CurSegNet: 3D Dental Model Segmentation Network Based on Curve Feature Aggregation

Jiafa Mao, Zhan Liu, Jingke Gu, Chunping Wang, Sixian Chan

Abstract: The precise segmentation of 3D dental models obtained from intraoral scanners (IOS) is a primary task in computer-aided orthodontic diagnosis and treatment. Existing dental model segmentation networks mainly focus on extracting local aggregation features and help to describe the local feature information of the model to some extent. However, they ignore the relationship between distant local aggregation areas, resulting in some segmentation regions with multiple or missing teeth. To address this issue, we propose a 3D Dental Model Segmentation Network Based on Curve Feature Aggregation (CurSegNet), which combines curve aggregation features to learn richer mesh feature semantic information. Specifically, we introduce a dual-stream branch that independently handles local features and curve features of 3D dental model. This design allows the network to acquire more comprehensive underlying spatial information at various scales. Furthermore, we propose a cross-channel cascaded module for extracting geometric features from the dual-stream branches and subsequently merging them. This approach leverages multi-scale underlying spatial information to improve the quality of 3D mesh features. Finally, experiments evaluate a clinical patient dental model dataset, demonstrating that our method is superior to the current state-of-the-art segmentation methods.