Automatic Detection And Instance Segmentation Of Tooth Based On CBCT Images

With the research on deep learning and smart medicine in recent years,digital dental consultation has been rapidly developed.Obtaining a complete 3D dental model is a key step in digital dentistry,and CBCT(Cone beam CT)dental data,because it contains the complete morphological structure information of teeth,is gradually becoming one of the common data in dental clinical diagnosis and treatment and clinical experiments,and the automatic and accurate CBCT dental example segmentation is important for the subsequent treatment tasks,especially orthodontic work.The automatic and accurate segmentation of CBCT dental examples is important for the subsequent treatment tasks,especially orthodontics.Many scholars have conducted many meaningful studies on CBCT tooth segmentation in recent years,and great progress has been made.Traditional tooth segmentation methods often ignore the rich morphological features of teeth,especially when facing complex clinical malocclusion cases,most tooth segmentation methods still fail to achieve better segmentation results.Although current deep learning methods can greatly improve the accuracy of tooth instance segmentation,the application of deep learning methods in this area is greatly hampered by the very small amount of CBCT data currently available for network training.Therefore,accurately capturing the complete geometry as well as morphological details of a single tooth remains a challenge for current tooth segmentation research.To address these issues,this paper focuses on the accurate automatic detection and instance segmentation of teeth from CBCT data,as follows.First,to address the problem that the complex geometric structures of teeth with nodal features such as roots of CBCT data are difficult to capture,which leads to poor accuracy of tooth segmentation,a method of CBCT tooth instance segmentation based on capture dependency and perceptual field adjustment is proposed,and the new method constructs a new two-stage deep learning framework,TSDNet(Tooth Segmentation Deeplearning Network)to achieve automatic and accurate instance segmentation of dental CBCT data.The first stage is used to obtain the accurate tooth center of mass to ensure accurate spatial localization of the tooth,and to obtain the instance information of a single tooth.In the second stage of the network,the coding process introduces a guidance module for tooth geometry information based on the 3D self-attention mechanism to enhance the network’s ability to capture the overall structure information of the tooth.The proposed multi-scale cavity convolution-based tooth feature integration module is also used to capture tooth surface and root details at multiple scales,which improves the tooth segmentation accuracy.We evaluate the proposed method using CBCT dental datasets obtained from real medical institutions,and experimentally demonstrate that TSDNet has high segmentation performance and accurate instance segmentation,which has certain advantages over existing advanced methods.Second,when only the center of mass of the tooth is used as the key point for tooth localization,the network training voxels point to the center of mass,and the overall tooth features are often difficult to capture when facing clinical malocclusion cases with more complex geometric structures,while the small amount of CBCT training data affects the model training effect.To address these problems,this paper proposes a tooth instance segmentation method(Tooth Morphology Segmentation Deeplearning Network(TMSDNet))with morphological data enhancement and morphology key feature points guidance in CBCT.The framework introduces a hierarchical tooth mesiodistal prediction network with hierarchical guidance of key points generated from the center of CBCT tooth data layer by layer,and the network training voxels point to the tooth mesiodistal axis layer by layer.We designed a multi-task learning mechanism to output instance-level tooth mesial axes as a secondary task to finally obtain accurate tooth segmentation results.At the same time,we introduced the 3D Thin Plate Spline(TPS)to distort the CBCT tooth data to simulate the generation of more complex clinical malocclusion cases and expand the training volume of the network data.Extensive evaluation,ablation studies and comparisons with existing methods show that our method achieves state-of-the-art segmentation results,especially in challenging complex clinical malocclusion cases,and we also show good segmentation performance,demonstrating the potential applicability of our method in clinical medicine.