Deep Learning Study Of Third Molar–Nerve Contact On Orthopantomogram

This study uses the contact relationship between the mandibular third molar (MM3) and the inferior alveolar nerve (IAN) on cone-beam computed tomography (CBCT) as the gold standard. It aims to develop a deep learning model to automatically assess the contact between MM3 and IAN on panoramic radiographs, minimizing false-positive contacts and reducing the need for CBCT scans while maintaining diagnostic accuracy. A total of 1,279 panoramic radiographs containing mandibular third molars (MM3), collected from patients between April 2024 and January 2025, were selected as the dataset for this study. Bounding boxes were manually annotated using the open-source software LabelImg. The dataset was divided into training, validation, and test sets in a ratio of 8:1:1. A deep learning model based on the ResNet-50 architecture was pretrained and fine-tuned using the validation set to construct an automatic assessment model. Model performance was evaluated on the test set using precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC). The developed ResNet-50 deep learning model demonstrated excellent performance in assessing the contact relationship between MM3 and IAN. When tested on 1,279 panoramic radiographs containing MM3, the model achieved an accuracy of 0.860, a recall of 79.1%, an F1-score of 0.800, and an AUC of 0.890. The ResNet-50-based deep learning model we developed demonstrates reliable performance and enables accurate assessment of the contact between MM3 and IAN on panoramic radiographs.