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

Abstract

Purpose Some advanced RF pulses, like multidimensional RF pulses, are often long and require substantial computation time because of a number of constraints and requirements, sometimes hampering clinical use. However, the pulses offer opportunities of reduced‐FOV imaging, regional flip‐angle homogenization, and localized spectroscopy, e.g., of hyperpolarized metabolites. Proposed herein is a novel deep learning approach to ultrafast design of multidimensional RF pulses with intention of real‐time pulse updates. Methods The proposed neural network considers input maps of the desired excitation region of interest and outputs a single‐channel, multidimensional RF pulse. The training library is, e.g., retrieved from a large image database, and the target RF pulses trained upon are calculated with a method of choice. Results A relatively simple neural network is enough to produce reliable 2D spatial‐selective RF pulses of comparable performance to the teaching method. For binary regions of interest, the training library does not need to be vast; hence, reestablishment of the training library is not necessarily cumbersome. The predicted pulses were tested numerically and experimentally at 3 T. Conclusion Relatively effortless training of multidimensional RF pulses, based on non‐MRI‐related inputs, but working in an MRI setting still, has been demonstrated. The prediction time of a few milliseconds renders real‐time updates of advanced RF pulses possible.

Multi‐task convolutional neural network‐based design of radio frequency pulse and the accompanying gradients for magnetic resonance imaging

“…We developed a multi‐task CNN that was proven to design spatially 2D excitation RF pulses. To our knowledge, this is the second work in DL‐based RF design, next only to Vinding et al, 15 and the first work in CNN‐based RF design for both RF and gradient waveforms. We aimed to provide real‐time RF design at each specific scan for optimized RF performance under the given imaging conditions, including the hardware restrictions (RF and gradient constraints and the number of transmission channels), 31 system imperfections (eddy currents and B 0 inhomogeneities), 32,33 imaging object‐dependent challenges (local magnetic field inhomogeneities in terms of both B 0 and B 1 ) 34 and prescription‐dependent requirements (FOV of interest).…”

“…By setting up this improved neural network, we have extended the initial work in the field 15 by 3‐fold in output dimensions (one RF dimension and two gradient dimensions versus a single RF waveform dimension), 2.5‐fold in time resolution of the waveforms (4 versus 10 μs) and ~10‐fold in the number of points in each waveform (~6000 versus ~600 points), yielding a total of 3 (‐fold dimensions) x ~10 (‐fold points in each dimension) = ~30‐fold the amount of data in each design. When compared with the neural network‐based RF design with prefixed gradient waveforms, the 3D design output has the potential for better cooperation between the RF and the gradient waveforms, as these waveforms together determine the weighting in the excitation k‐space, and the gradient waveforms alone determine the excitation k‐space sampling pattern, both of which are important to the excitation profile 5 .…”

“…For the baseline of the DL performance, we established a typical fully connected neural network (FC‐NN) with similar structures as in Vinding et al 15 As shown in Figure 2B in the red box, the FC‐NN consisted of an input layer, four fully connected layers with sizes of 8192, 6000, 5120 and 4092 with ReLU layers in between, and finally a regression layer of size 6000 with sigmoid activation to yield the output signal with matched size. A dropout rate of 0.5 was employed to reduce the risk of overfitting 30 .…”

“…Vinding et al recently reported an RF design method assisted by deep learning (DL), where a fully connected neural network was trained to produce the desired RF waveform that achieved 2D spatial selectivity, with predesigned gradient waveforms, 15 resulting in milliseconds of computation time and allowing for real‐time waveform generation. A DL approach could accelerate the application of optimized pulses by training a neural network that could design a broad spectrum of shaped pulses.…”

Predicting in vivo MRI Gradient-Field Induced Voltage Levels on Implanted Deep Brain Stimulation Systems Using Neural Networks

“…Computational modeling and simulations can be used to generate large synthetic datasets. Incorporating ML with computational modeling and simulations has a potential to provide better understanding of biological and physical phenomena, solve illposed inverse problems, optimize complex design problems in a variety of fields (Burrascano et al, 1999;Kim et al, 2007;Tolk, 2015;Hughes et al, 2017;Cohen et al, 2018;Ianni et al, 2018;Pérez et al, 2018;Deist et al, 2019;Kiarashinejad et al, 2019;Meliadò et al, 2019;Tahersima et al, 2019;Vinding et al, 2019). In this work, we have generated a dataset by modeling and simulating MRI gradient-field induced voltage levels on implanted DBS systems, using realistic MRI gradient coil models, six adult anatomical human models (Gosselin et al, 2014) and clinically relevant DBS implant trajectories.…”