Magnetic resonance image reconstruction based on graph convolutional Unet network

Magnetic Resonance Imaging (MRI) acquisition is a long time process that leads to patient discomfort and motion artifacts. Convolution network has been implemented in MRI image reconstruction to speed up MRI. Convolution network utilizes spacial local convolution kernel successfully, but non-local information with self-similarity is ignored. In order to effectively explore non-local information in MRI reconstruction, we propose graph convolutional Unet (GCN-Unet) for MRI image reconstruction. In the GCN-Unet, non-local information is represented by a patch graph extracted from MRI images, in which nodes are consist of image patches, and edges are self-similarities between nodes, reflecting non-local structure similarity intra an MRI image. We utilize graph convolution with a Unet framework to effectively represent features of non-local information with self-similarities and aggregate the extracted non-local features to reconstruct MRI image. Experiments demonstrate the effectiveness of the GCN-Unet, the Peak Signal-to-Noise Ratio (PSNR) and Structure Similarty Index Measure (SSIM) of reconstructions are improved.