Multi-Teacher Temporal Regulation Network for Surgical Workflow Recognition.

Area of Science:

• Computer Vision
• Medical Informatics
• Artificial Intelligence

Background:

• Automatic surgical workflow recognition is crucial for operating room efficiency.
• Existing methods struggle with extracting sufficient visual features and exhibit performance degradation due to training-testing feature inconsistencies.
• Surgical videos present challenges due to their complexity and extended duration.

Purpose of the Study:

• To propose a novel Multi-Teacher Temporal Regulation Network (MTTR-Net) to address limitations in current surgical workflow recognition methods.
• To enhance the accuracy and robustness of automatic surgical workflow recognition systems.

Main Methods:

• Implemented a "sequence of clips" training strategy with an auxiliary temporal regularizer for discriminative visual feature extraction.
• Developed a cross-mimicking strategy to generate consistent features and mitigate training-testing inconsistency.
• Utilized a temporal encoder trained on mimicked features for stable performance.

Main Results:

• MTTR-Net demonstrated superior performance compared to state-of-the-art methods across various metrics.
• The proposed methods effectively addressed insufficient feature extraction and feature inconsistency issues.
• Experiments were conducted on eight public surgical video datasets.

Conclusions:

• MTTR-Net offers a significant advancement in surgical workflow recognition.
• The developed strategies improve the reliability and accuracy of automated surgical video analysis.
• The approach holds promise for enhancing surgical training and real-time decision support in operating rooms.