Leveraging Graph Attention Networks for Targeted Feedback from Operating Room Surgery Videos

Providing targeted feedback in cataract surgery is essential for refining surgical techniques and supporting skill development. This paper introduces a framework for generating targeted feedback based on specific procedural steps in cataract surgery videos, using a specialized feedback catalog tailored to assess critical surgical actions. Our approach employs a Video Masked Autoencoder (VideoMAE) as the feature extractor, enhanced with a Graph Attention Network (GAT) to capture inter-label dependencies. This framework achieves an AUC of 0.839 on a cataract surgery video dataset, improving classification accuracy and specificity across multiple feedback criteria compared with various other methods. Our findings demonstrate its effectiveness in delivering structured, context-aware feedback, and highlight the potential of GAT-based architectures in advancing targeted feedback generation for surgical procedures.