Scientists have built an AI tool that reads a brain MRI and maps how quickly different regions are aging — a more detailed picture than a single brain-age number, and one that could help study dementia.
Researchers at the University of Southern California’s Leonard Davis School of Gerontology, led by associate professor Andrei Irimia, trained a deep-learning neural network on MRI scans from 14,748 cognitively normal adults aged 19 to 100, drawn from six public datasets including the UK Biobank. Published in PNAS.
What brain age means
The underlying idea is to train a model to predict a person’s chronological age from their brain scan alone, using only healthy individuals.
Applied to a new person, the model outputs the age their brain resembles. The difference from their actual age — the brain-age gap — is taken as a measure of whether the brain is aging faster or slower than expected.
The measure has been associated in prior work with cognitive decline, dementia risk and mortality. Its limitation is that it compresses an entire brain into one figure, which is where this work departs.
Why one number is insufficient
Rather than assigning one number to the whole brain, the model measures local brain age down to the voxel — a three-dimensional pixel — revealing which regions look biologically older than others.
The reason that matters is that neurodegenerative diseases are regionally specific. Alzheimer’s begins in particular memory-related structures. Frontotemporal dementia affects frontal and temporal regions. Parkinson’s involves distinct areas again.
A whole-brain average of a localised process is dominated by the unaffected majority, so early regional change is diluted below detection. Measuring locally preserves the signal that a global figure discards.
What normal aging looked like
Across participants, the frontal and temporal lobes appeared older than the parietal and occipital regions, and the right hemisphere aged slightly faster than the left.
The frontal-temporal finding accords with existing knowledge: those regions show the greatest volume loss with age and support the executive function and memory abilities that decline first in ordinary aging.
They are also, in evolutionary and developmental terms, the last to mature — frontal regions continue developing into the third decade of life. A pattern in which the last to mature is the first to decline recurs across neuroscience, though the reasons remain debated.
The hemispheric asymmetry is less easily explained and is the kind of observation that needs replication before being interpreted.
The dementia link
In people with mild cognitive impairment and Alzheimer’s disease, the model found significantly older local brain ages in structures neurodegeneration hits early — including the hippocampus and amygdala and other deep memory-related regions.
That result functions as validation rather than discovery. The regions identified are precisely those known to be affected early in Alzheimer’s, so finding them confirms the method detects real pathology rather than noise.
A model trained only on healthy brains that nonetheless flags the right structures in disease is doing something meaningful, because it was never shown what disease looks like.
What it could enable
The researchers say the approach could eventually help identify dementia earlier, track its progression, and test whether experimental therapies slow aging in specific brain areas.
The trial application may be the most immediately valuable. Neurodegeneration trials struggle for sensitive endpoints: cognitive tests are noisy and influenced by mood, effort and practice, while whole-brain volume changes slowly. A regional imaging measure that changes measurably over a trial period would let smaller studies detect effects, which matters in a field where large trials have repeatedly failed ambiguously.
The caution
The researchers note it needs validation on diverse clinical datasets and long-term follow-up studies before routine care.
Both points are substantive. UK Biobank and similar cohorts skew toward healthier, better-educated and less ethnically diverse participants than the general population, and models can perform worse on groups underrepresented in training.
What an aging map cannot tell an individual
The gap between a research tool and a clinical test is worth stating explicitly, because brain-age measures are already being marketed directly to consumers.
A model of this kind is trained and evaluated on populations, and population-level accuracy does not imply individual reliability. A measure that separates groups cleanly can still carry enough variation that any one person’s value is uncertain.
Scanner differences compound that. MRI measurements vary between machines, sequences and sites, and models trained on research-quality scans from a small number of centres can behave differently on a clinical scan from elsewhere.
Most importantly, an older-looking region is not a diagnosis and does not indicate what to do. There is no intervention established to reverse regional brain aging, so a person told their hippocampus looks older than expected has acquired anxiety rather than actionable information — which is why the researchers position this as a research and trial tool rather than a test to be requested.
Long-term follow-up is needed for a different reason: showing that an older-looking region predicts future decline requires waiting to see who declines. Cross-sectional agreement with existing diagnoses does not establish predictive value. Research news, not medical advice.