Deep Learning-Based Classification of Facial Profile Convexity from Profile Photographs in Orthodontics: An...

The soft-tissue facial profile is a cornerstone of orthodontic diagnosis and treatment planning, strongly influencing facial esthetics and patient satisfaction. This study aimed to develop and evaluate a deep learning-based framework for automated classification of facial convexity (convex, normal, concave) from standardized profile photographs, with an emphasis on transparent preprocessing and model interpretability. A dataset of 1200 natural head position (NHP) profile photographs (400 per class) was labeled by three experienced orthodontists using a consensus approach based on the soft-tissue facial convexity angle, operationally defined by the Glabella-Subnasale-Pogonion (G-Sn-Pg) landmarks. Images underwent cropping, background removal using U2-Net, silhouette generation, and contour extraction to emphasize geometric profile features while minimizing photometric and demographic confounding factors. A custom convolutional neural network (Contour-CNN) was trained using L2 regularization, dropout, cosine-annealing learning-rate scheduling, and Bayesian hyperparameter optimization. Model performance was evaluated using accuracy, precision, recall, F1-score, confusion matrix analysis, receiver operating characteristic (ROC) curves, and saliency-based interpretability measures. The proposed model achieved an overall accuracy of 98% on a held-out internal test set. Interpretability analyses suggested that model predictions were influenced by anatomically plausible facial regions, supporting the potential clinical plausibility of the decision-making process. Nevertheless, external validation, prospective clinical studies, and appropriate ethical oversight are required before routine clinical deployment.