Abstract
The heterogeneity of brain aging is a hallmark of neurological and psychiatric disorders, yet machine-learning tools used to characterize this process, including the 'brain age' paradigm, have largely relied on global metrics that lack the specificity to map these complex patterns. Here, we introduce BrainAgeMap, an interpretable deep learning framework that generates fine-grained, voxel-wise maps of brain-predicted age difference (brain-PAD) from T1-weighted magnetic resonance imaging scans. We provide converging lines of evidence for the framework's clinical, prognostic, and neurobiological utility. Disorder-specific topographies of accelerated aging were identified in Alzheimer's disease (AD), frontotemporal dementia, and schizophrenia. Longitudinal analysis of the hippocampus revealed accelerated aging in individuals with progressive versus stable mild cognitive impairment (MCI), demonstrating prognostic value. Regional brain-PAD in the temporal lobe correlated strongly with in vivo tau pathology measured by positron emission tomography in AD, linking the maps to underlying molecular pathology. Furthermore, regional brain aging in MCI and AD was linked to individual differences in episodic memory function. BrainAgeMap provides a robust tool to delineate disease-specific pathways of neurodegeneration, offering new opportunities for early diagnosis, patient stratification, and monitoring therapeutic interventions.