Tooth Shape Generation And Deformation Based On Deep Convolutional Neural Networks

The prosthodontic CAD function is an important part of the prosthodontic CAD/CAM system,this function determines the contour of the diseased tooth by matching the standard tooth model in the system,and then completes the design of the tooth shape by manual adjustment.however,this approach is time consuming for the physician and the final model designed is easily influenced by the physician’s experience.In order to solve the above problems,this paper proposes a convolutional neural network-based tooth shape automatic design method,which mainly solves the shape design and model deformation problems in dental model design work.The main work of this paper is as follows.To address the problem that the current dental shape design based on the CAD system for restorative dentistry still requires manual adjustment by the doctor,we design a tooth shape generation network which improves the lack of geometric information during tooth shape generation by introducing semantic guidance information.Also,this paper proposes a self-assessment mechanism using scoring of segmentation confidence for finding and correcting unreliable pixel prediction results.By testing the performance of this network on the Shining 3D dental dataset we constructed,the experimental results show that the method proposed in this paper outperforms other available methods in terms of accuracy.Among which the PSNR is increased by about 1.31,it achieves the goal of helping doctors to complete tooth shape design according to a flexible and efficient way.To address the problem that the current 3D deformation network only aggregates local neighboring voxel features and does not fully explore the interrelationship between non-local voxel features in the work of tooth shape model deformation,this paper proposes a method based on environmental information mining.The method can obtain environmental information from different spatial domains to improve the representation performance of the network,the main work includes the introduction of a self-attentive mechanism to improve the discrimination of voxel features by learning the non-local dependencies of different voxels in the feature space.As well as introducing a multi-scale learning method that uses an atrous convolution method with different expansion rates to obtain environmental information in different perceptual domains separately,providing richer contextual features for the model.In addition,this paper proposes a method that can adaptively combine the features obtained by the encoder and decoder,further improving the nonlinear mapping capability of the model.Finally,this paper constructs a dental dataset to solve the problem of lacking database when studying tooth deformation.Experimental results show that the accuracy of the proposed method is more excellent,and the MIo U is improved by 3.6% compared with other methods.