In this study, we aim to develop a deep complex-valued CNN with an unrolled network structure for general PF reconstruction by iteratively reconstructing PF sampled data and enforcing data consistency. Our proposed method can achieve a better recovery on magnitude and phase images without noise amplification compared to conventional PF reconstruction methods. Further, this approach can be extended to 2D PF reconstruction and joint multi-slice PF reconstruction with complementary sampling across adjacent slices. We train and demonstrate our approach in spin-echo and gradient-echo data, including SWI.
Publicly available datasets, including T1w brain data from Calgary-Campinas Public Brain MR Database1, T2w brain data from fastMRI database2 , and SWI data from OpenNeuro database3, were used for training and evaluating our proposed approach.
[1] https://sites.google.com/view/calgary-campinas-dataset/download?authuser=0
[2] https://fastmri.med.nyu.edu/
[3] https://openneuro.org/datasets/ds000221/versions/1.0.0