Accurate recognition of dental anomalies and pathologies is critical for personalized treatment planning across all domains of clinical dentistry. Cone-beam computed tomography (CBCT) is increasingly employed when detailed three-dimensional visualization is needed—for example, in complex endodontic cases, pre-implant assessment, or surgical planning. High-resolution CBCT aids multiple steps of care, including treatment design, risk assessment, and follow-up evaluation. Early detection of conditions such as caries, periapical periodontitis, or root resorption helps dentists intervene before irreversible damage occurs. Accurate identification of pre-existing dental restorations (e.g., composite fillings, ceramic crowns) is critical to prevent iatrogenic damage during subsequent orthodontic tooth movement or prosthetic rehabilitation procedures. CBCT can also reveal hard-tissue variants such as bone islands or retained root fragments, allowing practitioners to anticipate orthodontic tooth movement challenges. Despite these advantages, manual interpretation of CBCT volumes is time-consuming and subject to inter-observer variability (Charuakkra et al., 2023). Consequently, recent research has explored deep-learning approaches—particularly automated tooth segmentation—to streamline CBCT review and improve diagnostic efficiency and consistency.

Recent advancements in deep learning have established it as a powerful tool for automating CBCT segmentation tasks, including the segmentation of the pulp cavity in mandibular molars(Slim et al., 2024), localization of root resorption(Zheng et al., 2024), dentition segmentation(Kim et al., 2024), temporomandibular joint segmentation(Hu et al., 2024) and others. For instance, Altun et al. employed a deep learning model to segment both maxillary sinuses and associated maxillary sinus diseases(Altun et al., 2024). In their study, the model achieved recall, precision, and F1 scores of 1.00, 0.985, and 0.992 for total maxillary sinus segmentation; 1.00, 0.931, and 0.964 for healthy maxillary sinus segmentation; 0.858, 0.923, and 0.889 for mucosal thickening segmentation; 0.977, 0.877, and 0.924 for mucous retention cysts segmentation; and 1.00, 0.942, and 0.970 for sinusitis segmentation. Additionally, Cui et al. introduced a two-stage neural network that leverages hierarchical tooth morphological information for tooth instance segmentation using CBCT, achieving a mean Dice coefficient of 94.8 %(Cui et al., 2021b). Despite these advancements, current studies often rely on overly ideal datasets that may not encompass the spectrum of lesions encountered in real-world clinical scenarios. Therefore, validation using datasets with more diverse and representative clinical lesions is crucial for enhancing the robustness and generalizability of these deep learning models.

Building on the previously identified challenges, our study makes several significant contributions. First, we developed an extensive and diverse dataset encompassing a broad spectrum of clinical lesions, thereby ensuring robust validation of our model (Fig. 1). Second, we developed an approach to accelerating annotation of CBCT scans for model development: this approach leverages an initial deep-learning model to accelerate the labeling process. Third, we introduce an instance-level tooth-segmentation network that produces color-coded, per-tooth renderings. These overlays do not replace the multi-planar reconstructions that dentists rely on for definitive lesion assessment; rather, they give a quick, three-dimensional map that can draw attention to regions where caries, periapical change, or existing restorations might warrant closer multi-planar reconstructions review. Finally, we conducted a simulation comparison in which experienced dental clinicians interpreted a testing set with and without the system assistance; the segmentations reduced anomaly-detection time without compromising diagnostic accuracy, demonstrating clear value for routine dental radiology workflows.