Quantifying cross-site effect in AI-based dental age estimation: evidence from Brazilian panoramic radiographs

To quantify cross-regional generalization of dental age-estimation models, identify practical strategies to improve their performance, and report uncertainty in a transparent manner.

A total of 21 722 panoramic radiographs from 2 Brazilian regions were acquired using distinct equipment. A robust Inception-v4 model was evaluated under 4 scenarios: (1) training on Northeast data and testing on Southeast data; (2) fine-tuning using Southeast data only or both regions; (3) training from scratch on pooled data; (4) pooled training with augmentation. Model performance was assessed using mean absolute error (MAE), mean signed error (bias), R2, Bland-Altman analysis, and calibration metrics.

A marked performance drop was observed when the model trained on Northeast data was applied to Southeast radiographs (MAE 4.97 years vs 3.10 years in-region), with negative bias and wider Bland-Altman limits at older ages. Training with pooled regions and modest fine-tuning improved accuracy and calibration across both cohorts (MAE 3.24-3.69; R2 0.93-0.95). Data augmentation yielded only small additional improvements and did not eliminate large residual errors. Heatmaps highlighted clinically relevant anatomical structures commonly used by dental experts for age estimation.

Cross-site and domain shifts significantly impact the performance of AI models for dental age estimation. Multi-regional training combined with light model adaptation provides robust, well-calibrated, and interpretable results across regions, whereas data augmentation alone has limited effectiveness. This study offers a 2-region benchmark, code, and data-access protocols to support reproducible evaluation and guide clinical deployment.