Abstract 458: Development and validation of convolutional neural network to identify regions of interest in lumpectomy margins using optical coherence tomography

Abstract

Abstract Background/Objective: Optical coherence tomography (OCT) is the optical analog of high-frequency ultrasound and produces real-time, high-resolution images up to 2 mm deep. Multi-reader studies of OCT have shown differentiation of normal parenchyma from neoplasms, including DCIS and cancers, with >85% sensitivity and specificity. Intraoperative evaluation of breast lumpectomy margins (LMs) with OCT may help achieve negative margins at primary surgery and avoid re-excision. Artificial Intelligence can be trained to spot regions of interest (ROI) in OCT LM images suspicious for malignancy. The purpose of this study was to develop and validate an automated convolutional neural network (CNN) to screen OCT LM images for ROIs. Methods: Following IRB approval, LMs from 126 patients with ductal malignancy were OCT imaged. Images were compared to corresponding permanent histology and annotated by breast pathologists to create a training set of 25,000 control ROIs. A CNN algorithm was developed with 3 convolutional layers, a 3x3 kernel, and 3 fully connected layers to perform binary classification of images as “suspicious” or “non-suspicious” for malignancy. A weighted loss function was used to balance the training data for non-suspicious vs. suspicious images and to tune sensitivity and specificity. Once trained and weighted, the CNN was tested in a prospective study using OCT images of 29 LMs from 29 patients with biopsy-proven ductal carcinoma in situ (DCIS), invasive ductal carcinoma (IDC), or both. CNN results were compared to permanent histology. Results: The patient population was 61.5 ± 7.3 years old, 100% female, with Stage 0-1 disease. Disease included IDC (n=20), invasive lobular (n=2), DCIS (n=27), mixed (n=74), atypical ductal hyperplasia (n=24), as well as benign findings including atypical lobular hyperplasia (n=19), lymphatic invasion (n=13), lobular carcinoma in situ (n=12), usual ductal hyperplasia (n=35), and duct ectasia (n=17). Following primary surgery, LMs were scanned using OCT and images were CNN analyzed. Approximately 1.9 M OCT ROIs were assessed, identifying 101,099 suspicious ROIs. Three hundred and eighty-four (384) ROIs were correctly identified, yielding a 70% true positive and 5.2% false positive rate with 70% sensitivity and 96% specificity. The receiver operating curve is shown below. Conclusions: Automated analysis of OCT images using a trained CNN to identify ROIs suspicious for DCIS or IDC in LMs is feasible, demonstrating high concordance with permanent pathology. These findings indicate the utility of AI for screening OCT images with potential utilization for intraoperative evaluation of LMs. A pivotal prospective clinical trial will be necessary to evaluate breast specimens in real time to determine if this application may improve re-excision rates in lumpectomy. Citation Format: David Rempel, Andrew Berkeley, Chandandeep Nagi, Vladimir Pekar, Margaret Burns, Beryl Augustine, Alia Nazarullah, Ismail Jatoi, Kelly K. Hunt, Alastair Thompson, Savitri Krishnamurthy. Development and validation of convolutional neural network to identify regions of interest in lumpectomy margins using optical coherence tomography [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 458.