Deep learning based fully automated cephalometric skeletal classification

We aimed to develop a deep convolutional neural network model capable of directly classifying sagittal and vertical skeletal patterns from lateral cephalometric radiographs and to compare its performance with that of a commercially available fully automated landmark-based software (WebCeph). The dataset was created from 1500 cephalometric radiographs archived between 2019 and 2025. Balanced class labeling was performed using QuickCeph measurements under physician supervision, with a total of 6 classes in 3 sagittal classes and 3 classes vertical, 500 radiographs in each class. Sagittal classification was based on ANB (SI: 0–4°, SII: >4°, SIII: <0°); vertical classification was based on SN-GoGn (Normal: 26–38°; Horizontal: <26°, Vertical: >38°), moreover FMA and Y-axis were used as supporting criteria. The training of the deep learning model was conducted in two stages: in the first stage, preprocessing was applied, and in the second stage, feature extraction and classification were performed using a convolutional neural network. For comparison, the same dataset was also analyzed with WebCeph, a fully automated, landmark-based software. WebCeph's performance was evaluated against the expert-supervised QuickCeph results. The deep learning model achieved 87% accuracy in sagittal classification and 81% accuracy in vertical classification. However, on the same data, WebCeph's accuracy remained at 76.32% in sagittal classification and 68.2% in vertical classification. The deep learning model developed in our study showed higher performance compared to WebCeph. While overall high agreement was observed between WebCeph and QuickCeph in angular measurements, relatively low agreement was observed for FMA. In conclusion, a deep learning approach directly classifying lateral cephalometric radiographs can achieve higher accuracy, particularly in the sagittal direction, compared to a fully automated landmark-based method. This finding highlights the clinical benefit of end-to-end learning, which eliminates cumulative errors arising from landmark identification phase. However, cephalometric landmark analysis remains crucial for detailed treatment planning and long-term follow-up. To enhance the generalizability of the proposed approach, larger, multicenter, and independently validated studies from different centers/devices are recommended. Keywords: Artificial Intelligence, Deep Learning, Fully Automatic Cephalometric Analysis, Skeletal Classification