MGCBFormer: The multiscale grid-prior and class-inter boundary-aware transformer for polyp segmentation

The polyp segmentation technology based on deep learning could better and faster help doctors diagnose the polyps in the intestinal wall, which are predecessors of colorectal cancer. Mainstream polyp segmentation methods are implemented under full supervision. For these methods, expensive and precious pixel-level labels couldn't be utilized sufficiently, and it's a deviation direction to strengthen the feature expression only using the more powerful backbone network instead of fully mining existing polyp target information. To address the situation, the multiscale grid-prior and class-inter boundary-aware transformer (MGCBFormer) is proposed. MGCBFormer is composed of highly interpretable components: 1) the multiscale grid-prior and nested channel attention block (MGNAB) for seeking the optimal feature expression, 2) the class-inter boundary-aware block (CBB) for focusing on the foreground boundary and fully inhibiting the background boundary by combining the boundary preprocessing strategy, 3) reasonable deep supervision branches and noise filters called the global double-axis association coupler (GDAC). Numerous persuasive experiments are conducted on five public polyp datasets (Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, CVC-300, and ETIS-LaribPolypDB) comparing with twelve methods of polyp segmentation, and demonstrate the superior predictive performance and generalization ability of MGCBFormer over the state-of-the-art polyp segmentation methods.