Boundary Refinement Network for Polyp Segmentation With Deformable Attention

Abstract

Early and accurate polyp segmentation is crucial for the diagnosis and treatment of colorectal cancer. However, polyp segmentation faces many challenges: different polyp sizes, complex shapes, and ambiguous intestinal wall boundaries. To solve these problems, we propose a novel polyp segmentation network named DeformSegNet. Specifically, we first introduce a polyp perception module (PPM), which combines the dynamic multi-kernel spatial selection network (DMS-Net) and a transformer encoder to effectively locate polyps of different sizes. Next, we design a deformation-aware separable module (DSM), which consists of deformable attention that adaptively adjusts the sampling position, enabling the network to adapt to complex and diverse polyp boundaries. Finally, a cross-attention aggregation module (CAAM) effectively retains low-level features, further enhancing the boundary features and suppressing false positives. DeformSegNet achieves competitive segmentation accuracy on five polyp datasets, demonstrating excellent learning and generalization capabilities.