Learning-based reconstruction approaches require a suitable network architecture, training data, and a loss function, to measure the similarity between the reconstructed image and a reference image during training. In this talk, we give an overview of our current developments using variational networks for accelerated MR image reconstruction and discuss several challenges that are encountered during learning, focusing on how the design of the loss function influences reconstruction quality.

This presentation is part of Minisymposium “MS7 - Limited data problems in imaging (2 parts)”
organized by: Bernadette Hahn (University of Würzburg) , Gaël Rigaud (Saarland University) , Jürgen Frikel (OTH Regensburg) .

Authors:

Kerstin Hammernik (Graz University of Technology)

Florian Knoll (New York University)

Erich Kobler (Graz University of Technology)

Teresa Klatzer (Graz University of Technology)

Michael P Recht (New York University School of Medicine)

Daniel Sodickson (New York University)

Thomas Pock (Graz University of Technology)

Keywords:

deep learning, image reconstruction, inverse problems, machine learning, magnetic resonance imaging