Research On Tooth Point Cloud Completion Algorithm Based On PoinTr Model

With the improvement of substance levels,the incidence of periodontal disease in the population also gradually increases.The three-dimensional tooth model obtained by digital oral scanning equipment is a key factor in disease diagnosis and treatment.In orthodontics and implantology,it is necessary to diagnose the patient’s dental condition based on the imaging results to formulate corresponding treatment strategies.Currently,cone beam computed tomography(Cone Beam Computerized Tomography,CBCT)is one of the commonly used imaging methods in the diagnosis and treatment of dental diseases.However,due to the low resolution of the imaging device,the scanning process is easily affected by obstructions.The tooth surface structure is complex,and some parts are connected with the soft tissue.The three-dimensional tooth model obtained by scanning is often incomplete,resulting in partial deletion,which is difficult to meet the requirements of auxiliary diagnosis.Most of the existing point cloud completion studies are based on common objects,such as cars,tables and chairs,and have not been applied to dental data.In view of the above problems,this paper studies the automatic hole completion task of low-precision tooth point cloud based on the deep learning method.The main research contents and achievements are as follows:(1)To establish a tooth point cloud data set,this paper collected 130 cases of intraoral scan data extraction gold standard with high imaging quality.The CBCT does not match the oral scan data spatially,so it is necessary to make the missing data of the oral scan.Extract the sparse point cloud from oral scan to obtain the low-precision point cloud of the teeth to be completed.Due to the small sample size of dental data,8 kinds of objects were extracted from the public data set Shape Net,and 130 cases of data were selected for each object to make a small sample data set.Intuitively compared the completion effect of the network.(2)The CEC PoinTr(Combine Edge Convolution Diverse Point Cloud Completion with Geometry-Aware Transformers)point cloud completion network is designed.In view of the problem of the small amount of tooth point cloud data,polymorphic missing data augmentation processing was performed on the dentition.On the basis of the PoinTr network,the feature information fused with multi-level semantics was extracted by joint edge convolution.In this paper,a completion experiment is carried out on the small sample Shape Net dataset and the tooth point cloud dataset,and the chamfering distance between the output point cloud and the ground truth is calculated.The distance values of the improved CEC PoinTr network on the two datasets are reduced by 0.1 and 23.03,verifying its effectiveness.At the same time,the completion results of the CEC PoinTr network applied to the tooth point cloud are significantly better than other networks,and the tooth shape and missing parts can be restored.(3)The OP PoinTr(Offset Attention-Patch Variance Diverse Point Cloud Completion with Geometry-Aware Transformers)point cloud completion network is designed.Aiming at the problem that it is difficult to learn the details of the tooth surface,each dentition is processed in segments.A bias attention module with relative position features is introduced into the completion network,and the loss function is designed in combination with the chamfer distance and the uniformity of the point cloud itself.The experiments are carried out on the small sample Shape Net dataset and the tooth point cloud dataset respectively.The chamfering distance calculated by the improved OP PoinTr network is reduced by 1.23 and 52.27,respectively,reaching 12.76 on the smallsample Shape Net dataset.The evaluation index of the OP PoinTr network is better than other completion networks.