CS-SwinGAN: A swin-transformer-based generative adversarial network with compressed sensing pre-enhancement for multi-coil MRI reconstruction

Abstract

Magnetic resonance imaging (MRI) reconstruction from undersampled k-space data is a crucial area of research due to its potential to reduce scan times. Current deep learning approaches for MRI reconstruction often combine frequency-domain and image-domain losses, optimizing their sum. However, this approach can lead to blurry results, as it averages two fundamentally different types of losses. To address this issue, we propose CS-SwinGAN for multi-coil MRI reconstruction, a swin-transformer-based generative adversarial network with a Compressed Sensing Block for pre-enhancement. The newly introduced Compressed Sensing Block not only facilitates the separation of frequency-domain and image-domain losses but also serves as a pre-enhancement stage that promotes sparsity and suppresses aliasing, thereby enhancing reconstruction quality. We evaluate CS-SwinGAN in both standard MRI reconstruction tasks and under varying noise levels in k-space to assess its performance across diverse conditions. Numerical experiments demonstrate that our framework outperforms state-of-the-art methods in both conventional reconstruction and noise suppression scenarios. The source code is available at https://github.com/notmayday/CS-SwinGAN_MC_Rec.