Temporal ROI-Based Framework for Automated Multi-Criterion Critical View of Safety Assessment in Laparoscopic Cholecystectomy

Accurate assessment of the Critical View of Safety (CVS) during laparoscopic cholecystectomy is essential to prevent bile duct injuries. Automated CVS assessment from surgical videos remains challenging due to complex anatomy, instrument occlusion, and the temporal nature of dissection. However, existing methods rely on full-frame processing without focusing on anatomically relevant regions, limiting both efficiency and anatomical specificity. We propose a region-of-interest (ROI)-based spatio-temporal framework for automated multi-criterion CVS classification that explicitly models anatomical relationships and surgical progression using the multi-modal Endoscapes dataset. Rather than processing full frames, our pipeline extracts anatomically relevant ROIs guided by bounding box annotations, models intra-frame structural dependencies via a Graph Attention Network and captures surgical progression through a bi-directional LSTM with temporal attention. Complementary segmentation masks are leveraged to validate that model attention aligns with true anatomical boundaries. Evaluated on the Endoscapes test set, our method achieves 70.8% mean Average Precision (mAP), 74.6% balanced accuracy for multi-criterion classification, and an Area Under the ROC Curve (AUC) of 86.8% for overall CVS assessment, outperforming representative full-frame baselines. The model demonstrates high specificity (> 90%) and clinically interpretable attention maps that consistently localize to Calot's triangle and cystic structures. By decoupling spatial detection from temporal reasoning and exploiting multi-modal annotations, this work advances automated surgical safety assessment through clinically grounded ROI processing and expressive spatio-temporal modelling.