Synopsis

Keywords: Image acquisition: Sequences, Image acquisition: Machine learning

Artificial intelligence (AI) provides new tools to solve complex optimization problems, such MRI pulse sequence design. AI can be used to efficiently design the RF pulses and gradient waveforms to achieve the required image contrast and k-space trajectory or to optimize the schedule of acquisition parameters in MR fingerprinting, according to hardware constraints such as gradient amplitude and slew rate. This lecture will review techniques that use AI to design MRI pulse sequences such as AutoSeq, MRZero, and optimization of parameter schedule for MR fingerprinting. The application of ChatGPT for auto-generation of pulse sequence code will be also discussed.

Abstract

Pulse sequence design involves arranging RF pulses and gradient waveforms to achieve the targeted image contrast and k-space trajectory, according to hardware constraints (gradient amplitude and slew-rate, SAR). To maximize acquisition efficiency, the duration of RF pulses and gradient waveforms needs to be minimized. Given the large number of variables, the optimization problem is very challenging, and the usual approach is to use approximate and/or handcrafted solutions. The introduction of artificial intelligence (AI) algorithms, and particularly deep neural networks, enables to solve complex optimization problems, such MRI pulse sequence design. AUTOmated pulse SEQuence generation (AutoSeq) used reinforcement learning to generate a 1D pulse sequence (1). A follow-up version of AutoSeq used a multilayer fully-connected neural network to perform fast quantitative T1 and T2 parameter mapping (2). A step forward in the auto generation of pulse sequences was introduced in the MRZero technique, where a differentiable MRI simulation model was employed along with a reconstruction algorithm to train the network using a loss function defined between the reconstructed and target images (3). MRZero demonstrated artifact suppression and SAR reduction compared to conventional pulse sequences. MRZero was also applied to reduce blurring in fast spin echo acquisitions by optimizing the flip angle values using a T2 decay model during training (4). In the context of quantitative imaging using MR fingerprinting, AI can be used to optimize the acquisition schedule parameter to improve quantification accuracy and/or shorten the acquisition. For example, the flip angle schedule was optimized to reduce the acquisition time and improve the quality of MR fingerprinting chemical exchange saturation contrast (MRF-CEST) in brain tumors (5). More recently, the large language model ChatGPT was employed to autogenerate pulse sequence code based on PyPulseq in the GPT4MR technique (6). This lecture will cover the principles of using AI algorithms to design pulse sequences and to automatically generate pulse sequence programming code.

Acknowledgements

Ouri Cohen from Memorial Sloan Kettering Cancer Center; Moritz Zaiss and Florian Knoll from Friedrich-Alexander Universität Erlangen-Nürnberg; Matthew Rosen from MGH/Harvard Martinos Center