Automated Segmentation of Dental Structures in Panoramic Radiographs Using U-Net 3+

This study, based on deep learning, offers an approach that includes segmentation of dental structures from panoramic radiographs using the U-Net 3+ architecture. The proposed model is trained and validated using a dataset of 2,693 images sourced from Kaggle, which includes panoramic dental radiographs of children aged between 2 and 13 years. The proposed model leverages a combination of convolutional, BatchNormalization, and activation layers, with a design tailored to enhance the extraction of fine-grained details through multi-scale feature integration. Metrics such as accuracy, precision, recall, dice score and Mean Intersection over Union (Mean IoU) are used to evaluate the performance of the model using 5-fold cross-validation. The experimental findings highlight the model’s high efficacy, achieving a Dice Score of 0.9331, an accuracy of 0.9736, and a Mean IoU of 0.9182. The {U}-Net 3+ model exhibited strong generalization capabilities, evidenced by the low discrepancy between training and validation metrics, and a robust ability to distinguish between classes, as indicated by an AUC of 0.9611. These results underline the model’s potential for clinical applications, offering precise and reliable segmentation of dental structures, which is crucial for diagnostic and treatment planning purposes. The study’s findings suggest that integrating deep learning methodologies into dental imaging can significantly improve the accuracy and efficiency of dental diagnostics. Future work may involve further refinement of the model and expansion of the dataset to encompass a broader range of dental conditions and demographics.