Classification of morphological variations of mandibular condyle in panoramic radiographs with a deep learning approach

Abstract

in panoramic radiographs. Materials and Methods A total of 1,056 panoramic radiographs, containing 2,112 healthy mandibular condyles, were used in the study. The dataset was split into training (~80%), validation (~10%), and test (~10%) sets. Two experienced dento maxillofacial radiologists annotated the training images and classified the condyles into four morphological categories: Round, Angled, Diamond, and Crooked Finger-shaped. The YOLOv8 deep learning model was trained using transfer learn ing, hyperparameter tuning, and fine-tuning techniques. Performance was assessed using metrics including precision, recall (sensitivity), F1-score, mean Average Precision (mAP), and training time. True positives, false positives, and false negatives were evaluated based on bounding box localization and class assignments. Results The model demonstrated balanced performance across classes in the training dataset. On the test dataset, the model achieved an overall F1-score of 0.769 and mAP@0.5 of 0.786. The highest performance was observed for the Crooked Finger class (0.795 precision, 0.870 recall, 0.831 F1-score, 0.837 mAP@0.5) and the Angled class (0.723 precision, 0.860 recall, 0.786 F1-score, 0.808 mAP@0.5). The Round class showed moderate results with 0.677 precision, 0.870 recall, 0.761 F1-score, and 0.798 mAP@0.5. The Diamond class had the lowest performance, with 0.528 precision, 0.696 recall, 0.600 F1-score, and 0.661 mAP@0.5. Conclusion The model effectively distinguishes the Angled and Crooked Finger classes but faces challenges with the Dia mond and Round classes. Despite varied performance, the model demonstrates balanced performance overall, providing a foundation for further refinement and optimization.