Feature-guided multilayer encoding-decoding network for segmentation for 3D intraoral scan data.

Area of Science:

• Computer Vision
• Medical Imaging
• Orthodontics

Background:

• Accurate malocclusion segmentation is vital for orthodontic diagnosis and treatment planning.
• Existing deep learning methods struggle with robustness and feature confusion in malocclusion segmentation.
• This limits the reliability of current clinical applications.

Purpose of the Study:

• To develop a robust and accurate 3D dental model segmentation method for malocclusion.
• To enhance the reliability of deep learning in clinical orthodontic applications.
• To improve the semantic recognition of dental malformations.

Main Methods:

• A U-shaped 3D segmentation network with hierarchical feature guidance.
• A feature-guided deep encoder with a novel normalization method and a push-pull strategy for point cloud optimization.
• An inverted bottleneck global feature extraction flow and layer-by-layer decoding for high-resolution mesh reconstruction.

Main Results:

• Achieved 96.6% overall accuracy (OA) and 90.8% mean intersection over union (mIoU) on a custom malformed dental dataset.
• Outperformed existing methods like PointNet and MeshSegNet significantly.
• Demonstrated strong performance on public datasets (Teeth3DS, 3D-IOSSeg) with OA up to 96.4% and mIoU up to 94.5%.

Conclusions:

• The proposed method offers superior performance in malocclusion segmentation compared to existing approaches.
• It provides a robust and accurate solution for intelligent virtual orthodontics.
• The method enhances semantic recognition and spatial feature reconstruction for high-resolution dental meshes.