Ultra‐Low‐Field Paediatric MRI in Low‐ and Middle‐Income Countries: Super‐Resolution Using a Multi‐Orientation U‐Net

Abstract

Owing to the high cost of modern magnetic resonance imaging (MRI) systems, their use in clinical care and neurodevelopmental research is limited to hospitals and universities in high income countries. Ultra‐low‐field systems with significantly lower scanning costs present a promising avenue towards global MRI accessibility; however, their reduced SNR compared to 1.5 or 3 T systems limits their applicability for research and clinical use. In this paper, we describe a deep learning‐based super‐resolution approach to generate high‐resolution isotropic T2‐weighted scans from low‐resolution paediatric input scans. We train a ‘multi‐orientation U‐Net’, which uses multiple low‐resolution anisotropic images acquired in orthogonal orientations to construct a super‐resolved output. Our approach exhibits improved quality of outputs compared to current state‐of‐the‐art methods for super‐resolution of ultra‐low‐field scans in paediatric populations. Crucially for paediatric development, our approach improves reconstruction of deep brain structures with the greatest improvement in volume estimates of the caudate, where our model improves upon the state‐of‐the‐art in: linear correlation (r = 0.94 vs. 0.84 using existing methods), exact agreement (Lin's concordance correlation = 0.94 vs. 0.80) and mean error (0.05 cm3 vs. 0.36 cm3). Our research serves as proof‐of‐principle of the viability of training deep‐learning based super‐resolution models for use in neurodevelopmental research and presents the first model trained exclusively on paired ultra‐low‐field and high‐field data from infants.