A sample application designed for detection of teeth and jaw bone from cone-beam computed tomography images using deep learning methods.

objectivesCone Beam Computed Tomography (CBCT) provides detailed anatomical information for treatment planning in dentistry. However, manually identifying tooth and jawbone structures is time-consuming and can vary depending on the observer. The aim of this study is to analyze the performance of U-Net, DeepLab V3+, and YOLO V3 deep learning architectures for automatic segmentation of tooth and jawbone structures in CBCT images.

methodsThis study utilized CBCT data from seven different patients. A total of 1,155 axial images were expertly labeled in terms of tooth and jawbone regions. The dataset was trained with U-Net, DeepLab V3+, and YOLO V3-based semantic segmentation models. The learning rate, number of epochs, and batch size parameters of the models were optimized using the GridSearch method. Dice Similarity Coefficient (DSC), Intersection over Union (IoU), precision, and recall performance evaluation metrics were used in the performance assessment.

resultsSuccessful results were obtained in tooth and jawbone segmentation in all deep learning models used in the study. Among the models used in the article, the most successful performance was obtained from the U-Net architecture. The U-Net model achieved DSC=0.9289, IoU=0.8671, precision=0.9213, and recall=0.9365 values with a learning rate of 0.001, 70 epochs, and 16 batch size parameters. The DeepLab V3+ deep learning algorithm also yielded similar results, while YOLO V3 showed lower performance compared to other models.

conclusionDeep learning-based segmentation methods provide high accuracy in the automatic determination of tooth and jawbone structures in CBCT images. Among the deep learning models used in the study, U-Net was identified as the most successful model. The approach developed in this study has the potential to support treatment planning in dental applications and reduce the burden of manual segmentation. However, the method needs to be validated with larger, multi-center datasets before it can be put into clinical use.