MS ‐ FSSM : A Multi‐Scale Frequency‐Sensitive State Space Model for Brain Tumor Segmentation in Multi‐Modal MRI

Accurate automatic segmentation of brain tumors is clinically essential, yet remains challenging due to their irregular morphology, ill‐defined boundaries, and weak contrast in enhancing regions. To alleviate these challenges, we propose MS‐FSSM (Multi‐Scale Frequency‐Sensitive State Space Model). The architecture features three decoupled, parallel encoding streams—convolutional neural network (CNN), vision transformer (ViT), and state space model (SSM)—designed respectively to capture local texture, global semantic context, and diffuse long‐range structural dependencies. Following multi‐scale feature fusion, a frequency‐aware enhancement module is introduced to recover high‐frequency boundary details. In the decoding stage, semantic‐gated skip connections suppress background interference, while a state‐space refinement module integrated at the mid‐resolution level helps improve segmentation completeness and boundary continuity of tumor regions. Evaluated on two public benchmarks—the Medical Segmentation Decathlon (MSD) Task01 and BraTS 2025 Task02—MS‐FSSM achieves 0.87 Dice on enhancing tumor (ET) in both glioma (MSD) and meningioma (BraTS), with Surface Dice of 0.91 in MSD. Meanwhile, it only requires 11.36 M parameters and 2.06 GFLOPs. Experimental results indicate that our method achieves a favorable trade‐off between segmentation accuracy and computational efficiency, offering a promising pathway toward lightweight, clinically deployable brain tumor segmentation systems.