CAM-PoEm: Contour-Guided State-Space Modeling for Boundary-Aware Image Segmentation

Accurate gastrointestinal polyp segmentation is important for early colorectal cancer detection but remains challenging due to substantial morphological variation and ambiguous boundaries. While selective state-space models such as Mamba provide efficient long-range modeling, they lack explicit structural inductive biases for fine-grained boundary localization. We propose CAM-PoEm, an encoder–decoder architecture centered on a Contour-Aware Mamba (CAM) bottleneck. CAM incorporates contour information specifically through the state-to-output readout of the state-space model rather than modifying its recurrent state dynamics. A lightweight Learnable Contour Extractor combines Sobel gradients with learnable refinement to generate structural priors, while Multi-Kernel Positional Embedding provides complementary multi-scale local spatial encoding. CAM-PoEm achieves Dice scores of 91.65%, 91.50%, 88.31%, and 88.27% on Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and PolypGen2021, respectively, using 4.96M parameters. In a separate controlled three-run ablation on Kvasir-SEG, replacing standard Mamba with CAM improves Dice from 0.8873 ± 0.0044 to 0.9161 ± 0.0037. Further conditioning experiments show that applying contour priors specifically to the readout outperforms conditioning other state-space parameters. These findings provide evidence that structural priors can introduce boundary sensitivity at the readout stage while preserving efficient recurrent global modeling, motivating readout-conditioned state-space models for structure-aware dense prediction.