Background and Aims: Panoramic radiographs are routinely used for the assessment of dental conditions; however, their manual interpretation is time-consuming and highly dependent on the clinician's expertise. This study aimed to develop a deep learning-based U-Net framework for the automatic segmentation of four common dental conditions: dental caries, dental fillings, root canal-treated teeth, and impacted third molars.
Materials and Methods: This study was designed as the development of an applied model and was conducted in 2021 at the University of Tehran. Panoramic dental images were collected from publicly available sources. Two dentists collaboratively annotated the images and assigned labels corresponding to four dental conditions: dental caries, dental fillings, root canal treatment, and impacted wisdom teeth. To prevent data leakage, 10% of the entire dataset were initially reserved as an independent test set. Data augmentation were then applied exclusively to the training and validation data. Following augmentation, the remaining dataset were split into training and validation sets using an 80:20 ratio, respectively. A modified U-Net model was trained using preprocessed images resized to 512 × 512 pixels. Model performance was evaluated using Recall, precision, F1-score, and accuracy.
Results: The proposed model demonstrated promising performance in detecting the four target conditions. F1-scores, which were considered the primary performance metric because of potential class imbalance, were 73.56% for caries, 92.2% for fillings, 87.1% for impacted wisdom teeth, and 76.0% for root canal treatment. Precision values were 79.74%, 97.5%, 77.0%, and 61.2%, respectively, while accuracy values ranged from 98.4% to 99.5%.
Conclusion: The proposed U-Net-based framework demonstrated promising performance for the automated identification of dental conditions in panoramic radiographs and might facilitate computer-aided dental screening and treatment planning. Further validation using larger and more heterogeneous datasets are required to enhance the model’s generalizability.
