MRzero - Automated Discovery of MRI Sequences Using Supervised Learning

RF, Encoding and timing task. A, k-space sampling at different iterations. B, Flip angles over measurement repetitions. C, TR and TE over measurement repetitions. D, Simulation-based reconstruction at different iterations. E, Phantom measurement, F, In vivo brain scan. G, Training error curve. Target sequence: 2D transient gradient- and RF-spoiled GRE, matrix size 48, TR = 20 ms, TE = 3 ms, FA = 5 deg.

Magnetic resonance (MR) images can be created noninvasively using only static and dynamic magnetic fields, and radio frequency pulses. MR imaging provides fast image acquisitions which have been clinically feasible only since the discovery of efficient MR sequences, ie. time-efficient application of two building blocks: radio frequency pulses and spatial magnetic field gradients. The proper arrangement of these building blocks is the crucial step for MR sequence development. In the context of MR in medicine, generating contrast between tissues is of central importance. Here MR shows outstanding properties especially in soft tissues, leading to many applications in routine medical imaging with specialized MR sequences for a certain contrast of interest. The direct relationship between MR image contrast and the actual MR sequence with its many free parameters raises the question if both image and contrast generation can be performed in a completely automatic manner. Many recent works addressed this task by analytic optimization of RF pulses and image acquisition parameters.
In the present work, we propose a joint automatic sequence and reconstruction optimization framework, MRzero, based on supervised learning and a fully differentiable MR simulation process. This enables gradient descent in the sequence parameter space, as well as efficient generalization using multiple input and target samples. We show how basic image encoding as well as artifact suppression and SAR minimization can be achieved. Furthermore, we show how arbitrary target images can be used when using a neural network layer added to the reconstruction, and how possible overfitting can be tackled in an efficient manner. MRzero performs the optimization in a simulation environment—a differentiable simulated replica of the MR system. However, the learned sequences were implemented and tested on a real clinical MR system and scanned both in vitro and in vivo.

1.
A. Loktyushin, K. Herz, N. Dang, F. Glang, A. Deshmane, S. Weinmüller, A. Doerfler, B. Schölkopf, K. Scheffler, M. Zaiss:
MRzero - Automated discovery of MRI sequences using supervised learning.
Magn Reson Med, 2021 Aug; 86(2):709-724.
Go to Editor View