Structural-Prior Attention Network for Region-Aware WMH Segmentation

White matter hyperintensities (WMH) exhibit distinct spatial patterns, with differing prevalence and morphology in periventricular and deep regions. SPAN-WMH incorporates anatomical knowledge by generating probabilistic structural priors that reflect these region-specific distributions and embedding them into a spatially adaptive attention mechanism. The network modulates feature responses according to the anatomical context, enabling improved discrimination of subtle, region-dependent lesion patterns. Experiments on 1,100 subjects from WMH2020 show substantial improvements: deep-WMH Dice increases from 0.643 to 0.742 (+15.4%), periventricular Dice rises from 0.821 to 0.884 (+7.7%), and HD95 decreases by 28.3%. Small-region IoU improves by 12.1%, highlighting better detection of scattered deep lesions.