Comparison of denoising methods for synthetic low-field neurological MRI

Abstract

Over the past decade, the resurgence of low-field (LF) magnetic resonance imaging (MRI) sensor systems designed to operate up to 1 tesla range has proven well suited to inspire new solutions and design strategies to address frontline medical challenges where environmental factors are most extreme. Examples of successful low-field MRI devices in resource-limited environments include: (1) classification of infant hydrocephalus in Africa and (2) in remote small hospitals where the low-field scanner travels to the patient's bedside to observe volumetric changes in brain structure [1,2]. These low-field MRI design examples have been effective in leveraging MRI information in the setting where it is used with minimal available resources. In developing countries, access to high-field (HF) MRI is limited and requires support and infrastructure to be used. These lower-cost systems can potentially benefit from many developments that have occurred in higher fields, such as signal-to-noise dependence on static magnetic fields and hardware components (i.e., magnet, gradient coils, etc.) [3]. In addition, improvements in machine learning now provide superior noise reduction compared to traditional methods, resulting in improved performance with smaller size and lower power consumption. With improved access to medical imaging equipment, people around the world who cannot afford it due to the high cost of conventional MRIs will be able to obtain high- quality imaging data with improved contrast resolution and acquisition times. SNR (signal-to-noise ratio) is an important measure of the quality of a signal in low-field MRI. The amount of useful information in a signal compared to background noise directly affects the effectiveness of a low-field scan [4].