Domain Adaptive Multiple Instance Self-Training for Intraoperative Anomaly Detection.

Area of Science:

• Computer Vision
• Surgical Robotics
• Medical Imaging Analysis

Background:

• Intraoperative anomalies disrupt surgical workflows, increasing error risks.
• Automated anomaly detection is vital for safe assistive and autonomous surgery.
• Current systems face challenges with domain shifts and deformable environments.

Purpose of the Study:

• To develop a robust anomaly detection framework for surgical video analysis.
• To enhance adaptability across diverse surgical platforms and scenarios.
• To improve the reliability of automated systems in identifying intraoperative anomalies.

Main Methods:

• Proposed DA-MIST (Domain Adaptive Multiple Instance Self-Training) framework.
• Utilized a two-stage training strategy combining multiple instance learning and self-training.
• Implemented a scene-decoupled memory mechanism and a state-aware dual-branch attention module.

Main Results:

• DA-MIST demonstrated strong adaptability across heterogeneous surgical domains.
• The framework consistently reduced false alarms in anomaly detection.
• Enhanced anomaly localization accuracy was achieved on a large-scale endoscopic video dataset.

Conclusions:

• DA-MIST offers a robust solution for weakly supervised anomaly detection in surgery.
• The proposed methods improve system performance and generalizability.
• This work contributes to the advancement of safer autonomous surgical systems.