SGS-Mamba: Semantic Gap-Sensitive Mamba Network for Medical Small Lesion Segmentation

Accurate segmentation of small lesions is critical for the early diagnosis and intervention of severe diseases. U-shaped networks (U-Nets) are widely used models for lesion segmentation, renowned for their strong feature extraction capabilities and encoder–decoder architecture, which incorporates direct skip connection (DSC) to preserve spatial details. However, their decoders often exhibit information-specific biases toward the content provided by DSC due to the semantic gap between the encoder-decoder feature sets, rather than maintaining a strict one-to-one correspondence. This semantic mismatch reduces their effectiveness in segmenting small lesions with weak or indistinct features, thereby failing to meet high-precision clinical requirements. To address this limitation, we propose the Semantic Gap-Sensitive Mamba U-shaped Network (SGS-Mamba), enhancing sensitivity to small lesions across three dimensions: long-range interaction, guided filtering and semantic alignment. Specifically, an adaptive Visual State Space block is designed to focus on small lesions by leveraging long-range contextual modeling. A Self-Guided Filtering Attention (SGFA) at the bottleneck is employed to enhance small lesions through noise suppression and attention focus. The Semantic Gap-sensitive Skip Connection (SGSC), incorporating Cross-Guided Filtering module and Rényi divergence regularization, are designed to bridge the semantic gap through long-range interaction and semantic alignment. Experiments conducted on three small lesion datasets and one multi-label dataset demonstrate state-of-the-art performance, while maintaining competitive computational efficiency. Furthermore, when the SGFA and SGSC modules are integrated as plug-and-play components within other architectures, they consistently yield substantial improvements, underscoring that their benefits generalize effectively across diverse architectures and tasks.