GFACNet: 3D dental segmentation from intraoral scans integrating geometric features and anatomical constraints

Dental segmentation is a critical step in computer-aided orthodontic treatment planning, but accurate segmentation still faces numerous challenges due to complex tooth morphology, ambiguous gingival boundaries, and clinical issues such as malformed teeth, crowding, and malocclusion. This paper proposes GFACNet, a network that integrates geometric features and anatomical constraints for 3D dental segmentation from intraoral scan data. Our method comprises three key innovations: 1) a morphology-aware graph construction (MAGC) mechanism that adaptively constructs graph structures based on dental geometric characteristics, 2) a multi scale transformer (MST) feature integration module that processes features at different scales while capturing both local and global context, and 3) a hierarchical anatomical constraint loss (HACL) that incorporates multi level anatomical features to guide anatomically consistent segmentation. Experiments on real intraoral scanning datasets demonstrate that GFACNet significantly outperforms existing methods in handling complex dental morphologies, particularly in cases of malformed and missing teeth. Additionally, our method requires reduced computational resources while providing a more practical solution for clinical applications.