The preprint of our new paper: "Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast" is available on Arxiv. Using synthetic brain MRI scans generated from segmentations, we train a convolutional neural network that learns to generate 1 mm MP-RAGES from a set of clinical MRI scans of the same subject, acquired anisotropically with different slice thickness, spacings, and orientations. The synthesized scans can be subsequently processed with standard neuroimaging tools for registration and segmentation, e.g., FreeSurfer. Because the training data are synthetic, one can easily train a network for a new set of MR contrasts / resolutions without having to obtain any new training data. We believe that this is an important step towards quantitative clinical neuroimaging and towards exploiting huge amounts of existing clinical data in neuroimaging studies.
The preprint is available here: https://arxiv.org/pdf/2012.13340.pdf