Research On 3d Point Cloud Classification Of Tooth Based On Spatial Relationship Features

Digital orthodontics is an advanced computer-aided treatment technology that has emerged in the field of orthodontics in recent years.Compared with traditional orthodontics,which relies heavily on the dentist’s experience in orthodontic treatment,digital orthodontics is data-based and relies on various digital and intelligent technologies,which greatly reduces human resource input and alleviates the current human resource shortage problem that widely exists in the field of dentistry.The task of tooth classification is a key step in digital orthodontics.The digital orthodontic system has to perform further diagnosis on the basis of the classification results before proposing an orthodontic plan as well as implementing orthodontic treatment.Therefore,the accuracy of dental classification has a huge impact on digital orthodontics.However,the current dental classification methods used in the digital orthodontic process are less efficient and accurate,and they are generally only coarse-grained by function,which cannot be closely coordinated with the subsequent stages of digital orthodontics.In order to improve the intelligence of digital orthodontics,as well as the efficiency and accuracy of dental classification,this thesis investigates the intelligent and fully automated dental classification method.At present,the dental models widely used in clinical practice are 3D point clouds of teeth acquired by intraoral scanning equipment.According to clinical observation and data analysis,there are two major difficulties in classifying 3D point clouds of teeth:firstly,it is very difficult to classify these similar teeth at a fine-grained level because the geometries of tooth models with close locations or similar functions are highly similar to each other.Secondly,the classification task is oriented to teeth requiring orthodontic treatment,so there are a large number of anomalies in the data,such as wisdom teeth,missing teeth,malocclusions,etc.,which greatly affect the classification accuracy.These two major problems lead to the ineffectiveness of the general 3D dental point cloud model classification method.In order to solve the difficult problems in 3D point cloud classification of teeth,this thesis proposes a 3D point cloud classification network based on spatial relationship features,and its main innovations are as follows:(1)To address the problem of too much similarity between 3D point clouds of teeth,this thesis proposes to use the feature that different teeth have a fixed spatial distribution and encode it as a spatial relationship feature,while using a multi-scale point cloud feature extraction structure to extract point cloud features with stronger expressive power,and by fusing two features with complementary information to obtain a classification feature with multiple feature information at the same time,which will be used for the tooth point cloud classification task(2)This thesis also proposes the use of multi-classification Focal Loss as the loss function of the network for solving the problem of low accuracy of wisdom tooth classification due to the unbalanced quantity of tooth samples from different categories.Focal Loss greatly alleviates this problem by a special strategy.(3)For the problem of multiple anomalies in tooth classification,multiple anomaly handling methods are used to deal with the major anomalous conditions in the classification results.The main problems such as widely existing anomalies,missing teeth,and sticky-teeth are discussed to further improve the accuracy and robustness of the algorithm based on the point cloud classification network proposed in this thesis.