PrepNet: A Multi-Task Geometric Deep Learning Framework for Automatic Detection and Numbering of Prepared Tooth in 3D Dental Scan Models.

OBJECTIVES: This study aimed to develop and evaluate PrepNet, a multi-task geometric deep learning framework, for automatic detection and numbering of the prepared tooth in 3D dental scan models (DSMs). METHODS: In this study, we propose PrepNet, a multi-task geometric deep learning framework designed to simultaneously predict the target ROI centroid, estimate the ROI size, and classify the tooth number from 3D DSMs. PrepNet integrates local graph feature extraction using DGCNN to capture surface variations with global contextual modeling via a Point Cloud Transformer to encode long-range dental structures. RESULTS: PrepNet achieved the lowest mean radial error for ROI centroid detection (1.498± 0.884 mm) and the highest successful detection rate at 2.0 mm (80.29%) compared with the baseline models. For tooth numbering, PrepNet showed the highest classification accuracy of 0.836. For ROI size estimation, PrepNet achieved the lowest maximum deviation of 2.420 mm, with a mean absolute error (0.620 ± 0.480 mm) comparable to the best-performing baseline. CONCLUSIONS: PrepNet demonstrated the feasibility of multi-task geometric deep learning for the automatic detection and numbering of the single prepared tooth in 3D DSMs. CLINICAL SIGNIFICANCE: This approach may improve the efficiency of single-crown CAD/CAM design workflows by reducing manual target-tooth selection and ROI definition prior to dental crown design.