Voxelwise white-matter compositional connectivity mapping from diffusion MRI via reliability-aware hierarchical inference

Diffusion MRI tractography enables non-invasive mapping of brain connectivity but is commonly summarized as region-to-region graphs, obscuring the fine-grained organization of white-matter pathways. Voxelwise connection distribution maps (CDMs) preserve this information by representing the relative distribution of tractography-derived connections to anatomical targets at each voxel. However, normalized streamline-count profiles are noisy estimates of latent streamline-target compositions, and existing approaches neither quantify their uncertainty nor use population information to stabilize individual estimates. We propose individualized connection distribution mapping (iCDM), which extends CDM through an isometric log-ratio (ILR) representation, hierarchical empirical-Bayes inference of latent individual compositions, and quantification of inferential uncertainty and individual-population support. Individual streamline counts define the likelihood, while a spatially varying population prior is estimated from group data; its precision reflects between-subject dispersion, and its mean can be conditioned on independently learned subject characteristics. Simulations showed lower estimation error than naive and Dirichlet-multinomial estimates, with additional benefit from characteristic-conditioned priors when the characteristic explained substantial variation. Compositions remained similar under within-tractogram split-half sampling and tested changes in seeding and streamline budget. In independent scan-rescan data, data-information maps were reproducible, and greater data information was associated with lower compositional divergence. In a pediatric cohort, iCDM detected age-related compositional variation at voxel and tract levels. In tumor patients, posterior support shifted from predominantly population-based in lesion cores toward greater individual-data support with distance from the lesion. iCDM provides uncertainty-aware voxelwise connectivity inference for population-level and individualized analyses in basic and clinical research.