Research On Tooth Point Cloud Registration Methods Based On CBCT And Intraoral Scanning Data

Oral diseases can not only affect the physical and mental health of patients but also cause systemic diseases.With the development of oral digitization and 3D scanning technology,Cone Beam Computed Tomography(CBCT)and Intraoral Scanning(IOS)are widely used in clinical dental practice.CBCT 3D imaging can provide complete information about the interior of the mouth such as crowns and roots,but the scan resolution is low and there is more redundant information;IOS can capture high-resolution images of teeth in real time,but it only provides information about the crown structure.Therefore,combining information obtained from IOS and CBCT helps to achieve a faster,simpler and more detailed dental treatment process.However,the current researches about tooth point cloud data are mainly focused on classification,segmentation and three-dimensional reconstruction,and there is few research on end-to-end neural network for registration in the field of dental health care.Based on the above background,this paper studied the tooth point cloud registration method of CBCT and IOS data.The main research contents are as follows:(1)Construction of the dataset.In view of the lack of paired tooth point cloud public datasets,this paper collected the raw data of CBCT and IOS.First,extracted the periodontal structure of CBCT raw data with a medical image processing software called Mimics,and separated it into the upper and lower jaw which were exported to the STL format.And then the format of both data were converted into point clouds and down-sampled by Visual Studio.Z-score normalization was used to eliminate spatial differences between cross-source point clouds.Finally,CBCT-IOS tooth paired point cloud dataset was constructed.(2)Design of PCR-SA(Point Cloud Registration with Set Abstraction)tooth point cloud registration network.PCRNet only extracts global features of the point cloud,while ignoring the point-to-point connections.Therefore,PCR-SA integrated local and global features by introducing the Set Abstraction(SA)module.The point cloud was divided into several small overlapping regions by farthest point sampling and ball query.Global features of each small regions were extracted by MLPNet.The network was optimized on the basis of pre-training parameters to solve the small samples training problem of deep neural networks.PCR-SA achieved better registration results than the original network through comparative experiments.PCR-SA of 8 recursive structure(r PCR-SA)had higher accuracy,and the Chamfer Distance(CD)and the Earth Mover Distance(EMD)were 0.091,0.124 respectively.(3)Design of PCR-SIFT(Point Cloud Registration with Point SIFT)tooth point cloud registration network.The structure of PCR-SA is simple and the registration speed of it is faster,but the network ability of feature learning is weak.Therefore,Point SIFT module was added to PCR-SA to extract more detailed local features by stacked 8-neighborhood search and orientation-encoding convolution,and the scale invariance of features was realized by two ways of convolution stacking,identity shortcut connection(PCR-SIFT)and residual connection(Res-PCR-SIFT),which further improved the registration accuracy of the network.In the comparison experiments,Res-PCR-SIFT had higher registration accuracy than other networks,and faster than PCR-SIFT network,CD loss and EMD loss decreased to 0.042,0.118 respectively.The effects of different registration directions on tooth point cloud registration were verified on PCR-SA and PCR-SIFT networks,and the experimental data showed that the direction from IOS to CBCT had a better registration effect.