Nam Dang | Searcharxiv

arXiv2024

MR-zero meets FLASH -- Controlling the transient signal decay in gradient- and rf-spoiled gradient echo sequences

Abstract Purpose The complex signal decay during the transient FLASH MRI readout can lead to artifacts in magnitude and phase images. We show that target-driven optimization of individual rf flip angles and phases can realize near-ideal signal behavior and mitigate artifacts. Methods The differentiable end-to-end optimization framework MR-zero is used to optimize rf trains of the FLASH sequence. We focus herein on minimizing deviations from the ideally spoiled signal by using a mono-exponential Look-Locker target. We first obtain the transient FLASH signal decay substructure, and then minimize the deviation to the Look-Locker decay by optimizing the individual (i) flip angles, (ii) rf phases and (iii) flip angles and rf phases. Comparison between measurement and simulation are performed using Pulseq in 1D and 2D. Results We could reproduce the complex substructure of the transient FLASH signal decay. All three optimization objectives can bring the real FLASH signal closer to the ideal case, with best results when both flip angles and rf phases are adjusted jointly. This solution outperformed all tested conventional quadratic rf cyclings in terms of (i) matching the Look-Locker target signal, (ii) phase stability, (iii) PSF ideality, (iv) robustness against parameter changes, and (v) magnitude and phase image quality. Other target functions for the signal could as well be realized, yet, their response is not as general as for the Look-Locker target and need to be optimized for a specific context. Conclusion Individual flip angle and rf phase optimization improves the transient signal decay of FLASH MRI sequences.

physics.med-ph↗

arXiv2021

MRzero -- Fully automated discovery of MRI sequences using supervised learning

Purpose: A supervised learning framework is proposed to automatically generate MR sequences and corresponding reconstruction based on the target contrast of interest. Combined with a flexible, task-driven cost function this allows for an efficient exploration of novel MR sequence strategies. Methods: The scanning and reconstruction process is simulated end-to-end in terms of RF events, gradient moment events in x and y, and delay times, acting on the input model spin system given in terms of proton density, T1 and T2, and $Δ$B0. As a proof of concept, we use both conventional MR images and T1 maps as targets and optimize from scratch using the loss defined by data fidelity, SAR penalty, and scan time. Results: In a first attempt, \textit{MRzero} learns gradient and RF events from zero, and is able to generate a target image produced by a conventional gradient echo sequence. Using a neural network within the reconstruction module allows arbitrary targets to be learned successfully. Experiments could be translated to image acquisition at the real system (3T Siemens, PRISMA) and could be verified in the measurements of phantoms and a human brain \textit{in vivo}. Conclusions: Automated MR sequence generation is possible based on differentiable Bloch equation simulations and a supervised learning approach.

physics.med-ph↗