B2Q-Net: Bidirectional Branch Query Network for Surgical Phase Recognition.

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Workflow Analysis

Background:

• Surgical phase recognition (SPR) is crucial for analyzing surgical workflows and providing real-time guidance.
• Current SPR methods often aggregate frame-level data, limiting the integration of historical context into local temporal modeling.

Purpose of the Study:

• To introduce a novel network, Bidirectional Branch Query Network (B2Q-Net), for improved surgical phase recognition.
• To address the limitations of existing methods by enabling bidirectional information flow between phase-level and frame-level features.

Main Methods:

• The B2Q-Net reformulates SPR as a bidirectional query between phase-level and frame-level features.
• It incorporates historical information during phase query initialization and uses a dual-scale selector (DSS) for high-quality phase queries.
• A state space query (SSQ) module with learnable tokens preserves historical information.

Main Results:

• B2Q-Net demonstrated superior recognition accuracy compared to state-of-the-art methods across three datasets.
• The network achieved a high inference speed of 106 frames per second (fps).

Conclusions:

• B2Q-Net offers a significant advancement in surgical phase recognition by effectively integrating historical context.
• The method provides accurate and efficient real-time guidance for surgical procedures.