Magnetic Resonance Imaging (MRI) is one of the leading diagnostic modalities capable of providing high-resolution images. However, the scan cost of MRI is higher due to its slow acquisition time. The scan time can be accelerated through undersampling k-space, which would result in a degraded image. Recently, deep learning has shown promising results in the restoration of degraded images. However, the deep learning methods have some limitations. To begin with, the proposed networks haven't effectively utilized the image and k-space domain data. Second, the methods are yet to explore the benefits of related MRI sequences thoroughly. Third, the deep networks are not optimized for the metric, which correlates with the radiologist's opinion on image quality. Furthermore, the radiologists view on diagnosis has not been incorporated in the design of deep networks. Finally, the deep networks have become parameter intensive and require large memory storage.

In this work, we develop networks to address the limitations mentioned above. First, we propose novel MRI reconstruction architectures Deep Cascade ReconSynergyNet (DC-RSN) and Variable Splitting RSN (VS-RSN) for single and multi-coil acquisition, which effectively utilizes both k-space and image domain data. Second, we propose to improve the reconstruction quality of the T2 sequence through Gradient of Log Feature (GOLF) based T1 sequence assistance. Third, we propose a Perceptual Refinement network (PRN) to refine the reconstructions and improve the surrogate metric for radiologists' opinions. Furthermore, we propose Reconstruction Global Local Generative Adversarial Networks (Recon-GLGAN) to leverage the prior information obtained from radiologist's practice. Finally, we propose attention-based distillation and imitation loss to improve the reconstruction quality of the lesser parameter network. We have extensively validated all the proposed methods with suitable experiments.