FeDNet: Feature Decoupled Network for polyp segmentation from endoscopy images

Early detection and diagnosis of colorectal polyps are critical to the diagnosis and treatment of colorectal cancer. When it comes to polyp segmentation, previous methods have limited benefit whether starting with the global contextual information to maintain the consistency of the information within the polyp or starting with the edge information to refine the prediction results. Therefore, a comprehensive way to obtain good polyp segmentation performance is to optimize both simultaneously. Depending on the above analysis, we explore in this paper how to improve the performance of polyp segmentation by optimizing the body and edge simultaneously. Inspired by the feature decoupled method in Laplacian pyramid, we decouple the input feature into the body and edge feature explicitly in an effective and reasonable way, and subsequently perform targeted optimization by introducing a novel Feature Decoupled Module (FDM). Furthermore, combined with FDM (with only 0.08 m network parameters), our approach can significantly outperform previous state-of-the-art methods, attaining a further improvement over the baseline. Especially, we achieve 92.4% mean Dice on the large-scale Kvasir dataset. Not only that, it also demonstrates strong generalization ability. It obtains the top performance on three datasets, where it achieves 82.3% and 81.0% mean Dice on CVC-ColonDB and ETIS, respectively, far exceeding the competitors. Code will be released.1