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 scanners restrict their use to high-income settings. To address this, more affordable and energy-efficient ultra-low-field MRI scanners have been developed. However, the reduced resolution and signal-to-noise ratio of the resulting scans limit their clinical utility, motivating the development of super-resolution techniques. The current state-of-the-art super-resolution methods require 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 super-resolution 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 for the target population, this study explores the efficacy of using a deep learning model, the 3D UNet, to generate higher-resolution brain scans from just one ultra-low-field scan. The model was trained to receive a single ultra-low-field brain scan of 6-month-old infants and produce a scan of MRR quality. Results showed a significant improvement in the quality of output scans compared to input scans, including increased image quality metrics, stronger correlations in tissue volume estimates across participants, and greater Dice overlap of the underlying tissue segmentations to those of target scans. The study demonstrates that the 3D UNet effectively enhances the resolution of ultra-low-field infant MRI scans. Generating higher-resolution brain scans from single ultra-low-field scans, without needing paired high-field data, reduces scanning time and supports wider MRI use in low- and middle-income countries. Additionally, this approach allows for easier model training on a site- and population-specific basis, enhancing adaptability in diverse settings.