Abstract
BACKGROUND: To develop a deep learning-based model that is capable of automatically segmenting teeth in cone-beam computed tomography (CBCT) images and generating auxiliary diagnostic reports. METHODS: A two-stage pipeline model comprising a segmentation module and a classification module was designed. The segmentation module integrates the 3D TransUNet model and fine-tuned nnU-Netv2 framework to predict tooth location and numbering. The classification module uses a 3D DenseNet169 model to assess tooth conditions and detect dental diseases. A total of 450 CBCT datasets were collected, preprocessed, and annotated with tooth contours, numbering, and disease status. The dataset was randomly split into training, validation, and test sets at a 3:1:1 ratio. The model's performance was then evaluated by using multiple quantitative metrics. RESULTS: The segmentation module achieved a Dice similarity coefficient (DSC) of 0.9409 and an average symmetric surface distance (ASSD) of 0.5011. The classification module, which is based on the 3D DenseNet169 model, achieved an accuracy of 0.9297 and an F1-score of 0.9252. CONCLUSIONS: The two-stage pipeline model effectively integrates automated tooth segmentation and auxiliary diagnosis. The segmentation module demonstrates high accuracy, while the classification module exhibits strong diagnostic performance. The final output seamlessly combines segmentation and the diagnostic results, enabling the automated generation of structured auxiliary diagnostic reports with tooth numbering. By significantly enhancing the diagnostic efficiency, the proposed method offers substantial support for clinical decision-making in dentistry and holds great potential for real-world applications.