ERSA-Net:Encoder Networks based on Residuals and Self-Attention for Accelerating MRI Reconstruction

Accelerating magnetic resonance imaging (MRI) reconstruction with deep learning-based methods is a challenging task. This paper proposes ERSA-Net, an encoder-based residual self-attention network for MRI reconstruction. This network is designed to reconstruct high-quality magnetic resonance images from undersampled K-space data. In particular, the backbone network architecture we propose is based on an encoder. It employs a series of residual blocks to improve the utilization of feature information in each encoder layer. As each layer progresses, we always use a residual network to retain feature information in each layer. In addition, the proposed ERSA-Net utilizes a specialized feature extractor to extract important feature values. It then performs self-attention calculations on the extracted information to identify dependencies between feature values and enhance the performance of the network model. We validated the effectiveness of ERSA - N et on datasets with various undersampling ratios and demonstrated that it significantly outperforms state-of-the-art methods.