A Pointnet Network Model That Processes Intraoral Scanning Data Automatically

To develop a deep learning end-to-end model that can automatically complete tooth position identification, segmentation and completion of individual teeth, facilitating further targeted clinical diagnosis and academic research such as restoration scheme design and tooth morphology research. A dataset comprising 1200 cases of intraoral scanning datas with a 1:1 male-to-female ratio was partitioned into training, validation, and test sets in an 8:1:1 ratio. A deep learning network based on the PointNet architecture was employed for model training, incorporating multi-level point set sampling strategies, hierarchical feature fusion mechanisms, multi-scale convolutional operations, and normalization processing. During model optimization, the training loss demonstrated exponential decay while the validation loss exhibited an oscillatory decline with a serrated pattern, with both curves gradually converging, indicating favorable training dynamics. The optimized framework achieved clinically relevant performance metrics: average tooth position identification accuracy of 96.0%, segmentation accuracy of 99.35%, and maintained 91.2% diagnostic precision when processing complex cases involving severe malocclusion or partially erupted teeth. The end-to-end network architecture maintained efficient computational performance, completing the entire workflow from raw intraoral scanning data input to individual tooth files exportation within an average processing time of about 1.2 seconds per case. These experimental findings demonstrate that the proposed model effectively accomplishes automated processing of intraoral scan data, with performance metrics holding substantial promise for supporting large-scale clinical diagnostics and academic research applications.