BAMPolyp: Bi-Axial Mamba Bottleneck for Gastrointestinal Polyp Segmentation

Md. Farhadul Islam, Tashik Ahmed, Partho Chanda, Joyanta Jyoti Mondal, Meem Arafat Manab, Sarah Zabeen, Jannatun Noor; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 1082-1092

Accurate gastrointestinal (GI) polyp segmentation demands both fine boundary precision and long-range context, which many existing models fail to balance efficiently. Current bottleneck designs often lose spatial continuity when modeling global dependencies. To address this, we propose BAMPolyp, a novel segmentation framework that introduces a Bi-Axial Mamba bottleneck into a U-Net-style architecture. Unlike conventional sequence models that flatten spatial dimensions and operate globally, Bi-Axial Mamba applies state-space mixing along height and width axes independently. This enables each spatial token to capture global dependencies along one axis while preserving local spatial continuity along the other, effectively enhancing both boundary sensitivity and contextual coherence, critical for distinguishing polyps from mucosal folds. An EfficientNet-B0 encoder and deep supervision decoder ensure strong hierarchical feature extraction and optimization. Ablation studies confirm that axis-wise decomposition improves localization accuracy without compromising semantic understanding. Our results establish BAMPolyp as an efficient solution for polyp segmentation, achieving 0.9380/0.8881 (Dice/IoU) on Kvasir-SEG, 0.9437/0.8939 on CVC-ClinicDB, 0.9255/0.8659 on CVC-ColonDB, and 0.8683/0.8211 on PolypGen with minimal computational overhead. The code for the proposed BAMPolyp is available at https://github.com/farhad324/BAMPolyp.