Abstract
Undersampling the k-space data is widely adopted for acceleration of MagneticResonance Imaging (MRI). Current deep learning based approaches for supervisedlearning of MRI image reconstruction employ real-valued operations andrepresentations by treating complex valued k-space/spatial-space as realvalues. In this paper, we propose complex dense fully convolutional neuralnetwork ($\mathbb{C}$DFNet) for learning to de-alias the reconstructionartifacts within undersampled MRI images. We fashioned a densely-connectedfully convolutional block tailored for complex-valued inputs by introducingdedicated layers such as complex convolution, batch normalization,non-linearities etc. $\mathbb{C}$DFNet leverages the inherently complex-valuednature of input k-space and learns richer representations. We demonstrateimproved perceptual quality and recovery of anatomical structures through$\mathbb{C}$DFNet in contrast to its real-valued counterparts.