Automated Tooth Detection and Caries Identification in CBCT With Deep Learning.
Authors
Affiliations (4)
Affiliations (4)
- Department of Stomatology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China; School of Stomatology, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China.
- Department of Stomatology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China; School of Stomatology, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China. Electronic address: [email protected].
- Department of Stomatology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China; School of Stomatology, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China. Electronic address: [email protected].
Abstract
Deep learning networks have achieved significant progress in caries diagnosis, but automated localization and numbering of carious teeth in cone-beam computed tomography (CBCT) images remain underexplored. This study aimed to develop a two-stage deep learning framework for tooth detection, numbering, and caries identification in CBCT images, providing a technical basis for automated analysis in support of opportunistic caries screening. This retrospective study included CBCT images from 65 eligible patients. Axial slices were used for model development. For tooth detection, seven classification schemes were designed, and YOLOv3 was compared with Cascade R-CNN. For caries identification, three classification networks (DenseNet169, MobileNet_V2, and ResNet50) were trained and evaluated. Detection performance was evaluated using mean average precision (mAP) and average precision (AP). Identification performance was assessed using balanced accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), Matthews correlation coefficient (MCC), and area under the precision-recall curve (AUC-PR). YOLOv3 achieved superior detection performance compared with Cascade R-CNN across all classification schemes (P < .0001). DenseNet169 outperformed the other networks for caries identification. Despite class imbalance, it achieved a balanced accuracy of 0.7414 and an MCC of 0.6074, with high specificity (0.9828) and NPV (0.8976). The integrated two-stage framework showed acceptable overall performance. The proposed two-stage framework showed promising performance in detecting, numbering, and identifying caries in CBCT images, supporting the feasibility of adjunctive opportunistic screening on scans acquired for non-caries indications. This framework may help prioritize clinician review of CBCT scans for suspected carious lesions, and clinical utility requires external multi-center and prospective validation.