Automated whole-dentition CBCT phenotyping enables explainable sex classification and age estimation through distinct dental features.
Authors
Affiliations (7)
Affiliations (7)
- Key Laboratory of Shaanxi Province for Craniofacial Precision Medicine Research, College of Stomatology, Xi'an Jiaotong University, No. 98 XiWu Road, Xi'an, Shaanxi, 710004, China.
- Department of Orthodontics, College of Stomatology, Xi'an Jiaotong University, No. 98 XiWu Road, Xi'an, Shaanxi, 710004, China.
- Department of Forensic Medicine, Faculty of Basic Medical Science, Chongqing Medical University, Chongqing, 400016, China.
- Health Science Center, Xi'an Jiaotong University, Xi'an, Shaanxi, 710061, China.
- Key Laboratory of Shaanxi Province for Craniofacial Precision Medicine Research, College of Stomatology, Xi'an Jiaotong University, No. 98 XiWu Road, Xi'an, Shaanxi, 710004, China. [email protected].
- Department of Orthodontics, College of Stomatology, Xi'an Jiaotong University, No. 98 XiWu Road, Xi'an, Shaanxi, 710004, China. [email protected].
- Key Laboratory of Shaanxi Province for Craniofacial Precision Medicine Research, College of Stomatology, Xi'an Jiaotong University, No. 98 XiWu Road, Xi'an, Shaanxi, 710004, China. [email protected].
Abstract
Tooth size and morphology are established primarily during development, whereas secondary dentin deposition and pulp-space remodeling continue throughout adulthood, providing biologically distinct information for sex and age estimation. This study investigated whether automated whole-dentition cone-beam computed tomography (CBCT) analysis could capture these two signals. The initial cohort comprised 327 adult CBCT scans, of which 275 passed automated segmentation-completeness screening and entered the final analysis (141 female and 134 male participants; age range, 21-61 years). ToothSeg was used for automatic tooth segmentation and numbering. After excluding variables potentially reflecting image acquisition rather than dental biology, extreme gradient boosting (XGBoost) models were developed separately for sex classification and age regression. Performance was assessed using five-fold out-of-fold predictions, and feature importance was interpreted using SHAP. The sex model achieved a balanced accuracy of 0.807 and an AUC of 0.872, while the age model achieved a mean absolute error of 5.29 years and an R² of 0.645. Lower-right canine volume was the most important individual feature for sex classification and was greater in males (Cohen's d = 1.20). Mean incisor relative low-gray voxel fraction was the most important age feature and was inversely correlated with chronological age (Spearman ρ = -0.706). As an exploratory study, automated whole-dentition CBCT phenotyping showed internally predictive, biologically interpretable patterns for adult sex and age estimation. However, given the modest sample size and proxy low-gray metric, multicenter external and anatomical validation are required.