DeepControl: 2DRF pulses facilitating inhomogeneity and B₀ off‐resonance compensation in vivo at 7 T

Abstract

Funding information Kong Christian den Tiendes Fond; Harboefonden; Eva og Henry Fraenkels Mindefond; VILLUM FONDEN Purpose: Rapid 2DRF pulse design with subject-specific B + 1 inhomogeneity and B 0 off-resonance compensation at 7 T predicted from convolutional neural networks is presented. Methods: The convolution neural network was trained on half a million singlechannel transmit 2DRF pulses optimized with an optimal control method using artificial 2D targets, B + 1 and B 0 maps. Predicted pulses were tested in a phantom and in vivo at 7 T with measured B + 1 and B 0 maps from a high-resolution gradient echo sequence. Results: Pulse prediction by the trained convolutional neural network was done on the fly during the MR session in approximately 9 ms for multiple hand-drawn regions of interest and the measured B + 1 and B 0 maps. Compensation of B + 1 inhomogeneity and B 0 off-resonances has been confirmed in the phantom and in vivo experiments. The reconstructed image data agree well with the simulations using the acquired B + 1 and B 0 maps, and the 2DRF pulse predicted by the convolutional neural networks is as good as the conventional RF pulse obtained by optimal control. Conclusion: The proposed convolutional neural network-based 2DRF pulse design method predicts 2DRF pulses with an excellent excitation pattern and compensated B + 1 and B 0 variations at 7 T. The rapid 2DRF pulse prediction (9 ms) enables subjectspecific high-quality 2DRF pulses without the need to run lengthy optimizations. K E Y W O R D S 2DRF pulses, 7 T, artificial intelligence, deep learning, optimal control 1 | INTRODUCTION Multidimensional spectral-/spatial-selective RF pulses have applications in inner-volume imaging, 1,2 navigation, 3,4 fMRI, 5,6 DWI, 7 DTI, 8 MRS, 9 carbon-13 dissolution dynamic nuclear polarization, 10,11 flow imaging, 12,13 and suppression. 14 RF pulse designers in general navigate between matters of pulse types; spatial (and/or spectral) target profiles; inhomogeneous B + 1 and B 0 fields; underlying gradient waveforms; system imperfections 15 ; constraints (eg, power, 16 specific absorption rate [SAR], 17 and states 18,19); and finally, which pulse design tool and physics model to adopt for the task

Ultrafast (milliseconds), multidimensional RF pulse design with deep learning

“…Considering the vast training libraries required for DeepControl 22,23 , the non-parallelized, vectorized computation of the standard and midpoint approximate gradients is not a problem, when we exploit all available CPU cores with an outer parallelization over many independent pulses.…”

“…The choice for target categories reflect our previous [22][23][24] and future DeepControl experiments, which will be described in a subsequent publication. We acknowledge that routines operating in the so-called small-flip-angle regime pose a robust alternative to optimal control in terms of speed and accuracy due to the linear Fourier relation existing between the pulse waveform and the excitation pattern.…”

“…The four gradients types have been implemented 37 in the blOCh framework used in Refs. 2,4,13,14,22,23,25 .…”

“…We have recently proposed a neural network based method for generating RF pulses to alleviate the run time problem [22][23][24] . This DeepControl framework requires training libraries with hundreds of thousands of RF pulses generated, e.g., with optimal control, but a trained neural network predicts a pulse quickly, and with a hard guarantee on the wall clock time.…”

Deep reinforcement learning-designed radiofrequency waveform in MRI

“…In this study, we propose a new RF pulse design framework, DeepRF, which utilizes the power of the self-learning characteristic of DRL to explore novel mechanisms for magnetization manipulation beyond conventional methods. This feature clearly differentiates DeepRF from previously proposed deep learning-based RF design methods using supervised learning [16][17][18][19] . Moreover, unlike conventional RF design algorithms that are tailored for a specific RF pulse type, DeepRF can produce RF pulses of different goals via a customized reward function.…”