Automated assessment and detection of third molar and inferior alveolar nerve relations using UNet and transfer learning models.

Panoramic radiographs (PRs) are widely used in assessing the relationship between the mandibular third molar (MM3) and the inferior alveolar nerve (IAN). The relationship of MM3 and IAN is a critical consideration in oral and maxillofacial surgery due to the risk of nerve injury that can lead to anesthesia in the lower lip or chin. This study aimed to evaluate the performance of different transfer learning models in classifying these relations into four classes after detecting and segmenting the MM3 and the IAN using a new method with the UNet model. 714 PRs from more than 500 patients, representing four classes, were utilized in a UNet model to annotate the MM3 and IAN. Then, the annotations were extracted and augmented to yield 1021 images, thereby achieving a more balanced distribution of samples in each class. Different transfer learning models were evaluated in classifying the images using 10-fold cross-validation. The UNet model achieves an accuracy of 0.97 in annotating the images. The transfer learning models performed well. The best model, DenseNet121, achieved an average accuracy of 0.84, precision of 0.86, recall of 0.84, and F1-score of 0.85 across all folds and classes. The proposed approach demonstrates high performance compared to other works, classifying four classes and achieving similar or better performance than other techniques. It can also benefit dentists at the annotation and classification stages. Based on our results, this approach holds promise for opening several research directions in the future.Trial registration: This study was approved by the internal review boards (IRB) of the Jordan University Hospital (JUH) (Ref. No. 10/2025/3202).