Voxelwise correlation of neighbors as a hypothesis driven framework for characterizing white matter lesion heterogeneity

White matter hyperintensities (WMH) on T2-weighted brain magnetic resonance imaging (MRI) are common in aging and associated with small vessel cerebrovascular disease. Standard segmentation methods treat these lesions as uniform binary entities, fundamentally reducing WMH signal by flattening a complex spectrum of tissue damage into a single label. Most WMH methods threshold voxel intensities to estimate lesion volume, missing richer characterization achievable by combining fluid-attenuated inversion recovery (FLAIR) with diffusion MRI. We introduce Voxel-wise Correlation of Neighbors (VCON), a cross-modal framework that quantifies voxel-level relationships between intensity values on T2-weighted FLAIR scans and fractional anisotropy (FA) on diffusion MRI within individuals. VCON generates hypothesis-driven WMH labels by identifying regions where increased FLAIR signal is negatively correlated with FA, suggesting underlying microstructural damage. Using MRI data from over 2,500 participants in community-based aging cohorts, we validated VCON through multi-scale analysis, age-association modeling, scanner comparisons, and intensity-based clustering of WMH into spatially coherent zones with distinct microstructural profiles. VCON revealed a gradient of WMH signal variation that tracks with age and diffusion metrics across scanners and segmentation methods. These results demonstrate that binary WMH masks may obscure clinically important variation in lesion characteristics. VCON reframes lesion segmentation as characterizing microstructural heterogeneity, offering additional structure-informed characterization beyond conventional binary methods by leveraging multimodal MRI signal variation.