MambaPolyp : Gated State‐Space Modeling With Multi‐Scale Context and Boundary‐Refinement for Colorectal Polyp Segmentation

Abstract

Accurate segmentation of colorectal polyps in colonoscopy images is crucial for the early detection and prevention of colorectal cancer (CRC). However, polyp segmentation remains challenging due to variations in size, shape, texture, and boundary ambiguity caused by specular highlights, motion blur, and low contrast. In this paper, we propose MambaPolyp, a novel encoder‐decoder architecture that integrates efficient long‐sequence modeling, multi‐scale context aggregation, and fine boundary enhancement for precise colorectal polyp segmentation. The core component of our design is the MambaGate module, which leverages dynamic state‐space models (SSMs) to capture long‐range dependencies in spatial features while adaptively gating irrelevant context. To enhance multi‐scale semantic understanding, we incorporate an atrous spatial pyramid pooling (ASPP) module at the bottleneck, enabling the model to aggregate features from varying receptive fields. The decoder is equipped with a Multi‐Mamba refinement mechanism, where stacked MambaBlocks progressively refine the fused features at each resolution level. A Boundary‐Refinement Module is introduced to sharpen the polyp edges and improve boundary localization. Furthermore, a deep supervision strategy is adopted across multiple decoder stages to guide training with rich hierarchical information. The model is evaluated on five benchmark datasets, Kvasir‐SEG, CVC‐ClinicDB, CVC‐ColonDB, ETIS‐Larib, and CVC‐300, demonstrating that MambaPolyp outperforms several state‐of‐the‐art (SOTA) segmentation models in terms of Dice coefficient, Intersection over Union (IoU), and boundary‐based metrics, while maintaining lower computational efficiency, making it suitable for real‐time clinical applications during colonoscopy.