TRFSA-HQS: Transformer-based MRI Compressed Sensing Network with Frequency Domain Self-Attention

Compressive sensing (CS) can reconstruct undersampled magnetic resonance images, but its iterative optimization process tends to be computationally intensive and time-consuming. Convolutional neural networks (CNNs) have demonstrated impressive reconstruction performance through non-linear feature extraction and mapping capabilities. However, CNNs often struggle to effectively learn and capture global dependencies in the data. Therefore, constructing a MRI reconstruction method that meets clinical real-time imaging and captures dynamic global correlation is crucial in fast and high-precision MRI tasks. We propose a deep unfolding network (TRFSA-HQS) for fast and accurate CS reconstruction, which combines frequency domain self-attention (FSA) based Transformer and half-quadratic splitting (HQS) iterative optimization scheme. TRFSA-HQS adopts an iterative scheme based on HQS to effectively decouple and update optimization problems. The decoupled data subproblems are updated by minimizing the objective function with competitive terms, while the regularization subproblems are updated using the TRFSA deep prior network. The TRFSA module employs an asymmetric UNet architecture, where the encoder and decoder utilize a frequency-domain discriminative feedforward network (DFFN) and FSA. The DFFN selectively extracts deep features from different frequency components, while the FSA captures global dependencies in the frequency domain. The experiment demonstrate that the proposed model achieves an average reconstruction PSNR of 35.11dB on a 0.2 sampling rate test set on FastMRI knee joint data, with an inference time of 0.74s. It has good reconstruction performance and noise robustness, meeting the clinical real-time imaging requirements.