Multiple Attention and Lesion-aware Dynamic Kernel for Polyp Segmentation

Colorectal cancer ranks among the primary contributors to cancer-related mortality globally. Polyps growing in the colon and rectum are precursors to colorectal cancer, making efficient and accurate detection of polyps in images of great practical significance. However, due to variations in the shape and size of polyps and the low contrast between polyp boundaries and the surrounding intestinal textures, accurate segmentation of polyps becomes challenging. To address these issues, this paper proposes two attention mechanism modules and A scheme for dynamically generating and updating kernels to enhance lesion awareness. Specifically, an efficient self-attention module is used to capture long-range contextual relationships, while a cross-attention module based on the polyp boundary region is employed to perceive differences between the boundary and the surrounding mucosa, along with a dynamic kernel generation and update scheme to obtain global contextual features of the input image. We validated our algorithmic network on two publicly available polyp benchmarks, and the experimental results demonstrate that our method outperforms alternative cutting-edge techniques.