Quantitative Measurement of White Matter Hyperintensities in Brain MRI using a Multimodal Boundary-Aware Dual-Task Segmentation Framework

White matter hyperintensity (WMH) subtype segmentation in brain magnetic resonance imaging (MRI) remains challenging due to heterogeneous lesion sizes, indistinct boundaries, and severe class imbalance. We propose a multimodal boundary-aware dual-task U-Net for segmenting periventricular (PVWMH), deep (DWMH), and subcortical (SWMH) lesions. The framework integrates FLAIR–T1-weighted feature fusion, Vision Mamba-based feature modeling, an SWMH-specific boundary loss, and auxiliary boundary detection. The method was evaluated using patient-level five-fold cross-validation on a private multicenter cohort (n = 75) and separately on 60 publicly annotated MICCAI WMH Challenge subjects for binary segmentation. On the private cohort, the proposed framework achieved an overall Dice similarity coefficient (DSC) of 0.6081 and a 95th-percentile Hausdorff distance (HD95) of 14.80 mm, compared with 0.5638 and 18.25 mm for the FLAIR-only ablation baseline. It also improved SWMH DSC from 0.4610 with nnU-Net to 0.5215 (p = 0.008). On the public dataset, the mean binary DSC and HD95 were 0.6812 and 12.91 mm, respectively.Given the relatively large through-plane voxel spacing (8.4 mm), the framework processes 2D axial slices independently and does not explicitly model inter-slice context, which may limit its ability to characterize lesions extending across adjacent slices.