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

Abstract

Accurate segmentation of malocclusion is crucial in orthodontic diagnosis and treatment planning, but existing deep learning methods seriously affect the reliability of clinical applications due to poor robustness and feature confusion between neighboring tooth classes when dealing with malocclusion. To address this problem, a U-shaped 3D dental model segmentation method based on hierarchical feature guidance is proposed. First, a feature-guided deep encoder architecture is constructed, which introduces a normalization method that combines the local mean with the global standard deviation. And a push–pull strategy is employed to optimize point cloud density, adjusting the standard deviation variation to meet the point cloud density requirements of different regions. Second, an inverted bottleneck global feature extraction flow was designed to guide the encoder in learning the overall features of the dentition and jaw through dynamic scaling of deep-level features,thereby enhancing the semantic recognition of malformations. Finally, an interpolation method is used to decode the high-level dental semantic information layer by layer to reconstruct the spatial structural features of the high-resolution dental mesh. Experimental results on a self-constructed malformed dental dataset show that the proposed method achieves an overall accuracy (OA) of 96.6% and a mean intersection over union (mIoU) of 90.8%, respectively, which are 3.4% points and 8.2% points higher than that of PointNet, 11.3% points and 26% points higher than that of MeshSegNet, and 2.2% points and 5.6% points higher than that of PointeNet ,and the number of model parameters is only 1.54 M. Meanwhile, on the public datasets Teeth3DS and 3D-IOSSeg, the OA of the proposed method reaches 96.4% and 90.1%, and the mIoU reaches 94.5% and 86.8%, respectively. These performance advantages indicate that the proposed method can better meet the development of intelligent virtual orthodontics.