A Frequency Domain Network for Hybrid-Domain Super-Resolution MRI

Hybrid-domain networks with sequentially or parallelly connected Frequency-domain (FDN) and Spatial-domain (SDN) networks, have been extensively used for Compressed-Sensing MRI (CS-MRI). As CS-MRI scans involve randomly acquired high-frequency data through variable-density (VD) sampling, hybrid-domain networks are trained to distinguish incoherent structural artefacts from the true image features. A key difference of super-resolution MRI (SR-MRI) from CS-MRI is the absence of high-frequency measurements. Therefore, super-resolving the image structures from a truncated acquisition would entail estimation of the high-frequency data from a low-resolution (LR) scan. Due to this reason, the frequency-domain component of the hybrid-domain CNN architecture for recovering high-frequency (HF) information from a truncated Fourier acquisition should be designed to distinguish artefacts that are highly correlated to the underlying image structures. This is achieved by designing the frequency-domain CNN architecture with the goal of extrapolating the unknown high-frequency data points with improved filling factor. When used in a hybrid domain configuration, Frequency Domain Super-Resolution Network (FDSRN) is able to achieve improved super-resolution performance in comparison to the existing hybrid-domain networks; and some of the recent spatial domain deep networks by employing either multi-contrast training samples, or multi-streaming architecture and having selectively demonstrated capability of super-resolving LR images from a band-limited acquisition.