Learning a Spatiotemporal Dictionary for Magnetic Resonance Fingerprinting with Compressed Sensing

Abstract

Magnetic resonance fingerprinting (MRF) is a novel technique that allows for the fast and simultaneous quantification of multiple tissue properties, progressing from qualitative images, such as T1- or T2-weighted images commonly used in clinical routines, to quantitative parametric maps. MRF consists of two main elements: accelerated pseudorandom acquisitions that create unique signal evolutions over time and the voxel-wise matching of these signals to a dictionary simulated using the Bloch equations. In this study, we propose to increase the performance of MRF by not only considering the simulated temporal signal, but a full spatiotemporal neighborhood for parameter reconstruction. We achieve this goal by first training a dictionary from a set of spatiotemporal image patches and subsequently coupling the trained dictionary with an iterative projection algorithm consistent with the theory of compressed sensing (CS). Using data from BrainWeb, we show that the proposed patch-based reconstruction can accurately recover T1 and T2 maps from highly undersampled k-space measurements, demonstrating the added benefit of using spatiotemporal dictionaries in MRF.