Reproduce “Learning dissection trajectories from expert surgical videos via imitation learning with equivariant diffusion” on OpenResearch

About this paper

Hongyu Wang, Yonghao Long, Yueyao Chen, Hon-Chi Yip, Markus Scheppach, Philip Wai-Yan Chiu, Yeung Yam, Helen Mei-Ling Meng and 1 more

Researchers from The Chinese University of Hong Kong developed iDPOE, an imitation learning framework that leverages equivariant diffusion models to predict expert dissection trajectories from endoscopic surgical videos. This method achieves state-of-the-art accuracy and generalization for Endoscopic Submucosal Dissection (ESD), showing robustness to challenging visual conditions and maintaining efficient inference speed.

Key insights

• Implicit policy modeling with diffusion models effectively captures the stochasticity and multiple valid trajectories in expert surgical actions, outperforming explicit policy approaches.
• Incorporating geometric equivariance into neural networks significantly improves model generalization and robustness to varying camera orientations and instrument movements in surgical videos.
• A dedicated conditional sampling strategy, leveraging forward-diffusion guidance, is crucial for accurate inference when using unconditionally trained diffusion policies for state-conditioned tasks.

Summary by alphaXiv.