DBT-Bleed: Dual-Branch Temporal Modeling with Key-Frame Selection for Surgical Bleeding Detection

DBT-Bleed is a newly proposed dual-branch multi-scale temporal modeling framework designed for the detection of intraoperative adverse events (IAEs), specifically surgical bleeding. It utilizes layer-wise temporal adapters to differentiate between bleeding and normal states while introducing HiRED, a Hierarchical Entropy-Driven frame selection strategy to efficiently process long surgical videos. Benchmark results on the MultiBypass dataset show improvements of 6.53% in F1 score, 5.62% in Recall, and 9% in MCC for bleeding detection, and the framework also demonstrates robust zero-shot transferability on a novel EndoPit-IAE dataset, highlighting its practical relevance for enhancing surgical safety through improved IAE detection.