Mamba-EdgeNet: Learnable Edge-Guided State Space Model for Skin Lesion Segmentation

Skin lesion segmentation is one of the key tasks in computer-aided diagnosis. However, accurate boundary delineation remains challenging because of low contrast, irregular boundaries, and visual artifacts in skin lesion areas. To address these problems, we developed Mamba-EdgeNet. Mamba-EdgeNet is a hybrid encoder-decoder framework that selectively integrates ResNet50 with EfficientMamba2D state-space blocks. A resolution-adaptive scanning strategy dynamically alternates between multi-directional and sequential spatial scanning, enabling long-range contextual modeling with linear complexity relative to sequence length. To delineate fuzzy borders, a learnable edge-guidance branch is constructed. Structural priors are generated using a learnable edge enhancer and a shallow edge extractor and are subsequently refined through interactions with deep semantic features. A gated fusion decoder adaptively integrates these edge-aware cues with hierarchical semantic features to resolve feature ambiguity and restore fine-grained contours. Evaluations on the official ISIC 2016 and ISIC 2018 splits, together with an internal evaluation using a non-standard custom partition of ISIC 2017, indicate that the proposed architecture achieves competitive segmentation performance relative to recent representative models across multiple evaluation metrics. Controlled ablation studies further support the contributions of the functional components and loss supervision strategies. Additionally, in-domain evaluation on PH2 and bidirectional cross-dataset transfer between PH2 and ISIC 2018 provide complementary assessments of segmentation performance in within-dataset and cross-dataset settings. These findings indicate that Mamba-EdgeNet is a competitive approach for automated skin lesion segmentation with improved boundary representation.