Meta-weighted Diffusion Model for Reliable Online Surgical Phase Recognition

Long Tian, Yufei Li, Jirui Wu, Ruiying Lu, Xiaonan Liu, Quan Wang, Liming Wang, Xiyang Liu

Keywords: surgical phase recognition, diffusion model, meta learning

Abstract:

Surgical phase recognition has drawn great attention most recently thanks to its potential downstream applications closely related to human life and health. Despite deep network-based models have made significant advancement in capturing discriminative long-term dependency of surgical videos to achieve improved recognition, they seldom account for exploring and modeling uncertainty of surgical videos, which should be crucial for reliable surgical phase recognition. we categorize the sources of uncertainty into two types, imbalanced phase distribution and low-quality image acquisition, which are inevitable in surgical videos. To address this pivot issue, we introduce a meta-weighted diffusion model (MetaDiff) to take full advantages of meta-learning and deep generative model in tackling uncertainty. For uncertainty caused by image quality, we present a classifier-guided diffusion model to produce countable denoised recognition results, making it possible to measure uncertainty using statistical tools for each video frame. For uncertainty caused by phase distribution, we propose a meta-weighted objective function to optimize the classifier-guided diffusion model, making the classification boundary robust against surgical video uncertainty.
We demonstrate outstanding ability of our model through comprehensive benchmarks on Cholec80, AutoLaparo, M2Cai16, and CATARACTS. Experimental results reveal that MetaDiff significantly outperforms state-of-the-art methods, separately achieving accuracies of $95.3\%$, $85.8\%$, $92.2\%$, and $85.1\%$ on Cholec80, AutoLaparo, M2Cai16, and CATARACTS.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8858