PCNet: Accurate Surgical Workflow Recognition with Phase Consistency.

Area of Science:

• Robotics
• Computer Vision
• Surgical Technology

Background:

• Real-time surgical workflow recognition is essential for autonomous robot-assisted surgery.
• Current methods often overlook spurious phase changes, impacting online operational reliability.

Purpose of the Study:

• To develop a model that accurately recognizes surgical workflows in real-time.
• To address and mitigate the issue of spurious phase changes in surgical phase recognition.

Main Methods:

• A multi-module model integrating spatial and temporal features using a transformer-like architecture.
• Introduction of a feedback loop to enforce phase consistency in output predictions.
• Comparative analysis using the Cholec80 dataset, evaluating ResNeSt against ResNet for feature extraction.

Main Results:

• The proposed model demonstrates accuracy comparable to state-of-the-art methods.
• The model significantly outperforms existing approaches in output continuity, reducing spurious phase changes.
• Validation of ResNeSt as a superior feature extractor compared to ResNet in this context.

Conclusions:

• The developed model enhances real-time surgical workflow recognition by improving output continuity.
• The feedback loop mechanism is effective in enforcing phase consistency.
• Further research is needed to explore the trade-off between inference accuracy and the amount of phase change.