Abstract
Compressive sensing (CS) is widely used to reduce the acquisition time ofmagnetic resonance imaging (MRI). Although state-of-the-art deep learning basedmethods have been able to obtain fast, high-quality reconstruction of CS-MRimages, their main drawback is that they treat complex-valued MRI data asreal-valued entities. Most methods either extract the magnitude from thecomplex-valued entities or concatenate them as two real-valued channels. Inboth the cases, the phase content, which links the real and imaginary parts ofthe complex-valued entities, is discarded. In order to address the fundamentalproblem of real-valued deep networks, i.e. their inability to processcomplex-valued data, we propose a novel framework based on a complex-valuedgenerative adversarial network (Co-VeGAN). Our model can process complex-valuedinput, which enables it to perform high-quality reconstruction of the CS-MRimages. Further, considering that phase is a crucial component ofcomplex-valued entities, we propose a novel complex-valued activation function,which is sensitive to the phase of the input. Extensive evaluation of theproposed approach on different datasets using various sampling masksdemonstrates that the proposed model significantly outperforms the existingCS-MRI reconstruction techniques in terms of peak signal-to-noise ratio as wellas structural similarity index. Further, it uses significantly fewer trainableparameters to do so, as compared to the real-valued deep learning basedmethods.