TAPSeg: An Open-Source Deep Learning Tool for Instance-Level Tooth and Pulp Segmentation in CBCT.

OBJECTIVES: To develop an open-source one-tap CBCT automatic segmentation tool based on deep learning to perform integrated segmentation and reconstruction of teeth and pulp in the 3D Slicer software and verify its generalization ability on multiple datasets. METHODS: Three datasets were used for model training and evaluation. Tooth segmentation employed a three-stage V-Net collaborative framework (arch area positioning, single-tooth centroid detection, and single-tooth fine instance segmentation), while pulp segmentation used nnU-Net for 3D semantic segmentation. Model performance was assessed on independent external datasets, including immature permanent teeth with open apices, using Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), sensitivity, and precision. Fine-tuning and additional evaluations were performed for teeth with open apices. RESULTS: TAPSeg achieved robust performance across multiple test sets for tooth segmentation (n = 198) and pulp segmentation (n = 148), with DSCs ranging from 91.5%-94.2% and 91.0%-92.2%, and HD95 values of 0.700-1.449 mm and 0.704-1.008 mm for tooth and pulp segmentation, respectively, while maintaining high sensitivity and precision. For immature permanent teeth, hard-tissue segmentation achieved a DSC of 92.1% ± 6.5%. CONCLUSIONS: By integrating V-Net and nnU-Net, TAPSeg provides accurate, efficient, and generalizable tooth-pulp segmentation with one-click operation as a 3D Slicer plug-in, lowering the threshold for clinical use. CLINICAL RELEVANCE: TAPSeg enables clinicians to generate patient-specific 3D models from CBCT scans within minutes, supporting standardized quantification of root morphology and volume changes, follow-up of root development, regenerative endodontic procedures, and automated digital dental workflows.