An AI trained on 15,000 MRIs maps how each region of the brain ages

Different parts of the same brain age at different rates, with some regions reading as "older" than the person. A deep learning model trained on roughly 15,000 MRIs of cognitively healthy adults turns that pattern into a map, and when it is run on mild cognitive impairment (MCI) and Alzheimer's scans it lights up the regions that neurodegeneration attacks first.

The work comes from Andrei Irimia, PhD, and colleagues at the University of Southern California's Leonard Davis School of Gerontology, and is published in the journal PNAS under the title "Deep learning maps local brain aging in relation to cognition across human adulthood." The advance is spatial resolution rather than a new architecture: a "local brain age" (LBA) is computed for each region of the brain, so some areas can read as "older" while others stay close to chronological age. The map surfaces regional heterogeneity that whole-brain-age methods collapse.

The authors frame the work as a research instrument, not a clinical test. Brain age is a young biomarker with open questions about what it actually measures; the training cohort is healthy by design and not population-representative; the regional signal correlates with cognition rather than causing it. None of that is a reason to dismiss the map. It is a reason to read it as a way to localize where aging has gone off-script, not as a diagnostic.