Artificial Intelligence methods for ultra-low- field MRI enhancement comparing brain volumes and neurodevelopment scores for small vs appropriate for gestational age children

Globally, nearly 250 million children do not achieve their developmental milestones in early childhood. Perinatal factors including low birth weight and being born small for gestational age (SGA) are associated with low fine motor scores leading to poor school performances. The current study aimed to compare brain volumes derived from super-resolved ultra-low-field (SR-ULF) and high-field (HF) MRI and correlate neurodevelopmental scores with the SR-ULF images to assess differences among SGA and appropriate for gestational age (AGA) children. A total of 180 children (SGA=90; AGA=90) from the Doppler cohort conducted in 2021 were re-approached to participate of which 64 between the age of 3-5 years were included in the final analysis. Ultra-low-field scans acquired using the Hyperfine SwoopTM 64mT system and neurodevelopmental scores using Malawi Developmental Assessment tool (MDAT) were assessed. SR-ULF images were generated using the pre-trained GAMBAS deep learning model while standard axial T1 and T2 sequences were acquired using HF MRI on a subset of participants (n=10) for ground truth comparison. SR-ULF and HF brain volumes showed strong correlations and high concordance for total intracranial volume, white matter, cerebral cortex, basal ganglia, and thalamus. However, correlations between SR-ULF brain volumes and MDAT z-scores were not statistically significant. Portable scans enhanced with deep learning methods can be implemented reliably in low resource settings where conventional MRI is a challenge, including in at-risk populations such as infants born SGA. Future research should aim to optimize segmentation algorithms and extend applications of SR-ULF MRI in longitudinal studies with larger samples.