SAFE-Net: Shape-aware and feature enhancement network for polyp segmentation

Precise identification of colon polyps, which are the principal precursors of colorectal cancer, is crucial for accurate diagnosis and treatment. Current segmentation methods possess unique strengths and limitations, given the colon polyp's varying size, shape, color, and unclear edge with surrounding mucosa. These challenges result in the absence of a universally effective segmentation method. Addressing the aforementioned issues, this paper proposes a novel shape-aware and feature enhancement network (SAFE-Net) for polyp segmentation. SAFE-Net contains three innovative modules: the shape-aware module (SAM), the feature enhancement module (FEM), and the multi-modality attention module (MMAM). SAM leverages the disparities in low-level features within the backbone network to effectively filter out noise from colonoscopy image, thereby extracting distinct detailed features, including shape, texture, and edge, ultimately enhancing the precision of polyp segmentation. FEM enhances and enriches the semantic information of the backbone network's features, enables the capture of polyps varying in shape and size, and selects and refines relevant features. MMAM utilizes high-level feature prediction as a guide map to achieve attention on the background, foreground, and boundary, making the network focus more on suspicious and complex polyp regions. This paper conducts experiments on five colonoscopy image datasets: Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, CVC-T, and ETIS. Comparative analyses are conducted between the proposed model and existing models. The experimental results confirm that the proposed model exhibits superior segmentation capabilities and provides a reliable foundation for subsequent diagnostic procedures.