Deep-Learning-based MR Compressed Sensing using Non-randomly Under-sampled Signal in Nonlinear Phase Encoding Imaging

Optional image scaling, and hence aliasless image reconstruction, is feasible using a signal that violates the sampling theorem in MR phase scrambling Fourier transform imaging. In this method, the main and aliased image components are separated in the scaled space when a large scaling factor is selected. In the present study, a new fast imaging method, in which aliasing artifacts caused by undersampling of the signal, are removed in two steps: in the downscaled space introduced by aliasless reconstruction and through de-aliasing using a deep convolution neural network. The proposed method is shown to provide higher PSNR images compared to random sampling compressed sensing and has an advantage in terms of low-sampling-rate image acquisition.