Privacy-preserving deep learning for automated ANB estimation from lateral cephalometric radiographs with an exploratory extension to Cobb angle assessment

Abstract Automated cephalometric analysis requires accurate landmark localization and reliable transformation of radiographic data into clinically interpretable measurements. This study developed and validated a privacy-preserving deep learning pipeline for automated detection of Sella, Nasion, A-point, and B-point landmarks and ANB angle estimation on radiographs. Specifically, lateral cephalometric radiographs were used for ANB landmark detection and estimation. In addition, anatomical reference points used for Cobb angle measurement in scoliosis assessment were defined to ensure consistency in auxiliary spinal evaluations. Initially, 1286 raw DICOM images underwent technical quality control, protected health information audit, anonymization, DICOM-to-PNG fidelity validation, and strict landmark quality control, yielding 1282 final images. Each image was evaluated by three orthodontists to establish the reference landmark standard. Ground-truth ANB angles were geometrically computed from the annotated landmarks after y-up coordinate conversion. A UNet-based heatmap regression model was trained for four-landmark localization and evaluated using stable 5-fold cross-validation and an independent locked test set. In cross-validation, the model achieved a mean radial error of 1.03 ± 0.45 mm, ANB mean absolute error of 0.97 ± 0.61°, and class accuracy of 86.85 ± 4.90%. On the locked test set, mean radial error was 0.76 mm, ANB mean absolute error was 0.57°, class accuracy was 92.68%, and [email protected] mm was 96.34%. This audited pipeline provides a geometrically verified and de-identification-compliant framework for automated ANB estimation from radiographs. As an exploratory proof-of-concept extension, the shared architecture achieved a Cobb angle mean absolute error of 0.78° on a small locked test subset ( n = 36), suggesting potential for auxiliary spinal evaluation that warrants dedicated validation.