Answers are grounded in this paper and cite their sources.
Background Lateral cephalometric analysis is necessary for proper orthodontic diagnosis and treatment planning. Classically it depends upon manual landmark identification on radiographs, which is time consuming, relies on expert skills. Accordingly, this study aimed to create and evaluate a deep learning model for automated cephalometric assessment using lateral profile photographs as a radiation free alternative. Materials & methods A total of 1,350 dataset was used to train a dual task model using HRNetV2 combined with an Adaptive Multilayer Perceptron (MLP) to predict 19 landmark coordinates. The model was tested using 150 unseen photographs to evaluate the Mean Radial Error (MRE), Successful Detection Rate (SDR), Mean Absolute Error of the angular and linear measurements and classification accuracy. Results The HRNetV2 + MLP model combination achieved a mean radial error of 2.317 ± 1.80 mm, with SDRs at the 2 mm and 4 mm thresholds were 53.8% and 84.5%, respectively. The ANB angle demonstrated Mean absolute error of 1.46º. Bland-Altman analysis revealed minimal systematic bias between AI and Dolphin measurements across most cephalometric variables. Conclusion These findings demonstrate the feasibility of automated landmark detection and analysis from lateral facial photographs and support further investigation of the photographic-based approaches thus eliminating radiographic exposures.
No takes yet. Share an insight, caveat, or question.
Osman et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: