基于深度学习的牙齿嵌塞自动判别方法

食物嵌塞是口腔常见病征，容易引发局部牙龈红肿、溢脓、龋齿等口腔问题，给患者带来极大的痛苦和不便。目前临床上难以自动筛查嵌塞牙齿，且传统的锥形束CT重建方法的准确度及精度均有待提高。提出一种牙齿嵌塞自动化判断的方法，对牙齿模型进行单个牙体的精准分割，在U-Net网络的基础上使用KPConv卷积核代替二维卷积核来构建分割网络，并使用图割方法优化分割结果。同时，采用平面拟合的方法将分割后的牙齿模型投影到水平和竖直平面上，在平面上求出牙齿嵌塞特征，并利用支持向量机根据所求特征对牙齿的嵌塞情况进行判断。通过充分利用样本模型的几何结构信息，设计简化牙齿模型的几何采样及包含牙齿几何结构约束的图割方法提高网络模型的运算时间及精度。实验结果表明，该方法对牙齿模型的分割准确率为92%，对牙齿嵌塞的判断正确率为81%，能够为医生提供辅助诊断。

Automatic Identification Method of Tooth Impaction Based on Deep Learning

Food impaction is a common oral symptom, which easily causes oral problems, such as local gingival redness, pus overflow, and dental caries, bringing great pain and inconvenience to patients.At present, automatically screening impacted teeth in clinics is difficult, and the accuracy of the traditional cone beam CT reconstruction method must improve.An automatic judgment method of tooth impaction is proposed to accurately segment a single tooth in the tooth model.Based on the U-NET network, a KPConv convolution kernel was used, instead of a two-dimensional convolution kernel, to construct the segmentation network, and the graph cut method was used to optimize the segmentation results.Meanwhile, the segmentation model after the embedded teeth plug features according to the relevant judgment of whether to plug embedded teeth using the plane projection method from the horizontal to the vertical plane, the plane and the embedded teeth characteristics, and the support vector machine asking for the teeth embedded plug profile of judgment.The geometric sampling of the simplified tooth model and graph cutting method containing the constraints of the tooth geometric structure were designed by fully using the geometric structure information of the sample model, improving the operation time and accuracy of the network model.The results show that the segmentation accuracy of the tooth model can reach 92%, and the judgment accuracy of tooth impaction can reach 81%, which can provide doctors with auxiliary diagnoses.