Abstract
Objective: Intraoral scanners (IOS) provide high-precision 3D data of teeth and gingiva, critical for personalized orthodontic diagnosis and treatment planning. However, traditional segmentation methods exhibit reduced performance with complex dental structures, such as crowded, missing, or irregular teeth, constraining their clinical applicability. This study aims to develop an advanced 3D point cloud segmentation model to enhance the automated processing of IOS data in intricate orthodontic scenarios. Methods: A 3D point cloud segmentation model was developed, incorporating relative coordinate encoding, Transformer-based self-attention, and attention pooling mechanisms. This design optimizes the extraction of local geometric features and long-range dependencies while maintaining a balance between segmentation accuracy and computational efficiency. Training and evaluation were conducted using internal and external orthodontic datasets. Results: The model achieved a mean Intersection over Union (IoU) of 92.14% on the internal dataset and 91.73% on the external dataset, with Dice coefficients consistently surpassing those of established models, including PointNet++, TSGCN, and PointTransformer, demonstrating superior segmentation accuracy and robust generalization. Conclusions: The model significantly enhances tooth segmentation accuracy in complex orthodontic scenarios, such as crowded or irregular dentitions, enabling orthodontists to formulate treatment plans and simulate outcomes with greater precision-for example, optimizing clear aligner design or improving tooth arrangement efficiency. Its computational efficiency supports clinical applicability without excessive resource consumption. However, due to the limited sample size and potential influences from advancements in IOS technology, the model's generalizability requires further clinical testing and optimization in real-world orthodontic settings.