Abstract:Currently, most accelerated parallel magnetic resonance imaging (PMRI) iterative reconstruction algorithms use real-valued convolutions to reconstruct under sampled MR magnitude images or even real-part images. However, this approach does not take into account the fact that multi-coil MRI inherently involves complex-valued data, thus limiting the reconstruction accuracy of iterative algorithms for MR images. Given that MRI data is inherently complex-valued, and in order to be able to reconstruct multi-coil MRI images with more accurate texture details, a complex-valued attention-enhanced iterative network (CAEI-Net) is proposed in this study for simultaneous reconstruction of MRI images and sensitivity maps. The complex prior of MRI images and sensitivity maps is automatically learned using a complex-valued neural network. Experimental results on the FastMRI dataset show that the proposed reconstruction algorithm achieves higher reconstruction accuracy compared to previous algorithms and better reconstructs multi-coil MRI images.