TriQuery: A Query-Based Model for Surgical Triplet Recognition

Area of Science:

• Medical Artificial Intelligence
• Computer Vision
• Surgical Technology

Background:

• Surgical video action recognition is crucial for intraoperative understanding and decision support in intelligent surgical systems.
• Challenges include temporal continuity and the complex distribution of surgical action triplets (instrument, verb, target).

Purpose of the Study:

• To propose TriQuery, a query-based model for surgical triplet recognition and classification.
• To enhance the accuracy and interpretability of surgical video analysis.

Main Methods:

• TriQuery utilizes a multi-task Transformer framework with task-specific query tokens.
• It employs specialized attention mechanisms, a Multi-Query Decoding Head (MQ-DH), and a Top-K Guided Query Update (TKQ) module for temporal cues.

Main Results:

• TriQuery demonstrated improved performance over existing baselines on the CholecT45 dataset across multiple classification tasks.
• Attention visualizations confirmed that task queries focus on relevant spatial regions, enhancing model interpretability.

Conclusions:

• TriQuery effectively advances surgical video understanding for clinical applications.
• The query-based approach offers a promising direction for complex surgical action recognition.