Research On CBCT Tooth Segmentation Method Based On Deep Learning

With the rapidly development of computer vision and graphics,digital stomatology has become a reality.The key to digital dental care is to obtain and segment a complete3 D dental model.Cone Beam Computer Tomography(CBCT)is a new medical imaging technology developed in recent years,and it is one of the commonly used equipment for dental diagnosis.During dental imaging,the CBCT scanner rotates around the patient’s head and takes hundreds of images,providing comprehensive 3D volumetric information of all oral tissues,including teeth.Due to its high spatial resolution,CBCT images are suitable for 3D model reconstruction.Accurate tooth segmentation on CBCT images is the basic step for 3D tooth model reconstruction.However,the lack of common boundaries between teeth,the varying profile topologies,and their proximity to alveolar bone and jaw bone make it difficult to segment teeth from CBCT images.Current CBCT tooth segmentation methods mostly consider the prior knowledge and statistical shape model of teeth during segmentation.However,even if excellent manual initialization is used,these methods will always face many false appearance or failure.To solve the above problems,this article proposes a CBCT tooth segmentation method based on Mask Scoring R-CNN deep learning framework.The specific work of this thesis is as follows:1.Since there is no public dental CBCT image data set at present,it is necessary to collect the dantal CBCT image data set and mark the teeth in the CBCT image one by one.In the process of marking,the full manual outlining method is adopted to construct 20 groups of dental data sets of oral CBCT images for deep learning tooth segmentation research.2.This thesis improves the deep learning network Mask Scoring R-CNN,and proposed a CBCT tooth segmentation algorithm based on deep learning,which can achieve automatic and accurate tooth segmentation.The core of this method is a two-stage network: in the first stage,multi-frame CBCT images were input for feature extraction,so that the model could learn a high diversity of features;in the second stage,built the Network and the Region Proposal Network(RPN)together to extract the candidate regions,and finally performed the instance segmentation task and scored the segmentation effect.The experimental results on the CBCT tooth data set constructed in this article show that compared with the traditional tooth segmentation algorithm,the algorithm in this thesis can detect the root information more sharply,and the segmentation effect is more accurate.Compared with the existing CBCT segmentation algorithm based on deep learning,the algorithm in this thesis can segment tooth data with metal artifacts more accurately and has stronger robustnessis.In addition,this thesis proves the effectiveness of the multi-frame input algorithm through ablation experiments and parameter sensitivity experiments.