A hierarchical residual Swin transformer with dual-level locality compensation for accelerated MRI reconstruction

Magnetic resonance imaging (MRI) provides high spatial resolution, excellent soft-tissue contrast, and reliable anatomical discrimination, and is therefore widely used in clinical diagnosis. However, long acquisition time still limits imaging efficiency and clinical throughput. Although convolutional neural network (CNN)-based deep learning approaches have shown considerable promise for reconstructing images from undersampled k-space data, existing methods frequently fail to preserve fine structural details and texture continuity under aggressive undersampling conditions. In this study, we propose a novel reconstruction network based on the Swin Transformer, which integrates three key components: (i) long residual connections to facilitate the recovery of high-frequency details; (ii) a dual-level locality compensation mechanism to restore local spatial information; and (iii) a channel recalibration module to explicitly capture global dependencies among channels, thereby enhancing global feature extraction. Extensive experiments conducted on the IXI and MICCAI 2013 datasets demonstrate that the proposed method consistently outperforms several compared generative adversarial network (GAN)-based approaches for MRI reconstruction across all undersampling mask types, achieving a PSNR improvement of approximately 0.6-1 dB over the best competing method under various settings. The implementation code is available at https://github.com/cvgjhuor/HR-SwinGAN.