Tooth Segmentation Method Based On Deep Learning

With the rapid development of 3D digital imaging technology in recent years,digital oral technology has attracted more and more attention,and its convenience and efficiency also affect the oral medical industry.Segmenting a single tooth from a three-dimensional dental model is one of the important steps of the virtual orthodontic system.At present,most orthodontic software in the industry adopts a tooth segmentation method that requires interactive marking.A seed point is selected on each tooth of a three-dimensional dental model through human-computer interaction,which is inefficient.To solve this problem,the use of deep learning to solve the automatic selection of seed points in the process of tooth segmentation is proposed.It is proposed to use an improved full convolutional network to identify the tooth bounding box,build a self-built dental jaw plane tooth bounding box image data set,intercept the jaw plane image,and mark the bounding box coordinates of each tooth.There are a total of 1000 teeth in the data set.Jaw plane image;in order to improve the accuracy of network recognition,the SE module is embedded in the network;the final experiment shows that the two-dimensional deep learning method can accurately find the bounding box of the tooth,and select the center of the bounding box as the seed point to segment the dental model,But due to the difference between the center point of the bounding box of the bounding box and the seed point of the incisor,the wrong segmentation of the incisor occurs.In order to solve the problem of incorrect segmentation of two-dimensional images,it is proposed to use FeaStNet graph convolution to identify tooth seed feature points.First,by analyzing the position,of the seed point and the final segmentation effect of each tooth type,setting up unified rules,self-built a dental model seed point feature data set,by simplifying segmentation effect of each tooth type to establish uniform rules model's seed point dataset;then,a new multi scale network structure was constructed using feature-directed graph convolution to identify the 3D The feature information on the tooth and jaw model is deepened to better fit the teeth;finally,the network model is deepened to include the information that has been collected through the network The found feature points are used as base points,and the nearest point to the base point on the tooth and jaw model is found as a seed point,if the seed point If the location is accurate,the tooth is divided from the gum based on the seed point.For results with inaccurate seed point locations,the seed point locations were corrected by manual manipulation and then segmented.The proposed automatic seed point selection method can automatically select tooth seed points and solve the problem of cross-marking in tooth segmentation.It basically realizes the automation of tooth segmentation,which is suitable for the segmentation of all kinds of malformed tooth patient models.