Automated Intraoperative Lumpectomy Margin Detection using SAM-Incorporated Forward-Forward Contrastive Learning

Tyler Ward, Braxton McFarland, Sahar Nozad, Talal Arshad, Hafsa Nebbache, Jin Chen, Xiaoqin Wang, Abdullah-Al-Zubaer Imran

Keywords: Breast, Forward-Forward Contrastive learning, Lumpectomy, Positive margin, SAM, Specimen radiography

Abstract: Complete removal of cancerous tumors with a negative specimen margin during lumpectomy is essential to reduce breast cancer recurrence. However, interpretation of 2D specimen radiography (SR) by radiologists, the current method used to assess intraoperative margin status, has limited accuracy. This study aims to improve positive margin detection performance on SRs by leveraging cutting-edge deep learning models. We developed a novel lumpectomy margin assessment method using an innovative pre-training technique, Forward-Forward Contrastive Learning (FFCL), followed by few-shot segmentation training leveraging the Segment Anything Model 2 (SAM 2). Experimental results on independent annotated breast SRs demonstrate the effectiveness of the proposed FFCL-SAM method in classifying and segmenting positive margins in SRs.

Submission Number: 115