%0 Journal Article
%T 基于双边结构多尺度动态图的牙齿点云分割<br>Tooth Point Cloud Segmentation Based on Multi-Dynamic Graph of Bilateral Structurest
%A 王桢
%A 陈胜
%J Modeling and Simulation
%P 2467-2477
%@ 2324-870X
%D 2024
%I Hans Publishing
%R 10.12677/mos.2024.133225
%X 牙齿隐形矫治装置设计的关键就是将牙齿区域从口腔内三维点云模型中实现精准分割，并在尽量降低人为干预的需求下，对牙齿区域中的单颗牙齿实现全自动分割。传统分割技术需要依赖于专家的先验知识以及大量的人机交互辅助操作，分割性能易受到牙齿形状和位置变化的影响，无法实现全自动分割。因此本文提出了一种Multi-Dynamic Graph of Bilateral Structurest (M-DGB)模型。模型使用双边结构将点云特征信息分别输入坐标边与法向量边，首先利用特征转换模块获得不同尺度的初始全局特征，之后利用多尺度动态图模块中的K近邻图与改进的动态图卷积EdgeConv  模仿卷积神经网络渐近地增加感受野的方式，对局部几何特征实现分层次多尺度学习，进一步提取增强后的局部特征。最后将先前得到的局部增强特征与全局特征密集融合，以获取更具有表达能力的多属性特征。此外，改进了一种混合损失函数，加强牙齿与牙龈的边界分割。将该模型在自制数据集上进行实验，与现有点云分割模型PointNet、PointNet  、MeshSegNet相比，分割精度提高，平均Dice系数为0.972、PPV为0.964、SEN为0.987。<br />
The key to the design of the invisible orthodontic device is to accurately segment the tooth area from the 3D point cloud model in the oral cavity, and to achieve automatic segmentation of a single tooth in the tooth area with minimal human intervention. Traditional segmentation technology relies on the prior knowledge of experts and a large number of human-computer interaction auxiliary operations, and the segmentation performance is easily affected by changes in tooth shape and position, and fully automatic segmentation cannot be realized. Therefore, a Multi-Dynamic Graph of Bilateral Structurest (M-DGB) model is proposed. The model uses a bilateral structure to input the point cloud feature information into the coordinate edge and normal vector edge respectively, firstly uses the feature transformation module to obtain the initial global features of different scales, and then uses the K-nearest neighbor graph in the multi-scale dynamic graph module and the improved dynamic graph convolutional EdgeConv   to imitate the way of convolutional neural network asymptotic increase of the receptive field, realizes hierarchical and multi-scale learning of local geometric features, and further extracts the enhanced local features. Finally, the previously obtained local enhancement features are intensively fused with the global features to obtain multi-attribute features with more expressive ability. In addition, a mixed loss function has been improved to enhance the boundary segmentation of teeth and gums. Compared with the existing point cloud segmentation models PointNet, PointNet   and MeshSegNet, the segmentation accuracy is improved, and the average Dice coefficient is 0.972, PPV is 0.964, and SEN is 0.987.
%K 点云分割，三维牙齿模型，双边结构，动态图卷积，混合损失函数<br>Point Cloud Segmentation
%K 3D Tooth Model
%K Bilateral Structurest
%K Dynamic Graph Convolution
%K Mixed Loss Function
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=87125