Generalizable Quantitative Bone Age Assessment From Three-Dimensional Cervical Vertebral Morphology: A Multi-Institutional Study

Introduction and Aims Bone age assessment (BAA) is essential for evaluating skeletal maturity and guiding growth-related treatment. Conventional methods rely heavily on expert assessment and show limited generalizability. This study aimed to develop and validate a quantitative BAA approach using 3-dimensional cervical vertebral morphology, focusing on transferability across diverse attributes and demographic populations. Methods This multi-centre retrospective study analyzed 702 cone-beam computed tomography (CBCT) images from 5 Chinese institutions, including patients under 19 who underwent both hand-wrist and CBCT imaging within 30 days (2022-2025). Twenty-two 3D cervical vertebral parameters were extracted as input features for 8 machine learning (ML) models. Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), and explained variance. External validation, subgroup analyses across demographic and imaging parameters, and comparisons with junior clinicians were performed. Feature contributions were interpreted using SHAP analysis. Results The dataset included 355 individuals for training, 182 for internal validation, and 165 for external testing (11.78 years ±2.41 [SD]). In the internal validation cohort, CatBoost and Random Forest (RF) demonstrated the best performance with accuracies of 99.45% and 96.69%, respectively. In the external validation cohort, RF and CatBoost maintained superior predictive ability (both 91.97% accuracy, R² = 0.96). CatBoost consistently outperformed the other models across demographic and equipment-based groups, with significant differences from junior clinicians’ predictions ( P < .05). SHAP analysis highlighted the key features of the anterior height of the third vertebral body and the posterior height of the fourth vertebral body. Conclusion The 3D ML model provides a scalable, reliable BAA solution with high accuracy and generalizability, reducing the need for expert assessments and enabling widespread adoption. Clinical Relevance The proposed approach enables objective bone age assessment from existing CBCT scans, supporting consistent orthodontic growth evaluation and treatment timing across diverse clinical settings.