Artificial Intelligence-Based Multi-Stage System for Automated Angle's Classification of Malocclusion from Intraoral Images in Orthodontics.

This study presents a multi-stage deep learning pipeline for automated Angle's classification of occlusion using intraoral images in orthodontics. The pipeline integrates three key stages: (1) a binary Occlusion Side Classifier (OSC) to determine whether the input image represents the right or left side of the patient's dentition, (2) side-specific bounding box detection models (MolarBBoxNet-R or MolarBBoxNet-L) to localize the molar region, and (3) a unified classifier (AngleClassifier-R50) to predict Class I, II, or III occlusal relationships. A dataset of 8909 lateral intraoral occlusion images from three orthodontic centers was used, including patients older than 6 years, ensuring the presence of fully erupted first molars. Images were annotated by two experts, with discrepancies resolved through consensus. The pipeline achieved perfect occlusion side classification (accuracy 1.00) and high molar classification accuracy (97.41% internal, 94.3% external), with sensitivity and specificity on the external validation set (383 previously unseen images) of 99.0%/98.9% for Class I, 88.2%/98.7% for Class II, and 97.2%/93.4% for Class III. Grad-CAM visualizations confirmed the model's focus on clinically relevant molar regions, enhancing interpretability. The system processed images in ~ 0.11 s, significantly faster than expert annotation (p < 0.001). By operating solely on intraoral photographs, without radiographs, the approach demonstrates practical potential for general practice and tele-orthodontics. It may reduce diagnostic variability, support large-scale screening, and alleviate clinical triage burden. Future work should expand the framework to incorporate additional occlusal parameters and multimodal inputs such as radiographs or 3D scans for comprehensive orthodontic diagnostics.