Video Polyp Segmentation Using Implicit Networks

Abstract

Polyp segmentation in endoscopic videos is an essential task in medical image and video analysis, requiring pixel-level accuracy to accurately identify and localize polyps within the video sequences. Addressing this task unveils the intricate interplay of dynamic changes in the video and the complexities involved in tracking polyps across frames. Our research presents an innovative approach to effectively meet these challenges that integrates, at test time, a pre-trained image (2D) model with a new form of implicit representation. By leveraging the temporal understanding provided by implicit networks and enhancing it with optical flow-based temporal losses, we significantly enhance the precision and consistency of polyp segmentation across sequential frames. Our proposed framework demonstrates excellent performance across various medical benchmarks and datasets, setting a new standard in video polyp segmentation with high spatial and temporal consistency. Our code is publicly available at https://github.com/AviadDahan/VPS-implicit.