Colorectal cancer ranks among the leading causes of cancer-related mortality worldwide, with early polyp detection and accurate segmentation serving as critical interventions for prevention. Despite remarkable progress in deep learning-based polyp segmentation, existing methods continue to face challenges including ambiguous boundaries, diverse morphologies, and scale variations. In this paper, we propose MambaPolyp, a hierarchical network based on Vision Mamba that integrates a similarity aggregation mechanism to achieve high-precision polyp segmentation. MambaPolyp employs a pretrained Vision State Space Model backbone to extract multi-scale features, utilizes a Cascaded Fusion Module to aggregate high-level semantic information, leverages a Hierarchical Scan Mamba (HSM) to enhance shallow detail representation, and employs a Similarity Aggregation Module with windowed cross-attention to intelligently fuse dual-stream features. The HSM incorporates three parallel branches-global long-range, meso-scale regional, and local multi-directional scanning|along with a morphology-guided Region of Interest mechanism to focus on challenging regions. Extensive experiments on multiple benchmark datasets, including Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and ETIS-Larib, demonstrate that MambaPolyp outperforms state-of-the-art methods across key metrics such as mean Dice and mean Intersection over Union while maintaining linear computational complexity. Ablation studies further validate the effectiveness of each module. This work presents an efficient and accurate new paradigm for polyp segmentation.
Xue et al. (Fri,) studied this question.