基于GAN和TV正则化的MRI超分辨率重建算法

为解决磁共振成像(MRI)超分辨率重建图像失真、组织细节模糊及速度较慢等问题,提出基于生成对抗网络(GAN)框架和TV正则化的MRI超分辨率重建算法。根据MRI重建图像特点,改进GAN框架中的生成器损失函数,在GAN对抗损失的基础上,为保证低分辨率MRI图像到高分辨率图像的映射一致性,引入生成图像与原始图像残差的L2范数损失,为提高重建图像的细节信息,引入生成图像的全变分正则化损失,基于改进的损失函数,采用随机梯度下降法进行训练。实验结果表明,该算法在保证重建质量的同时能较好地恢复组织细节、加快重建速度,超分辨率重建效果明显。

Total variation regularized MRI super-resolution algorithm with generative adversarial networks

To solve the problem of image distortion, tissue blur and slow reconstruction in MRI super-resolution, an algorithm of total variation regularized MRI super-resolution with generative adversarial nets’ frame was presented. Based on the characteristics of MRI reconstructed image, the loss function of generator in GAN was improved. On the basis of GAN adversarial loss, the L2 norm loss was introduced to ensure the mapping consistency between low- and high-resolution MRI image, and the total variation regularized loss was introduced to improve the detail information. Based on these, stochastic gradient descent was used to train the networks. Experimental results show that the proposed algorithm can restore the details of the organization and speed up the reconstruction speed while guaranteeing the quality, and the effect of reconstruction is obvious.