Automatic Generation Method of Lumbar Spine Using Virtual Scoliosis Cases
Statistical Shape Model(SSM)describe the morphological variations in deformable objects and are widely used in Three-Dimensional(3D)liver segmentation,3D spine generation,and other fields.However,the existing SSMs yield inaccurate results and demonstrate unsatisfactory applicability when generating virtual Adolescent Idiopathic Scoliosis(AIS)lumbar spine cases.An automatic generation method of lumbar spine using virtual AIS based on a 3D Variational autoencoder Generation Adversarial Network(3D-VGAN)is proposed in this study.The 3D-VGAN model comprises an encoder,a generator,and a discriminator.The encoder and generator extract data characteristics based on a Variational AutoEncoder(VAE)model and combine the spatial attention mechanism.Meanwhile,residual block is used to solve network degradation during neural network training.To overcome the insufficient stability of the network model and the mismatch in training speed between the generator and discriminator,a threshold training method is used to train the network model to improve its stability.Experimental results on a dataset based on a 3D lumbar spine model of 43 AIS cases show that the Structural Similarity(SSIM)coefficient of this model is 0.999 62,which is 0.080 09％,0.002 00％,and 0.122 20％higher than those of 3D-VAE,3D-Variational AutoEncoder Generation Adversarial Network(VAEGAN),and SSM models,respectively.The Frechet Inception Distance(FID)coefficient of the case generation experiment is 275.653 48,which is 8.420 55％,0.977 73％,and 7.319 27％lower than those of 3D-VAE,3 D-VAEGAN,and SSM models,respectively.Thus,the proposed model outperforms existing models in lumbar spine data generation using virtual AIS case by leveraging attention mechanism,residual block,and threshold training methods.
国家科技期刊平台
NSTL
万方数据
