结合稀疏八叉树卷积神经网络的牙齿预备体颈缘线提取方法  

Extraction of Tooth Preparation Line Based on Sparse Octree Convolutional Neural Network

作　　者：张贝[1] 戴宁[1] 田素坤 袁福来 俞青[2] Zhang Bei;Dai Ning;Tian Sukun;Yuan Fulai;Yu Qing(College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016;Department of Prosthodontics,Nanjing Stomatological Hospital,Nanjing 210008)

机构地区：[1]南京航空航天大学机电学院,南京210016 [2]南京市口腔医院口腔修复科,南京210008

出　　处：《计算机辅助设计与图形学学报》2019年第12期2129-2135,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(51775273);国家重点研发计划项目(2018YFB1106903);江苏省重点研发计划项目(BE2018010-2);江苏省自然科学基金(BK20161487);江苏省六大人才项目(GDZB-034)

摘　　要：针对传统手工提取牙齿预备体颈缘线需要交互标记特征点,操作复杂,效率低的问题,提出了一种基于稀疏八叉树的卷积神经网络自动提取牙齿预备体颈缘线的方法.首先利用稀疏八叉树的空间划分,牙齿预备体模型被预处理为带有标签信息的稀疏点云,构建牙齿预备体数据集;其次利用已训练的卷积神经网络模型将牙齿预备体点云分割为2部分;然后采用密集条件随机场优化分割点云的边界,再将边界点拟合及插值获取新的边界点集;最后连接边界点在预备体模型上对应的投影点形成牙齿预备体颈缘线.在牙齿预备体数据集上的实验结果表明,卷积神经网络模型的预测准确率达到97.23%,通过对该方法提取的预备体颈缘线与专业医生提取的颈缘线之间的曲线偏差进行对比分析,验证了该方法的有效性.The traditional tooth preparation margin line extraction requires marking feature points interactively,which is complicated and inefficient.Aiming at the problem,a method is proposed to extract the preparation line by convolutional neural network.Firstly,in order to construct a dental preparation dataset,preparation models are processed as the sparse octree point cloud with labels through the spatial division of the octree.Secondly,the trained network model is used to divide the preparation point cloud into two parts.Thirdly,the dense conditional random field is used to optimize boundary points.Then,the new boundary points can be achieved by the fitting and interpolation.Finally,the corresponding projected points of the boundary points are connected to form a preparation line.Experiments on dental preparation datasets show that the accuracy of label predicted by network model can reach 97.23%,and the validity of the method is verified by comparing the curve deviation between the automatic constructed preparation line and the manual extracted preparation line.

关 键 词：预备体颈缘线 稀疏八叉树 卷积神经网络 条件随机场 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]