Surgical Workflow Analysis: An Explainable Approach

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning in Healthcare

Background:

• Manual surgical workflow analysis is inefficient and inconsistent.
• Need for automated solutions in catheterization labs for improved efficiency and patient safety.
• Existing methods lack interpretability and adaptability.

Purpose of the Study:

• To develop an explainable, automated two-stage model for surgical workflow analysis.
• To enhance real-time classification of workflow phases using computer vision and machine learning.
• To validate the model's generalizability and clinical implementation potential.

Main Methods:

• Utilized ceiling-mounted cameras for data acquisition.
• Implemented a two-stage model combining YOLOv8 object detection and Gaussian Mixture Model - Hidden Markov Model (GMM-HMM).
• GMM-HMM modeled spatial-temporal dynamics for real-time workflow phase classification.

Main Results:

• Achieved high accuracy: 95.2% on RdGG dataset and 95.4% on HH Tampere dataset.
• Demonstrated generalizability across different hospital environments.
• Highlighted model explainability, robustness, and precise phase transition identification.

Conclusions:

• The developed model offers a precise, explainable, and adaptable solution for automated surgical workflow analysis.
• Potential for clinical implementation due to real-time application and cross-hospital adaptability.
• Future work includes improving occlusion robustness and integrating multi-modal data.