was obtained via the rule of farthest point from the diagonal in the Receiver Operating Characteristics (ROC) curve, respectively.
A training data set of 1052 positive frames and 918 negative frames was built by flipping and rotating the images generated by the multimodal PAT system. These frames were determined as positive or negative based on the reading of ultrasound and photoacoustic images by a trained radiologist, also they were correlated to pathology H&E staining read by a pathologist. Two independent CNN models were trained with only ultrasound and only photoacoustic data, separately. Then the two models were combined with an OR method, that was, an image was classified as positive if predicted as positive by either ultrasound model or photoacoustic model ( Figure 5). A testing set was built with 76 positive frames and 70 negative frames, of which the

Accepted Article
This article is protected by copyright. All rights reserved.

Discussion
The goal of breast conservation surgery is to resect all cancer tissue while preserving as much normal tissue as possible for optimal cosmetic outcome. Achieving clear margin status of the excised breast tumor tissue, a key predictor of local recurrence [48,49], necessitates a fast and accurate intraoperative margin assessment tool. Current methods either need long procedure time, or lack sensitivity and specificity. Here, we demonstrate an intraoperative multimodal ultrasound and PAT system for high-speed and accurate assessment of breast tumor margins in 20 patients.
The ideal intraoperative tool for tissue margin assessment should image the entire tissue surface in a fast manner. In breast conservation surgery, the excised tissue can be of arbitrary shape and large size, which usually has a tissue surface area ranging from 1 cm 2 to 100 cm 2 . The tissue surface inevitably has a lot of fluctuations and irregularities, e.g. 10 mm height difference between tissue peaks and valleys. Conventional margin assessment tools, such as electrical resonance spectroscopy and OCT, have limited imaging depth of < 2 mm, and cannot adapt to such tissue surface fluctuation in scanning mode, which is significant to reduce the assessment time. To overcome this challenge, we developed two novel components. Firstly, our collinear imaging probe was able to provide 6 mm

Accepted Article
This article is protected by copyright. All rights reserved. imaging depth and adapts to tissue surface fluctuation up to 13 mm. Secondly, we developed an imaging chamber that significantly simplified and shortened the tissue preparation time for ultrasound/photoacoustic imaging. For conventional scanning of a specimen, agarose gel solution is needed to be prepared beforehand, poured into a tissue container, naturally cures to fix the tissue, and water is later added to couple the signal from the tissue to the imaging probe. Such tissue preparation method is not only complex, but also time-consuming. Here, we designed a tissue cartridge and tissue container to prepare the tissue for scanning and reconstruct 3D images within 5 minutes in 4 steps: 1) Before the imaging, the fresh excised breast tissue is rinsed by 0.9% saline solution, which takes 1 minute; 2) Ultrasound gel is applied on the surface of the tissue, which takes less than 1.5 minutes (depending on the tissue size); 3) The tissue cartridge with the tissue was inserted into the imaging chamber, which takes less than 0.5 minute; 4) By pressing one button, the tissue surface was scanned within 2 minutes. Thus, the user was able to visualize the tissue margins of one surface within 5 minutes. In this study, two opposite faces were imaged for one tissue, taking less than 10 minutes, which was the fastest intraoperative breast tumor margin assessment tool to date. Also, 3D images of the breast tissue could be reconstructed, showing the tumor location and size. By correlating the suspicious area to the lumpectomy cavity, the surgeon will be able to locate the corresponding area of concern for future resection.

Accepted Article
This article is protected by copyright. All rights reserved. interpretation of thresholds for calling an image positive, and there is inter-reader variability (Cohen's Kappa Coefficient between reader 1 and reader 2, between reader 1 and 3, and between reader 2 and 3 is 0.565, 0.340, and 0.205, separately). However, the most conservative reading still has a sensitivity of 71% and a specificity of 92%, which is better than the current radio frequency spectroscopy (71% sensitivity and 68% specificity) and intraoperative X-ray (49% sensitivity and 73% specificity). The reported optical coherence tomography achieves 100% sensitivity and 82% specificity, but the decision criterion is margin on surface ink for all the tumor types in the reading.
However, the new-released consensus guideline on breast margins requires at least 2 mm for DCIS, which leads to a sensitivity of only 63% for a recent clinical study [50] with the afore-mentioned optical coherence tomography probe because of its limitation in imaging depth. High frequency ultrasound itself provided a decent sensitivity in evaluating the breast tumor margins, especially in IDC specimens. However, with photoacoustic tomography, the detecting specificity of breast margins improved in IDC specimens, which would avoid unnecessary tissue excisions. Moreover, by adding the photoacoustic imaging channel, the multimodal ultrasound and PAT system improved both detection sensitivity and specificity for in situ cancer within 2 mm margins (Supplementary Table 4

-6).
Machine learning-based algorithm may improve the speed, consistency and accuracy of margin assessment for breast conserving surgeries, intraoperatively. Our results have shown that CNNbased model achieved higher accuracy and adaptability than other automated approaches. The ideal CNN-based classification algorithm will require little pre-processing and provide general margin assessment for all tumor types in real time. The CNN-based classification algorithm reported here was able to distinguish between positive and negative margin status of 2 mm margin area for in-situ tumor at a speed of 0.7 s/frame. We anticipate that the accuracy of our classification algorithm can be further improved by: (1) collecting more patient data (including both positive and negative cases); (2) reducing background noise and fine-tuning parameters of image pre-processing. We also expect to obtain a more generalized classification algorithm that can be used for more types of breast

Accepted Article
This article is protected by copyright. All rights reserved. tumors by: (1) training more CNN models focusing on different tumor types; or (2) adding another layer of pre-classifier before margin assessment.

Conclusion
As the prevalence of breast cancer screening, more lesions are identified in an early stage, leading to the popularity of using breast conservation surgeries. Achieving a negative or clear margin is essential for improving the clinical outcomes. However, an unmet need still exists to achieve timeefficient and highly-sensitive intraoperative evaluation of breast cancer margins during surgical procedures. Here, we demonstrate the first application of a customized mutlimodal ultrasound and PAT system for intraoperative breast tumor margins assessment using a compact and portable Raman-laser-based system and fresh lumpectomy specimens from 66 patients. Fresh lumpectomy specimens from a total of 36 breast cancer patients were used to improve the design, especially the collinear imaging probe and automatic tissue scanner.

Accepted Article
This article is protected by copyright. All rights reserved.

Experimental Section
Clinical Study Design: The objective of this study was to evaluate the potential of the multimodal ultrasound and PAT system to intraoperatively assess the breast tumor margin status in breastconserving surgeries. Fresh excised human breast tumor tissues (n = 66) with different tumor types (IDC and DCIS) at Indiana University Health Simon Cancer Center were obtained and imaged by the imaging system. Patients undergoing mastectomies, or having infectious diseases, were excluded from this study. No patients were excluded based on age, ethnicity, race, or weight (Table S2) The spectral profile of the generated Raman laser indicated the central wavelength of 1197.6 nm ( Figure S2C), which lay in the second overtone absorption peak of lipid. The pulse duration of the output was measured to be 10.75 ns ( Figure S2D), which met the thermal and stress confinements to efficiently generate photoacoustic signals. The maximum laser output could reach to 142 mJ when the pump energy was 294 mJ, corresponding a conversion efficiency of 48.3% ( Figure S2E). The

Accepted Article
This article is protected by copyright. All rights reserved.
continuous laser power output within three hours was shown in Figure S2F. It indicated an average output of 1 W with 1.8% instability, which guaranteed sufficient pulse energy to excite the stable photoacoustic signals.
Design of the Automatic Tissue Scanner: The automatic tissue scanner ( Figure S3A) was capable of 2D scanning of 10 x 10 cm 2 area within 2 minutes. The integrated imaging probe was placed on the top cover. After each breast conservation procedure, the top cover was lifted, and the fresh tissue was put on the tissue holder on the bottom ( Figure S3B). Then, the top cover was closed, and a water reservoir covered with a plastic film contacted with the applied gel. Distilled water which was reserved in a water bottle was directed to the water reservoir as acoustic-coupling medium. The imaging head then performed a 2D scan of the tissue within the water reservoir ( Figure S3C). After one surface scanning, the top cover was lifted, and the tissue was flipped for the other surface scanning. The whole procedure took < 10 minutes, which met the current economic need.
Design of the Collinear Imaging Probe: The schematic of the collinear imaging probe was shown in Figure S4. It was comprised of a custom-built ultrasound transducer array, a fiber bundle, a pair of cylindrical lenses, and two glass slides. The laser light propagating from the fiber bundle went through one glass slide to the tissue surface to excite ultrasound signals. The generated signals were reflected by the two glass slides, and then received by the ultrasound transducer array. This collinear design was able to generate a collinear laser beam with a size of 2.6 mm x 10 mm at a distance of 10 mm ( Figure S5), which was beneficial to adapt the tissue surface roughness. With a fresh excised breast tumor tissue, collinear design could reach 6 mm imaging depth with 13.4 mm adaptation, which was two times better than the traditional bifurcation design ( Figure S6).
Histology Evaluation: Based on the suture orientation, the fixed breast tumor tissue was first inked with different colors for the following margin identification. Then, it was grossed into several blocks.
Through gross inspection, only suspicious areas from each block were incised and placed into a cassette for further standard H&E staining. Histology slides were digitized with a light microscope