Automated diagnosis and classification of skeletal malocclusion using artificial intelligence on 2D facial photographs: A multi-modal deep learning approach

Objective: This study aims to establish and validate a two-stage deep learning framework that enables reliable and cost-effective identification of sagittal and vertical skeletal malocclusions based on 2D facial photographs. Methods: 2854 photographs were retrospectively collected from 1427 patients, along with corresponding lateral cephalograms and metadata (sex and age). A two-stage deep learning model (FaceDSM-Net) was developed using MobileNetV3-Large as the backbone, integrating facial photographs with metadata. The model was trained and tested internally, with performance evaluated by accuracy, precision, AUC, recall, and F1-score. Grad-CAM was used for interpretability. The generalizability was assessed by an independent external test set, and the performance was further compared with that of 72 raters with varying clinical experience on the same dataset. Results: FaceDSM-Net demonstrated excellent performance in sagittal (AUC 0.96, accuracy 0.85) and vertical (AUC 0.95, accuracy 0.86) classifications. Its robustness was improved by incorporating sex and age. Heatmaps showed FaceDSM-Net focused on key facial structures associated with skeletal malocclusion. The model demonstrated comparable reliability compared to human raters. Conclusion: FaceDSM-Net effectively classified sagittal and vertical skeletal malocclusions using multimodal data, demonstrating good interpretability and moderate generalizability, particularly in the frontal-vertical task. It holds potential for clinical application in skeletal malocclusion prediction. Clinical significance: This model significantly reduces resource and cost burdens while maintaining diagnostic performance. It supports early screening, individual and family-level diagnosis, and longitudinal follow-up. Its comparable performance over expert raters highlights its potential as a reliable clinical decision support tool, particularly in non-radiographic or resource-limited settings.