Seeing Through the Face: Robust Facial Diagnosis Enhanced by Landmark Detection and Pose Normalization

The treatment of oral‐maxillofacial deformities not only addresses aesthetic concerns but also helps mitigate potential risks to the digestive, circulatory, and nervous systems, as well as adverse effects on adolescent growth and development. Timely screening and prevention are crucial; however, reliance on experienced professionals has hindered effective implementation. Here, a novel approach with excellent robustness and generalization capability is proposed. First, 33 landmarks on 3,135 lateral photos are located, and the Combined Heatmap Regression and Coordinate Regression network is trained, achieving 99% landmark location accuracy. Using these landmarks and X‐ray data as the gold standard, a multilayer perceptron is trained, achieving a 94.7% diagnostic accuracy in diagnosing skeletal facial deformities. Finally, using the Intelligent Networked System for Human Identification and Graphical Health Tracking (INSIGHT), we screened 440 students for facial deformities without the involvement of medical professionals, achieving an accuracy rate of over 90%. The model uses landmark detection for facial soft tissue measurement, making the AI‐generated preliminary diagnosis more interpretable and reducing the impact of shooting devices and postures. Application of INSIGHT in primary hospitals, clinics, telemedicine, and community surveys could reduce reliance on medical staff, enhance screening efficiency, and improve healthcare in developing regions.