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

Researchers introduce DBT-Bleed, an AI framework for detecting intraoperative bleeding during surgery by using dual-branch temporal modeling and intelligent frame selection. The system significantly outperforms existing methods on bleeding detection while demonstrating cross-procedure generalization capabilities, alongside a new neurosurgery dataset for adverse event research.

DBT-Bleed addresses a critical gap in surgical safety technology by solving the challenge of distinguishing active bleeding from residual blood through advanced temporal reasoning. Current video analysis systems struggle with this distinction because they lack sophisticated temporal context modeling, creating safety risks in operating rooms where bleeding detection speed directly impacts patient outcomes. The research team developed a dual-branch architecture that processes short and long-term bleeding progression patterns independently, then combines these insights for more accurate classification.

The innovation extends beyond model architecture through HiRED, a computational efficiency breakthrough using entropy-driven frame selection. Surgical videos often contain extended periods of visual redundancy, and attempting to process every frame becomes prohibitively expensive. HiRED intelligently removes these redundant segments while preserving temporal information critical to bleeding detection, enabling practical deployment on resource-constrained surgical systems.

The 6-53% F1 score improvement demonstrates tangible progress toward clinical viability, while cross-procedure generalization results show the model learns generalizable bleeding patterns rather than memorizing dataset specifics. The introduction of EndoPit-IAE marks the first neurosurgery dataset annotated for intraoperative adverse events, establishing a foundation for future AI safety research in this surgical specialty. This development signals growing institutional commitment to AI-assisted surgical monitoring, with applications extending to real-time decision support systems.

Investors tracking healthcare AI should note this represents validated progress in a regulated domain where clinical evidence translates to adoption barriers. The cross-procedure generalization success suggests scalability potential across surgical specialties, supporting business models around universal surgical monitoring platforms. Future developments will likely focus on real-time implementation, regulatory certification, and integration with existing OR infrastructure.

• →DBT-Bleed achieves 6.53% F1 improvement over baseline methods in detecting intraoperative bleeding during surgery
• →HiRED frame selection strategy enables efficient processing of long surgical videos without sacrificing temporal precision
• →Cross-procedure zero-shot evaluation demonstrates 6% F1 gains, indicating the model generalizes bleeding detection beyond training data
• →EndoPit-IAE introduces the first neurosurgery dataset annotated for intraoperative adverse events, expanding research opportunities
• →Dual-branch temporal modeling architecture disentangles bleeding from visually similar residual blood through enhanced temporal reasoning