AxialFuse-Mamba: boundary-oriented axial state-space fusion for multimodal brain tumor segmentation

Accurate multimodal brain tumor segmentation must preserve local boundary information while incorporating long-range context across MRI sequences. Balancing local boundary localization with long-range contextual modeling remains challenging in 3D multimodal MRI. We present AxialFuse-Mamba, a 3D framework combining modality-specific local encoding, axial state-space modeling, spatially disentangled multimodal fusion, and boundary-oriented decoding. We evaluated the framework on held-out BraTS2021 and MSD BrainTumour test splits against CNN–, Transformer-, and Mamba-based baselines, including nnU-Net v2 and SegMamba. We assessed overlap and boundary-related performance using Dice, HD95, average symmetric surface distance (ASSD), Surface Dice, and case-wise analyses. On BraTS2021, AxialFuse-Mamba achieved a mean Dice of 89.979 % and a mean HD95 of 2.206 mm. On MSD BrainTumour, mean Dice was 82.367 % and mean HD95 was 2.559 mm. nnU-Net v2 achieved the highest mean Dice, whereas AxialFuse-Mamba obtained the lowest observed mean HD95. Complementary metrics showed region- and metric-dependent patterns across the evaluated methods. Missing-modality, threshold-sensitivity, cumulative ablation, and difficult-case analyses characterized model behavior and limitations. These findings position AxialFuse-Mamba as a boundary-oriented framework within the evaluated public BraTS-like settings. Source code for the evaluated framework is publicly available at https://github.com/Kuuga4/AxialFuse-Mamba .