Classifying legal age of majority (≥18 years) from panoramic radiographs with transfer learning: Benchmarking ViT and...

To assess whether transfer-learning models applied to panoramic radiographs (PANs) can classify individuals at the threshold of legal majority (≥18 years).

Vision Transformer (ViT) and EfficientNetV2 models were trained on PANs from Bosnian and Lebanese individuals aged 14-24.99 years (n = 1764), considering pooled and sex-specific datasets with and without augmentation. Binary classification, multiclass classification, and regression models were trained and evaluated. Model performance on an internal test set derived from the same sample was summarized using accuracy, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve (ROC AUC). For regression and multiclass models, legal majority classification was additionally assessed by thresholding predicted ages or age categories at 18 years. External validation employed an independent Brazilian dataset (n = 1579; 14-24.99 years). Formal statistical comparison between internal and external performance employed two-proportion z-tests for accuracy and DeLong's test for ROC AUC.

On the internal test set, the best-performing binary classification model, EfficientNetV2 with augmentation on the pooled dataset, achieved an accuracy of 0.90, sensitivity of 0.92, specificity of 0.86, F1 score of 0.91, and ROC AUC of 0.93. Using the same pooled, augmented configuration, thresholded regression predictions achieved accuracies of 0.85 for EfficientNetV2 and 0.83 for ViT, whereas thresholded multiclass predictions achieved accuracies of 0.84 and 0.73, respectively. Compared with direct binary classification, these thresholded outputs showed no clear advantage, and multiclass models generally showed higher specificity but lower sensitivity. This same pattern was retained when thresholded models were evaluated on the external validation set, with thresholded regression remaining comparatively stable and thresholded multiclass performance declining more markedly, especially for ViT. Visualization maps (gradient-weighted class activation mapping [Grad-CAM] and occlusion sensitivity) confirmed attention to relevant dental structures. On external validation, the same binary EfficientNetV2 configuration achieved an accuracy of 0.81, sensitivity of 0.90, specificity of 0.61, F1 score of 0.87, and ROC AUC of 0.85. Sex-specific models performed similarly, showing no clear advantage over pooled training. Statistical testing confirmed significant AUC degradation on external validation for both models (EfficientNetV2: p = 0.004; ViT: p < 0.001), while the accuracy drop was significant only for ViT (p = 0.035).

This study demonstrated that transfer learning with EfficientNetV2 and Vision Transformer can distinguish minors from adults using PANs with high internal performance and acceptable external generalization. Direct binary classification provided the most robust approach for legal age assessment in the present dataset. The findings support the forensic potential of these models, while also indicating the need for further work to improve robustness and real-world applicability.