Graph-Channel Attention U-Net for Undersampled MRI Reconstruction

Due to the lengthy acquisition time required for magnetic resonance imaging (MRI), researchers have proposed various methods to shorten this time. However, these methods often overlook the non-local similarity within the images, resulting in the inability to reconstruct images with clear structures. To address this issue, we present a Graph Channel Attention U-Net (GCA-Unet).The input to GCA-Unet consists of three images with different resolutions. Each encoder is specifically trained on input images of varying scales, as well as feature information from the downsampled preceding decoder. The architecture employs a Multi-Head Graph Attention Module (MGAM) to capture non-local features within the images. In the final layer of the decoder, we output the reconstructed image. Given that the input comprises images of different scales, we also introduce a Multi-Scale Feature Fusion Module (MFFM) to effectively integrate features across multiple scales. Extensive experiments conducted on brain and knee datasets demonstrate that our proposed GCA-Unet significantly outperforms other MRI reconstruction methods across various sampling rates. These results validate the effectiveness of GCA-Unet in reconstructing images with clear structures.