Abstract:Skin lesion segmentation plays a crucial role in intelligent diagnosis and therapeutic assessment. However, it remains challenging due to the complex lesion morphology, blurry boundaries, and low contrast between lesions and surrounding skin. To address these challenges, we propose an efficient hybrid-architecture segmentation model, termed FECNet (feature enhancement and contrastive semantic network). FECNet integrates several key modules to enhance structural representation, boundary modeling, and semantic discrimination in complex lesion scenarios. A feature enhancement module strengthens the modeling of multi-level local textures and global structures. The multi-scale feature fusion module aggregates enhanced features across different semantic levels, forming a more comprehensive and structurally consistent representation, effectively handling large variations in lesion appearance. The feature decoupling module explicitly separates foreground, background, and uncertainty regions to improve boundary coherence. Furthermore, a contrastive semantic context modulation module dynamically captures semantic discrepancies, enhancing foreground activation while suppressing background interference, thereby improving discriminative ability in low-contrast or visually confusing cases. Experimental results on multiple public datasets demonstrate that FECNet achieves state-of-the-art performance, showing superior segmentation accuracy and robustness, especially on images with fuzzy structures and low contrast.