Fully automated Pell & Gregory classification on panoramic radiographs

This study proposes a fully automated deep learning system based on the U-Net architecture for classifying mandibular third molars using the Pell & Gregory method. Novel anatomical landmarks were introduced and automatically detected on panoramic radiographs by the model. These landmarks were then used to determine the classification through their spatial relationships. The system was trained and evaluated using panoramic radiographs collected from different patients. Two independent datasets were constructed according to the side of mandibular third molar impaction: 373 images for the left jaw (teeth 37–38) and 328 for the right jaw (teeth 47–48). For the Pell & Gregory classification, the proposed approach achieved a classification accuracy of 93.24% for the left jaw and 91.30% for the right jaw, demonstrating consistent and reliable performance across both datasets. The model effectively localized anatomical points and classified third molars without manual input. This automated approach enhances diagnostic consistency and reduces observer variability, offering practical utility in clinical environments. Overall, the study demonstrates the potential of artificial intelligence to improve diagnostic workflows by providing a reliable tool for the automated classification of impacted third molars according to the Pell & Gregory system.