Learning dissection trajectories from expert surgical videos via imitation learning with equivariant diffusion

Area of Science:

• Medical Robotics
• Surgical Simulation
• Machine Learning in Medicine

Background:

• Endoscopic Submucosal Dissection (ESD) is a key technique for removing epithelial lesions.
• Predicting dissection trajectories in ESD videos can significantly enhance surgical training.
• Existing imitation learning methods struggle with uncertainty, geometric symmetries, and generalization in surgical scenarios.

Purpose of the Study:

• To develop an advanced imitation learning technique for predicting dissection trajectories in ESD videos.
• To address the limitations of current methods in handling variability and generalization.
• To improve the accuracy and robustness of surgical skills training through better trajectory prediction.

Main Methods:

• Proposed a novel Implicit Diffusion Policy with Equivariant Representations for Imitation Learning (iDPOE).
• Utilized a diffusion model for policy learning to capture stochasticity and improve training efficiency.
• Integrated equivariance to handle geometric symmetries and enhance generalization.
• Developed a forward-process guided action inference strategy for conditional sampling.

Main Results:

• The iDPOE method demonstrated superior performance in dissection trajectory prediction compared to state-of-the-art approaches.
• The model showed improved accuracy and generalization capabilities across diverse endoscopic views.
• Experimental results were validated on a dataset of nearly 2000 ESD video clips.

Conclusions:

• This research presents the first imitation learning-based approach for surgical skill learning focused on dissection trajectory prediction in ESD.
• The proposed iDPOE method offers a robust and accurate solution for predicting surgical trajectories.
• The findings have significant implications for advancing surgical training and simulation technologies.