Felix Glang1, Alexander Loktyushin1, Kai Herz1,2, Hoai Nam Dang3, Anagha Deshmane1, Simon Weinmüller3, Arnd Doerfler3, Andreas Maier4, Bernhard Schölkopf5, Klaus Scheffler1,2, and Moritz Zaiss1,3
1High-field Magnetic Resonance Center, Max-Planck Institute for Biological Cybernetics, Tübingen, Germany, 2Department of Biomedical Magnetic Resonance, Eberhard Karls University Tübingen, Tübingen, Germany, 3Friedrich-Alexander Universität Erlangen-Nürnberg (FAU), Neuroradiology, University Clinic Erlangen, Erlangen, Germany, 4Pattern Recognition Lab, Friedrich-Alexander-University Erlangen-Nürnberg, Erlangen, Germany, 5Empirical Inference, Max-Planck Institute for Intelligent Systems, Tübingen, Germany
1High-field Magnetic Resonance Center, Max-Planck Institute for Biological Cybernetics, Tübingen, Germany, 2Department of Biomedical Magnetic Resonance, Eberhard Karls University Tübingen, Tübingen, Germany, 3Friedrich-Alexander Universität Erlangen-Nürnberg (FAU), Neuroradiology, University Clinic Erlangen, Erlangen, Germany, 4Pattern Recognition Lab, Friedrich-Alexander-University Erlangen-Nürnberg, Erlangen, Germany, 5Empirical Inference, Max-Planck Institute for Intelligent Systems, Tübingen, Germany
The
proposed approach allows fully differentiable supervised learning of MRI
sequences, including parallel imaging, non-Cartesian trajectories and flip
angle optimization. By that, it extends the recently proposed MRzero framework.
Figure 3. Optimization of free non-Cartesian k-space
trajectory and RF flip angles for parallel imaging (R=4) at various
intermediate iterations, including the final result at the last iteration. (A)
Reconstructions, (B) difference maps to the target image (Figure 2A) and (C)
current trajectory (colors corresponding to shots) and (D) current flip angles
across the iterations. (E) Loss curve, i.e. normalized root mean squared error to
target image over iterations. An animated version of the Figure can be found at https://tinyurl.com/ydgb9gf8
Figure 2. Optimization of free non-Cartesian k-space
trajectory for parallel imaging (R=3) at various intermediate iterations,
including the final result at iteration 2550. (A) Reconstructions, (B)
difference maps to the target image (Figure 2A) and (C) current trajectory
(colors corresponding to shots) across the iterations. (D) Loss curve, i.e.
normalized root mean squared error to the target image over iterations.
