Improvement of Image Reconstruction for MRI Using Phase-Scrambling Fourier Transform and Dual-Domain Strategy

Compressed sensing MRI (CS-MRI) has been proposed as one of the methods to accelerate MRI. In recent years, CSMRI methods using deep learning (DL) model have become the mainstream, and the quality of reconstructed images has improved significantly. The DL model-based CS-MRI method proposed by Ito et al. reconstructs regularly undersampled phase-scrambling Fourier transform imaging (PSFT) signals using ADMM-CSNet. However, this method had problems with smoothing and residual artifacts in the reconstructed images, and was insufficient for complex-valued images. To solve these problems, we propose an MR image reconstruction method using PSFT signal and dual-domain strategy. The proposed method adds the k-space convolution layers to ADMM-CSNet and further increases the number of convolution layers in both the k-space and space domain. To verify the effectiveness of the proposed method, the simulation experiments were conducted using the complex-valued images. As a result, the PSNR of the reconstructed images with the proposed method was 1 dB higher than that with ADMM-CSNet for PSFT and Fourier transform imaging signals. In addition, the reconstructed images showed reduced smoothing and residual artifacts, and also showed improved performance in reconstructing detailed structures.