Introduction

Breast cancer is the leading cause of cancer death among women. Unfortunately, reoperation rates for breast-conserving surgery due to positive margins remain high, affecting over 20% of patients. Wide Field Optical Coherence Tomography (WF-OCT) shows potential for intraoperative margin visualization. Given its novelty, WF-OCT presents an underexplored opportunity for augmenting image review with an adjunct AI-driven clinical decision support system, serving as both a productivity tool and a decision aid. Here, we present a computationally efficient convolutional neural network (CNN) based binary classifier that identifies suspicious regions for further review. Independent testing on pathology confirmed 155 margins (31 are positive) from 29 patients resulting in the area under the receiver operating characteristic (AUROC) value of 0.976, sensitivity of 0.93, and specificity of 0.98 at a decision threshold of 75%. The model correctly identified 96.8% of pathology-positive margins.

Methods

WF-OCT images and the corresponding pathology data were collected during an IRB-approved clinical trial that accumulated data between 2019 and 2021. The WF-OCT data was divided into three sets. The training and validation sets included 586 WF-OCT margin scans from 151 subjects (ages 63 ± 11.7). An independent test set included 155 margin scans (31 positive and 124 negative) from 29 subjects (ages 58.5 ± 9.1) with pathology-confirmed status. Each margin scan is comprised of several hundred OCT images depending on the size of the specimen. Each image was divided into smaller overlapping patches of size 420x188 pixels. Each patch was then labeled by two subject matter experts using pathology results as the ground truth. Conflicts between the two labelers were resolved in a follow-up meeting. The finalized model is designed to be computationally efficient and consists of five convolutional layers and three fully connected layers, totaling around 1,589,000 parameters.

Results

For the test set, the model achieved an area under the receiver operating characteristic (AUROC) value of 0.976, sensitivity of 0.93, and specificity of 0.98 at a decision threshold of 75%. According to the ISO/IEC TS 4213, the area under the precision-recall curve (AUPRC) is well-suited to handle scenarios, such as ours, of data imbalance, wherein optimizing performance on the positive class is prioritized. The AUPRC is measured as 0.812. Model interpretability also helps to foster model-clinician collaboration. An example of two suspicious annular structures indicative of ductal carcinoma in situ that are being highlighted by the model. In 1000 runs, the average inferencing time was 10.34ms for a single WF-OCT image of 420x2400 pixels using the mobile variant of the NVIDIA GeForce RTX 3070 GPU. Considering a typical margin will have 200 to 400 b-scan images, the total inferencing time is 2 to 4 seconds per margin, making the algorithm feasible for use intraoperatively.  

Conclusion

This work demonstrated the proof of concept for margin visualization through WF-OCT (Perimeter Medical Imaging AI Inc., Dallas, TX), augmented by a deep learning-driven clinical decision support system*. The aim is to assist surgeons intraoperatively by offering suspicious feature identification. This CNN model has recently been integrated into an investigational WF-OCT device that is being evaluated in an ongoing prospective, multicenter, randomized, double-arm trial focused on evaluating its influence on positive margin rates in breast conservation surgery (ClinicalTrials.gov Identifier: NCT05113927). Data generated from this trial will inform future development work.

*Caution: Investigational Device. Subject to U.S. Law. ImgAssist is not available for sale in the United States.