MRI contrast synthesis from low-rank coefficient images

Abstract

Synthetic contrasts are commonly derived from parameter maps via Bloch simulation.Typically, model imperfections, in particular partial volume effects, cause artifacts in those images. Recently, it has been proposed to overcome this problem by mapping directly from MR-Fingerprinting data to synthetic contrasts with neural networks. Those methods, however, face the MRF-typical undersampling artifacts, as well as the computational burden of hundreds of input images. We propose to first reconstruct images in a low-rank sub-space, which maintains the correct partial volume contrast, but allows for removal of undersampling artifacts, and to map from this space to synthetic contrasts with a neural network.