Interpretable deep learning reveals spatiotemporal MRI features of brain aging that align with neurodegeneration.

Cortical thinning and atrophy are hallmarks of brain aging that have been characterized using magnetic resonance imaging (MRI). Brain aging involves many neuroanatomic features whose effects on brain structure remain unexplored. To address this challenge, we trained interpretable deep neural networks (DNNs) to estimate brain age (BA) from T1-weighted (T1w) MRI. By identifying MRI features unapparent to humans, DNNs can find aging-related structural alterations above and beyond cortical thickness and atrophy. Using a novel approach to DNN interpretability, we mapped brain aging progression in 25,539 cognitively normal UK Biobank adults aged 45-83 years. Cortical aging is found to involve anatomic features becoming prominent during the 50s within frontolateral, mesolimbic, cuneal, and occipitotemporal regions. From these foci, aging-related features propagate to adjacent areas at rates peaking in the 60s. Cortical thinning and atrophy do not trend closely with neurodegeneration, but important DNN-identifiable anatomic features have spatiotemporal dynamics that match those of amyloid or tau. Our results challenge the assumption that MRI cannot map anatomic features trending with neurodegeneration. Interpretable DNNs can empower MRI to quantify anatomic aging as a process of spatial feature expansion from focal regions into nearby structures in the sequence of neurodegenerative pathology. Traditional morphometrics explain only ∼1% of variance in DNN-identifiable features, which clarifies why the former are insensitive to anatomic changes involving neurodegeneration. Our results conceptualize brain aging in the context of spatial and temporal parallels between anatomic senescence and neuropathology. These findings may help to map cognitively normal adults' neurodegenerative anatomy even without PET measurements.