Super Resolution Techniques Applied to Low-Field MRI

Abstract

This work is part of the low-fieldMRI project, which aims to bring portable, affordable, low-fieldMRI scanners to low-income countries. Replacing the superconducting magnets of conventional scanners with standard ones can significantly reduce the costs, but it also has a negative impact on the Signal-to-Noise Ratio (SNR). In order to circumvent this problem, Super Resolution (SR) techniques may be used. In this thesis, standard and Deep Learning (DL) SR techniques are presented. For standard SR, two types of regularization are considered: Tikhonov and Total Variation (TV). Then, the problemis solved using CGLS and ADMM algorithms respectively. From our analysis, we could conclude that TV outperforms Tikhonov regularization, yielding promising results. Then, two 2D DL models, SRCNN and ReCNN, were implemented and trained on two commonly used SR datasets: T91 and Kirby21. Both networks managed to reconstruct the LR scans surprisingly well, with ReCNN yielding the best results when trained on both datasets. DL methods evidently outperform standard SR and can achieve a visual quality comparable to the one of a scan directly acquired in higher resolution. A 3D extension of these networks was also considered, but, although it led to an improvement, it did not perform as well as the 2D models. We attribute this to the lack of time, which did not allow us to extensively explore this possibility.