Critical View of Safety Assessment in Laparoscopic Cholecystectomy via Segment Anything Model

Laparoscopic Cholecystectomy (LC) is a minimally invasive surgery for the removal of diseased gallbladders. Compared to traditional open cholecystectomy, LC procedures is associated with significantly shorter recovery period, but has an increased chance of bile duct injuries (BDIs). Critical view of safety (CVS) is an important validation method and safety protocol which has a set of conditions that can be visually identified during LC surgeries. In this paper, we approach the problem of automated CVS prediction by combining state-of-the-art object detection methods and prompting the Segment Anything model to achieve more accurate localization of anatomical structures and classification of CVS conditions. When evaluated on our dataset of 5,750 frames with CVS annotations, our method achieved competitive results on frame-level CVS condition prediction, and around 20% improvement on video-level CVS assessment compared to previous SoTA LG-CVS.