Abstract
This paper presents a generative model for super-resolution in routineclinical magnetic resonance images (MRI), of arbitrary orientation andcontrast. The model recasts the recovery of high resolution images as aninverse problem, in which a forward model simulates the slice-select profile ofthe MR scanner. The paper introduces a prior based on multi-channel totalvariation for MRI super-resolution. Bias-variance trade-off is handled byestimating hyper-parameters from the low resolution input scans. The model wasvalidated on a large database of brain images. The validation showed that themodel can improve brain segmentation, that it can recover anatomicalinformation between images of different MR contrasts, and that it generaliseswell to the large variability present in MR images of different subjects. Theimplementation is freely available at https://github.com/WCHN/CA_MTV-preproc.