Recovery of high-resolution magnetic field distribution inside the brain from limited MRI data using machine learning prior

Abstract

High-resolution field maps in brain magnetic resonance imaging (MRI) scans provide the field distribution information inside the brain which is essential in reconstructing high-quality MR images with no artifacts and distortions. These high-quality images are highly desired in clinical applications. However, the high-resolution field maps, which are used to obtain high-quality MR images, come with the cost of scan time. Recent advances in deep neural networks, particularly the generative adversarial networks (GANs), can learn the prior information through examples and generate the high-resolution field map using only one low-resolution field map counter- part. In this work, we apply the deep learning methods to solve the field map super-resolution problem and show that our GAN-based approach has the potential to generate the high-resolution field maps as a post-processing step and to speed up many clinical MRI applications.