Research On 3D Dental Model Segmentation Based On Deep Learning Method

With the development of digital technology,computer-aided diagnosis and treatment system has been widely used in the field of Stomatology.As the key step in this system,3D dental model segmentation can provide important information for clinical diagnosis.Existing 3D dental model segmentation methods usually apply deep learning technology to automatically segment teeth,which have better robustness and generalization than the traditional segmentation methods.However,due to the complicated oral structure of patients in clinical environment,there usually exists misaligned and missing teeth in 3D dental model.This makes existing segmentation methods still face great challenges.Considering the above background,this thesis summarizes the issues of existing methods,and proposes two 3D dental model segmentation model to address these issues.The specified content is as follows:Firstly,existing methods can not fully mine the relationship between meshes in local region,therefore,they fail to extract the local detail features from 3D dental model.To solve this problem,this thesis proposes a graph attentional convolution network.This network consists of two branches,one branch applies the attention-based graph convolution to extract multi-scale local details,and the other branch applies the max-pooling to learn the global feature from 3D dental model.Finally,these two features with different scales are fused to obtain the final representation for teeth segmentation.We evalute the method on a real-patient 3D dental model dataset,the experimental results show that the method achieves 94.84% segmentation accuracy,which is the best among four compared methods,and the proposed method can better distinguish the boundaries of teeth.Secondly,existing methods ignore the differences between spatial attributes in 3D dental model,which causes the network learn confused feature.To solve this problem,this thesis proposes a two-stream graph convolution network.This network first adapts two parallel branches to learn multi-scale features from coodinates and normal vectors respectively,which can effectively avoid the confusion between different attributes.After that,a self-attention mechanism is applied to fuse the output features produced by two branches to obtain more comprehensive feature information for each mesh.Finally,the network segments each tooth based on the fused feature.The proposed method not only can make full use of different attributes,but also effectively solve the feature confusion between attributes.The experimental results show that this method achieves 96.96% segmentation accuracy,which is better than the first method in this thesis,and the proposed method can better address the isolated false predictions.