Abstract: Exploiting the sparse nature of data to reconstruct image from randomly undersampled K-space is the main method to solve the problem of limited application of magnetic resonance imaging(MRI) due to long acquisition time.However,the loss of textures is serious when the existing methods are used to reconstruct the highly undersampled MRI.Aiming at this problem,referring to the adversarial learning idea of generative adversarial networks(GAN),a novel method for highly undersampled MRI reconstruction based on semantic-contrast generative adversarial network(SC-GAN) is proposed.This method consists of two successive parts.In the first part,the MRI image with Cartesian highly random undersampled is input into the U-NET-based ge-nerator to compete with the discriminator for generating the initial reconstructed image,so as to construct the reconstruction subnet.The other part is the semantic contrast subnet,which compares the semantic information between the initial reconstructed image and its fully-sampled image with VGG-16 network.Comparison results are fed back to the first part for parameter adjustment until the best reconstructed image is generated .Experimental results show that the excellent reconstructed results are verified by subjective and objective evaluation when the acceleration factor is up to 7(14%).Compared with the state-of-art methods,the proposed SC-GAN has lower memory consumption,faster convergence speed and more textures,and can provide algorithm support for the development of next-generation MRI machines.

Key words: Deep learning, Generative adversarial network, Highly undersampled images, MRI reconstruction, Semantic contrast, Sparsity