虚拟脊柱侧凸病例腰椎自动生成方法

Automatic Generation Method of Lumbar Spine Using Virtual Scoliosis Cases

统计形状模型(SSM)是一种描述可变形物体形态学变化的模型,被广泛应用于三维(3D)肝脏分割、3D脊柱生成等任务。针对现有SSM在生成虚拟青少年特发性脊柱侧凸(AIS)腰椎病例数据时存在生成结果不真实且对AIS病例适用性较差的问题,提出一种基于三维变分自编码生成式对抗注意力网络(3D-VGAN)的虚拟AIS腰椎自动生成方法。3D-VGAN模型由编码器、生成器和判别器组成,编码器和生成器在变分自编码器(VAE)模型的基础上结合空间注意力机制提取数据特征,同时利用残差模块解决了神经网络训练过程中的网络退化问题。为了克服生成-判别结构的3D-VGAN模型中存在的稳定性不强、生成器与判别器训练速度不匹配的问题,使用阈值训练法对3D-VGAN模型进行训练,提升了稳定性。在基于43例AIS病例腰椎3D模型的数据集上的实验结果表明:3DVGAN模型在病例重建实验中的结构相似性(SSIM)系数为0.99962,相比3D-VAE、3D-变分自编码生成式对抗注意力网络(VAEGAN)和SSM模型分别提高了0.08009%、0.00200%、0.12220%;在病例生成实验中的弗雷歇初始距离(FID)系数为275.65348,相比3D-VAE、3D-VAEGAN和SSM模型分别降低了8.42055%、0.97773%、7.31927%。上述实验结果验证了3D-VGAN模型通过注意力机制、残差模块和阈值训练法取得了更好的虚拟AIS病例腰椎数据生成性能。

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 0.99962,which is 0.08009%,0.00200%,and 0.12220%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 275.65348,which is 8.42055%,0.97773%,and 7.31927%lower than those of 3D-VAE,3D-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.