DEAMNet: a dual-branch mamba-CNN encoder with wavelet boundary awareness for colorectal polyp segmentation

To address the challenge of simultaneously capturing global contextual information and local boundary details in colorectal polyp segmentation, this paper introduces a dual-branch edge-aware Mamba network (DEAMNet) built upon the effective integration of established visual recognition components. Drawing upon the prevalent Mamba-CNN hybrid paradigm, a Mamba-CNN dual encoder (MCDE) is constructed, where the state space branch captures global contextual information and the convolutional branch extracts local boundary details, with cross-attention employed to align and complementarily fuse heterogeneous features. Furthermore, by integrating data-driven learned boundaries with prior knowledge-guided wavelet boundaries, a wavelet fusion edge extraction module (WFEEM) is formulated to provide enhanced supervision for boundary segmentation, with boundary-aware cross-layer feature alignment leveraged to improve feature representation. Comparative experiments against nine state-of-the-art methods on five benchmark datasets demonstrate that DEAMNet achieves favorable performance in 13 out of 15 comparisons on both in-domain and out-of-domain datasets, indicating the practical effectiveness of this task-oriented architectural adaptation for colorectal polyp segmentation.