KDI-Net: dual-domain mapping network for undersampled magnetic resonance imaging reconstruction

Magnetic resonance imaging (MRI) is an imaging technique that provides comprehensive anatomical and functional information on the human body, but prolonged acquisition of fully sampled MRI images causes patient discomfort and motion artifacts. In recent years, deep learning (DL) has made significant progress in accelerating MRI image reconstruction. At present, MRI image reconstruction is concentrated on single domain, but hybrid domain reconstruction has more advantages. Here, we propose a hybrid domain reconstruction method based on AUTOMAP—KDI-Net, which divided into three parts: K-Block, D-Block and I-Block. We evaluated the performance of our KDI-Net model for MRI image reconstruction using PSNR, RMSE, and SSIM metrics and three reduction factors (2, 3, 4) on two publicly available MRI datasets (the AUTOMAP brain dataset and the Calgary Campinas single-coil brain dataset). The results show that KDI-Net performs better than AUTOMAP and Complex AUTOMAP. Compared to AUTOMAP, KDI-Net basically improved each metric by more than 6%. Compared to Complex AUTOMAP, KDI-Net basically improved by more than 8% on each metric. For reduction factor 2, the reconstructed MRI image is very close to ground true. Our specially designed KDI-Net can extract the sparsity from single channel MRI K-space, which achieve 2 folder acceleration.