Magnetic resonance imaging (MRI) is prone for subject motion during the examination. Despite prospective approaches try to avoid, diminish or adapt to the motion during acquisition not all methods are always applicable, lead to increased acquisition times or still have residual artifacts. Retrospective methods are applied after the examination to reduce artifacts by adapting the acquired data.

By modelling the motion distortion and the ideal image, a highly non-linear optimization problem can be proposed, which searches for the optimal image and the motion parameters. This method is widely discussed in literature as it gradually approaches the artifact free image based on a physical model. One disadvantage is the convergence speed, which makes it not suitable for clinical use.

This research project investigates several ways how to overcome this limitation speed using machine learning and neural networks.