Anatomical Interfaces as a Complementary Substrate for Structural MRI Representation in Alzheimer's Disease

Background Structural MRI analysis of Alzheimer’s disease (AD) commonly relies on region-of-interest (ROI) summaries or voxel-level deep networks, neither of which explicitly quantifies the physical interfaces shared by adjacent brain regions. We investigated whether these anatomical interfaces provide structural information complementary to conventional regional measures. Methods T1-weighted MRI was segmented into 79 anatomical regions using a pretrained nnU-Net. Each region was represented as a graph node with six morpho-textural descriptors, while edges were defined only between regions sharing voxel-level face contact and characterized by Structural Contact and Boundary Appearance features. Results Correlation analyses showed generally limited redundancy between node- and interface-level descriptors, supporting partially distinct representational information. Node-controlled analyses further showed that selected interface features remained significantly associated with disease stage after adjustment for the morpho-textural characteristics of both connected regions, indicating that interface-level variation was not fully reducible to regional abnormalities. White matter hypointensity interfaces repeatedly emerged among the retained stage-associated effects. Selected Structural Contact features also showed convergent associations with clinical severity across CDR-SB, ADAS-Cog13, and MMSE. Representation ablation demonstrated that disrupting anatomical adjacency or removing interface attributes reduced classification performance, supporting the contribution of anatomically organized relational information. The full representation also improved multi-class classification relative to conventional ROI-vector and voxel-level baselines, with this advantage preserved on the independent NACC cohort without fine-tuning. Conclusions These findings indicate that anatomical interfaces provide measurable and complementary structural information beyond conventional region-wise morphometry, offering an interpretable framework for representing inter-regional structural relationships in neurodegenerative disease.