RAVEN: Robust, generalizable, multi-resolution structural MRI upsampling using Autoencoders

Abstract

Due to their high inter-tissue contrast, Magnetic resonance images (MRIs) can reflect neuroanatomical changes related to healthy aging and pathological processes. However, standard brain MRI acquisition resolutions hinder the ability to measure the more subtle changes that occur in early disease stages. Increasing the resolution during acquisition poses multiple challenges, including increased noise, higher acquisition times and cost, and discomfort of the scanned individual. In this work, we propose a robust, generalizable single-image super-resolution network for brain MRIs named Resolution Augmentation with Variational auto-Encoder Networks (RAVEN) with generative adversarial networks (GANs). We show RAVEN is capable of upsampling in-vivo and ex-vivo MRIs of diverse modalities (e.g. T1-weighted, T2-weighted, and T2*) and varying field strengths (3T to 7T) to target voxel sizes as small as 0.5mm isotropic using arbitrary upsampling factors. RAVEN achieved state-of-the-art performance against deep learning and non-deep learning methods, best preserving true anatomical information. We have also made RAVEN open access, with the source code as well as training and evaluation scripts available and ready to use at: https://github.com/waadgo/raven.