Classification of mandibular impacted third molar teeth on panoramic radiographs using deep learning algorithms

Résumé fourni par la source

This study aimed to evaluate the image-based technical performance of deep learning-based models in classifying impacted mandibular third molars on panoramic radiographs. Accurate identification of angular impaction patterns is essential for treatment planning, yet manual interpretation is time-consuming and subject to observer variability. Deep learning approaches may offer a more efficient and standardized alternative. This retrospective single-center study included 961 impacted mandibular third molars identified on 569 panoramic radiographs obtained from the archive of Dicle University Faculty of Dentistry. Each tooth was categorized into mesioangular, horizontal, distoangular, or vertical positions according to the Winter classification. Ground-truth labels were established through consensus assessment. Images were annotated using Roboflow, and the dataset was divided into training, validation, and test subsets. YOLOv8n and Faster R-CNN (ResNet-50 FPN) models were trained with pre-trained weights in the Google Colab environment. Model performance was evaluated using precision, recall, AP@0.5, AP@0.5–0.95, and mean Average Precision (mAP). The YOLOv8n model achieved an mAP@0.5 of 0.955, mAP@0.5–0.95 of 0.785, precision of 0.88, and recall of 0.93, whereas the Faster R-CNN model achieved an mAP@0.5 of 0.9026, mAP@0.5–0.95 of 0.6995, precision of 0.87, and recall of 0.93. Deep learning–based classification of impacted mandibular third molars on panoramic radiographs appears feasible under the conditions of the present dataset. YOLOv8n yielded numerically higher performance metrics than Faster R-CNN under the conditions of the present dataset; however, no formal inferential comparison was performed. Further external validation and clinician-assistance studies are required before potential clinical implementation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.