Abstract
This study developed a deep learning model for the automated detection and classification of impacted third molars using the Pell and Gregory Classification, Winter's Classification, and Pederson Difficulty Index. Panoramic radiographs of patients treated at Sivas Cumhuriyet University between 2014 and 2024 were retrospectively analyzed. Impacted teeth were manually classified, and annotations were created using the CVAT software with bounding-box labeling. The dataset included 2,300 images for training, 765 for validation, and 765 for testing, encompassing 7,624 impacted teeth for training, 2,580 for validation, and 2,493 for testing, with 98 unique labels. The YOLOv11 model was trained using optimized hyperparameters (learning rate: 0.01, batch size: 4, up to 1,000 epochs) and data augmentation. Performance metrics included precision (0.980), recall (0.948), F1 score (0.974), mAP@50 (0.990), and mAP@50:95 (0.974). Specific labels, such as 48-Distoangular-C-III (F1: 0.633), exhibited lower F1 scores. The model demonstrated high accuracy and efficiency, addressing the limitations of manual classifications. Enhancing dataset diversity and refining challenging labels could further improve outcomes. This model automates the complex classification of impacted third molars, offering a reliable, efficient decision support system for clinical applications, streamlining workflows, reducing variability, and improving diagnostic precision.