Adaptive multimodal fusion via Gated Parallel Mamba architecture for ultra-high-precision stroke lesion segmentation in medical imaging

The accurate delineation of ischemic stroke lesions in magnetic resonance imaging (MRI) is impeded by heterogeneous lesion morphology and the computational expense of modeling global context in three‑dimensional data. In cerebral infarction assessment, diffusion‑weighted imaging, apparent diffusion coefficient, T2‑weighted imaging and T2star sequences (including susceptibility weighted image processing) each offer complementary information, yet existing fusion strategies often fail to adapt to missing modalities or capture long‑range dependencies efficiently. Here we present GPMNet, a lightweight convolutional framework that integrates an adaptive multimodal feature fusion module-employing dynamic cross‑attention to spatially weight and merge signals from all four MRI sequences-and a gated parallel state‑space module that models global voxel interactions in linear time via dual gated branches. We trained the network end-to-end on the ATLAS R2.0 dataset and our own dataset collected at HuanHu Hospital (Tianjin, China), labeled as HHD. The training used a combined Dice-binary cross-entropy and TOPK10 loss, and the outputs were refined using ensemble inference and connected-domain filtering. GPMNet achieved Dice coefficients of 0.6604 and 0.7171 on the two cohorts respectively, achieving superior results compared to other state-of-the-art algorithms. Moreover, the Grad-CAM-based interpretability analysis confirms that the model's attention corresponds to true ischemic areas across modalities, offering visual evidence of its diagnostic reliability and enhancing the transparency of the segmentation process. Our approach delivered rapid, high‑precision stroke segmentation and establishes a scalable paradigm for resource‑efficient clinical imaging applications.