MOLAR: MRI-based Opportunistic Localization and Recognition of teeth

Tooth loss is associated with cognitive decline, yet dental status is rarely available in large neuroimaging cohorts, although routine T1-weighted (T1w) brain MRI captures the dentition. We present MOLAR (MRI-based Opportunistic Localization And Recognition of teeth), an open pipeline that isolates the dental region (the "bite box") from T1w brain MRI and counts the teeth, without additional scanning or commercial software. Three raters annotated the usable OASIS-3 bite boxes (1,261 rated by all three; 63,829 point markers; count intraclass correlation coefficient, ICC(2,1) = 0.909). We reformulate counting from point labels as instance segmentation: each tooth becomes a compact capsule whose in-plane footprint adapts to voxel adjacency, so a standard three-dimensional nnU-Net counts teeth as connected components. On held-out scans the model detected teeth more consistently than raters agreed with one another (F1 0.873 against 0.787) and, after leave-one-fold-out threshold calibration, counted with mean absolute error 1.96 teeth, recovering 86% of the gap between an uninformed floor (3.57) and the inter-rater ceiling (1.70). A second network replaces registration by predicting the dental region, changing counts by 1.06 teeth; less than two raters disagree (1.58), whereas classical registration failed on 18% of scans. In an independent cohort (Aging Brain Cohort; n = 233; different scanner and raters), automated counts reproduced the manual association with the Montreal Cognitive Assessment (MoCA; Pearson r = +0.372 against +0.414, both p < 0.001), and after adjustment for age, sex, and race, the association was equivalent across counting methods. The software will be released as open source.