| Three-dimensional tooth mesh segmentation refers to the classification of teeth and gums in a 3D dental scan mesh model and assigning clinically relevant labels,such as canine and central incisor.Three-dimensional tooth mesh segmentation is a critical step in digital orthodontic simulation and intervention,providing dentists with a reference to clinical diagnosis.Therefore,the demand for obtaining the position of teeth in the oral cavity with high precision and full automation always remains.However,due to the lack of clinical labeled data,the large differences in individuals,and complex mesh feature extraction difficulties,the clinical applicability of 3D tooth mesh automatic segmentation still remains limited.This thesis focuses on how to design a set of dual-stream feature extraction and effective 3D dental mesh automatic segmentation algorithm with good performance on few-shot learning.The main work and innovation of this thesis are as follows:(1)A supervised 3D tooth mesh segmentation method based on dual-stream feature extraction using structural and spatial features is designed to address the problem of feature confusion in mesh segmentation of single-stream input dental mesh model.Due to the complexity and irregularity of 3D dental mesh data,the original features of a single-stream input mesh would ignore the complementarity of the dental mesh structure and spatial features,resulting in the loss of important semantic information required for tooth segmentation.Therefore,this thesis distinguish between structural features and spatial features of dental mesh data to avoid feature confusion issues.The structural feature stream is designed to extract more precise local tooth details by building a local graph neighborhood search and attention aggregation mechanism from coordinate and normal information.The spatial feature stream is designed to ensure the robustness of the model to complex dental arches such as missing teeth,misaligned teeth,and crowded teeth by using global graph neighborhood search and average pooling.Experimental results from a 15-class clinical dataset show that our method achieved DSC,SEN,and PPV of 97.08%,91.98%,and 91.63%,respectively.(2)Aiming at the problem of excessive reliance on labels in supervised dualstream feature extraction networks,a self-supervised 3D tooth mesh segmentation based on an augmented network for improving dual-stream feature extraction is designed,and for the first time,self-supervision is applied to 3D tooth mesh segmentation tasks.This method uses the supervised network segmentation results as the baseline to construct a self-supervised network of a comparison formula for original data-enhanced data.Using the commonalities and differences obtained through comparison as selfsupervised pseudo label signals,learn transferable knowledge from a large amount of unlabeled data.Experimental results from a 15-class clinical dataset show that our selfsupervised method has an increase of 3.01 percentage points in SEN value compared to the latest self-supervised method.In addition,this thesis designs a post-processing algorithm based on Graph Cut,and deeply optimizes the isolated error prediction and non-smooth boundary generated by the neural network to meet the special needs of clinical applications. |
