This study addresses the challenges of generalizability and scalability in computer-vision-based surgical guidance systems. The researchers propose an equipment-agnostic framework for real-time surgical decision support, using laparoscopic cholecystectomy as a use case. Lightweight U-Net and SegFormer models were trained on a large, diverse multicenter dataset and tested on a separate dataset. A web platform enabled real-time inference from any edge device, optimized for varying network speeds. The models achieved promising performance metrics (Dice scores, precision, recall) for predicting safe/dangerous dissection zones, and the platform delivered a prediction stream of at least 60 fps with minimal delay even on low-bandwidth connections. The findings demonstrate the feasibility and cost-effectiveness of deploying machine learning models for real-time surgical guidance in diverse settings.
