Deep learning super-resolution of paediatric ultra-low-field MRI without paired high-field scans

Abstract

Brain magnetic resonance imaging (MRI) is essential for diagnosis and neurodevelopmental research, but the high cost and infrastructure demands of high-field MRI limit its use to high-income settings. Ultra-low-field MRI scanners offer a more affordable and energy-efficient alternative, but their reduced resolution and signal-to-noise ratio restrict research and clinical utility, prompting the need for super-resolution techniques. Current super-resolution methods rely on either three anisotropic ultra-low-field scans acquired at different orientations (axial, coronal, sagittal) to reconstruct a higher-resolution image using multi-resolution registration (MRR) or the training of deep learning models using paired ultra-low- and high-field scans. Since acquiring three high-quality ultra-low-field scans is not always feasible, and paired high-field data may not be available, this study explores the efficacy of using a deep learning model to generate scans of MRR quality from a single ultra-low-field input scan. Results demonstrated significant enhancement in the quality of output scans, including improved image quality metrics, stronger tissue volume correlations, and greater Dice overlap of tissue segmentations. Generating higher-resolution brain scans from single ultra-low-field scans, without paired high-field data, reduces scanning time and further widens MRI accessibility in low- and middle-income countries. This approach also facilitates site-specific model training, which an exploratory external validation suggests may be necessary to address potential domain shifts across scanning sites.