Feasibility of predicting vertical cephalometric angles from panoramic radiographs using deep learning

Abstract

Introduction Early identification of vertical skeletal discrepancies is essential for orthodontic diagnosis and treatment planning. Since panoramic radiographs (OPGs) are more routinely obtained than lateral cephalometric radiographs (LCR), this study evaluated whether artificial intelligence could predict vertical skeletal angles on OPGs. Methods LCRs and OPGs of 715 patients were retrospectively collected from four imaging centre (2022-2025). LCRs were traced using WebCeph to obtain reference measurements of the Frankfort-mandibular plane angle (FMA), gonial angle, and Sum of Bj & ouml;rk. Multiple convolutional neural network (CNN) architectures (EfficientNet-B3, DenseNet121/169, ResNet-50/101, VGG16/19) were trained to predict these parameters from corresponding OPGs. Ensemble averaging was also employed as a non-learned aggregation strategy. Model performance was evaluated using mean absolute error (MAE) and the coefficient of determination (R-2). Wilcoxon signed-rank test assessed the differences in predictive performance. Inter-model agreement was quantified using intraclass correlation coefficients (ICC). Gradient-weighted Class Activation Mapping (Grad-CAM) was used for model interpretability. Results Ensemble averaging achieved the highest predictive accuracy across all angular parameters, with MAE values of 2.53 degrees +/- 0.08 for FMA, 3.16 degrees +/- 0.10 for the gonial angle, and 3.04 degrees +/- 0.09 for the Sum of Bj & ouml;rk. High agreement was observed among the CNN architectures for all measurements (P < 0.001). Grad-CAM visualizations indicated that predictions primarily relied on the gonial angle region, followed by the condylar area and mandibular ramus. Conclusion Deep learning demonstrates promising potential for estimating vertical angular measurements on OPGs. Although current prediction errors preclude replacement of cephalometric analysis, the incorporation of larger datasets, geometry-aware models, and external validation will help improving predictive accuracy.