M-Mamba^S: Multimodal Mamba for small lesion segmentation

Precise segmentation of small lesions is crucial for early disease diagnosis and timely clinical intervention. While multimodal imaging provides complementary information to enhance detection, achieving deep cross-modal feature interaction and effective fusion remains challenging. This difficulty stems from inter-modal discrepancies and the subtle structural characteristics of lesions, which are often obscured in individual imaging sources. To tackle the challenge, we propose M-Mamba^S, a novel multimodal Mamba architecture that incorporates three key innovations to advance small lesion segmentation. First, we introduce the VSSBlock^S module, which enhances the original VSSBlock with three specialized designs to effectively capture fine-grained small lesion features in complex imaging scenarios. Second, we introduce the Differential Feature Attention (DFA) module, a novel cross-modal interaction mechanism specifically designed to improve the segmentation of small lesions. By incorporating a multi-level feature interaction mechanism coupled with differential attention, DFA strengthens cross-modal feature synergy. This enhanced interaction enables the effective capture of complementary diagnostic cues, thereby optimizing the extraction of subtle yet critical lesion features. Finally, we present the Cross-Coupled State Fusion (CCSF) module. It leverages a cross-coupled state transition mechanism to align and fuse multimodal features, ensuring consistent propagation and reducing modality-specific noise for robust small lesion segmentation. The module also incorporates low-rank decomposition to suppress artifacts while preserving critical features, thereby enhancing multimodal integration and segmentation precision. Extensive experiments show that M-Mamba^S achieves state-of-the-art performance in segmenting small lesions, improving mIoU by 18.52% (from 31.14% to 49.66%) on ISLES2020, 14.93% (from 54.28% to 69.21%) on MM-WHS2017, and 7.59% (from 61.58% to 69.17%) on our Lymph dataset. Code: https://github.com/Erin668/M-Mamba.