Automated tool counts teeth from routine brain MRI
The model counted teeth with a mean absolute error of 1.96 and matched manual associations with cognitive scores in a 233-person cohort.
medRxiv
In human brain MRI scans from the OASIS-3 study, researchers developed an automated pipeline to extract dental information from routine neuroimaging. The tool, called MOLAR, isolates the dental region from T1-weighted scans and counts teeth using a three-dimensional nnU-Net. Across 1,261 human scans annotated by three raters, the model identified teeth with an F1 score of 0.873, compared with 0.787 between human raters, and had a mean absolute error of 1.96 teeth. A second network located the dental region directly, avoiding classical registration failures that occurred in 18% of scans. In an independent cohort of 233 people, automated tooth counts reproduced the manual correlation with Montreal Cognitive Assessment scores (r = +0.372 versus +0.414, both p < 0.001), remaining equivalent after adjusting for age, sex, and race.
Why it matters
Tooth loss correlates with cognitive decline in older adults, but large neuroimaging datasets rarely record dental status. This approach allows researchers to study links between oral health and brain aging opportunistically from existing scans without additional clinical visits.
Caveats
This work is a preprint and has not yet undergone peer review. The automated method still carries an average error of nearly two teeth compared to expert manual counts.
The paper
MOLAR: MRI-based Opportunistic Localization and Recognition of teeth
University of South Carolina
medRxiv · 21 Sep 2026 · Preprint, not peer-reviewed