Supervisors: Kerem Tezcan, Prof. Ender Konukoglu
Magnetic Resonance (MR) data acquisition is an inherently slow process. Shorter scan times are desirable, for reasons of patient comfort or artifact reduction due to motion. It might even be that the MR experiment tries to resolve motion, for instance in cardiovascular imaging applications. To this end an extensive list of methods were proposed to speed up scan times, by simply acquiring less data to begin with and remove artifacts that are induced by the procedure in some clever way. This report investigates a modern approach to reconstruction leveraging the recent success of high dimensional statistics. Reconstruction is formulated in a Bayesian perspective assuming a probability measure on the space of MR images. This distribution is learned from data using a multiscale architecture built from blocks of Continuous Normalizing Flows (CNF). The importance of hyper - parameters of this multiscale architecture are studied and reconstructions performed with the best performing models.