Mamba-Constrained Inter-Slice Consistency Network for Stroke Lesion Segmentation

Conventional 2D or pseudo-3D models often produce anatomically inconsistent stroke lesion masks across slices due to the lack of explicit inter-slice dependency modeling. MSC-Mamba addresses this problem by introducing a bidirectional Mamba backbone tailored to capture long-range interactions along the slice axis. The model encourages coherent lesion boundaries across adjacent slices and reduces abrupt fluctuations caused by noise or low contrast. A topology-aware decoder further improves structural continuity and prevents fragmentation. Experiments on ATLAS v2 (655 subjects; 520 for training and 135 for testing) show that MSC-Mamba achieves a Dice score of 0.856, outperforming nnUNet (0.806, +6.2%) and TransBTS (0.821, +4.3%). HD95 is reduced from 17.1 mm to 10.5 mm (−38.6%), and slice-to-slice contour variance decreases by 19.4%. On a low-contrast subset of 182 cases, MSC-Mamba maintains a Dice of 0.812, which is 9.1% higher than TransBTS. Testing on a secondary clinical cohort reveals a 7.6% Dice improvement and a 13.2% reduction in lesion fragmentation.