GSC-Mamba: unified global modulation, stochastic refinement and cross-scale gating for medical image segmentation

Medical image segmentation is vital for diagnosis and treatment, yet its accuracy is often constrained by challenges in modeling long-range dependencies, high computational complexity, and poor generalization against image domain shifts. Inspired by Kolmogorov-Arnold Networks (KANs) for interpretability and State-Space Models (SSMs) for efficient sequence handling, we introduce GSC-Mamba, a novel U-shaped network architecture designed to significantly boost segmentation accuracy, robustness, and generalization. GSC-Mamba achieves substantial performance gains through the innovative integration of three core components: (1) Global Appearance Modulation (GAM) to enhance robustness against domain shifts; (2) Stochastic Sequence Refinement (SSR) to boost feature representation generalization; (3) an extensively deployed Cross-Scale Attention Gate (CSAG) to ensure precise, attention-weighted feature transfer between the encoder and decoder. GSC-Mamba significantly surpasses existing state-of-the-art methods across multiple medical image segmentation datasets (ISIC17, ISIC18, CVC, BUSI, and GLAS). This superior performance, achieved with minimal parametric overhead and high computational efficiency, demonstrates GSC-Mamba’s excellent balance of accuracy and computational cost. This study establishes a new, efficient, and highly interpretable paradigm for medical image segmentation, offering innovative solutions to handle complex clinical image variability.