Leveraging Deep Learning to Enhance MRI for Brain Disorders

The limited availability and high cost of 7 Tesla (7T) structural MRI hinder its widespread application despite its superior imaging quality. This study introduces a High Frequency-Generative Adversarial Network (HF-GAN) to predict three-dimensional 7T-equivalent (P7T) images from standard 3T structural MRI scans, offering a cost-effective alternative. HF-GAN was trained on paired 3T and 7T MRI data and validated on external datasets, including STRATIFY/ESTRA (N=671) and ADNI2 (N=643), covering psychiatric and neurodegenerative disorders. Results indicate that P7T images generally exhibit enhanced contrast and preservation of fine structural details comparable to 7T and better than 3T, including improved sensitivity in detecting disease-related differences in key brain regions such as the thalamus, caudate, putamen, and frontal cortical areas. The partial 2 values revealed that P7T explained a higher proportion of variance compared to 3T in several comparisons, highlighting its improved sensitivity to disease-related structural changes. These findings demonstrate that HF-GAN effectively enhances 3T MRI data quality, providing a scalable solution for research and clinical applications in neurodegenerative and psychiatric disorders. Additional validations in brain and other organ systems are warranted to further advance clinical translation.