1. Chandarana H et al. Invest Radiol. 2011 Oct;46(10):648-53
2. Feng L et al. Magn Reson Med 2014 2014 Sep;72(3):707-17
3. Chandarana H et al. Invest Radiol. 2013 Jan;48(1):10-6
4. Block KT et al. ISMRM 2013. p 3809
5. Feng L et al. Magn Reson Med 2016 2016 Feb;75(2):775-88
6. Piccini D et al. Magn Reson Med 2017 Apr;77(4):1473-1484
7. Feng L et al. Magn Reson Med 2018 Feb;79(2):826-838
10. Benkert et al. ISMRM 2017. p 902
11. Feng L et al. Magn Reson Med 2017 2018 Jul;80(1):77-89
12. Feng et al. ISMRM 2017. p 4001
13. Benkert et al. Magn Reson Med 2017 2018 Jul;80(1):286-293
14. Feng L et al. JMRI 2019 2019 Feb;49(2):411-422
15. Feng L et al. Magn Reson Med 2020 Jan;83(1):94-108
GRASP MRI
Over the past decade, I have been the main driver for the development, optimization and extension of a rapid MRI technique called GRASP (Golden-angle RAdial Sparse Parallel imaging, see above the roadmap of this project). In 2010, my colleagues Drs. Hersh Chandarana and Tobias Block demonstrated the clinical use of stack-of-stars sampling for free-breathing 3D MRI. The stack-of-stars trajectory was relatively new at that time and the so-called golden-angle sampling was not widely used yet. Since then, this trajectory has attracted substantial attention and interest in the MRI research community and has been implemented by different vendors now. I stared involving in the project since 2011 and developed the GRASP technique by combining compressed sensing, parallel imaging and golden-angle stack-of-stars sampling. GRASP introduces a new and simple way of performing dynamic MRI. It enables free-breathing continuous data acquisition and employs sparse reconstruction that allows a flexible choice of temporal information, i.e., different reconstructions with user-defined parameters, such as temporal resolution and the number and position of temporal frames.
Like many iterative reconstruction techniques, the reconstruction time of GRASP is long and it is not practical to directly integrate into clinical MRI scanners. The development of the Yarra framework by Dr. Tobias Block was a milestone in the project. Yarra is a platform that can translate offline iterative reconstruction approaches into clinical MRI scanners in a smart way (see http://yarraframework.com). Since then, GRASP has been used in routine clinical MRI exams and has been evaluated for a wide range of clinical studies (see Feng L et al. JMRI 2017 Apr;45(4):966-987 for some examples).
I have led the extension of the GRASP framework to a number of improved variants. These include XD-GRASP (GRASP with eXtra Dimensions) to improve motion management, Koosh-ball-GRASP to combine GRASP with 3D golden-angle radial trajectories, variable-density stack-of-stars to improve the distribution of radial spokes in three dimensions, Unstreaking-GRASP to remove undersampling-induced residual streaking artifacts, RACER-GRASP (Respiratory-weighed and contrast enhancement-guide GRASP) to perform motion-weighted and dynamic contrast enhancement-guided reconstruction for capturing the most desired contrast phases, GROG-GRASP to improve reconstruction speed using the self-calibrating GRAPPA operator gridding (GROG) algorithm (see MRM 2008 Apr;59(4):930-5 for more details about GROG), and UTE-GRASP to combine GRASP with 3D golden-angle ultra-short TE radial sampling. These new developments have led to several novel applications, such as the 5D whole-heart sparse MRI, respiratory-resolved 4D whole-heart coronary MRA and 4D sparse lung MRI. In addition, stack-of-stars sampling has been extended to multi-echo acquisition (Dixon-RAVE: Dixon RAdial Volumetric Encoding) for free-breathing fat/water separation by my colleagues Dr. Tobias Block and Thomas Benkert. Dixon-RAVE has also been combined with the XD reconstruction strategy and has been tested for DCE-MRI.
During the past a couple of years, GRASP has been extended for further innovations. It has been extended to GRASP-Pro and XD-GRASP-Pro, which, as can be seen from the names, aim to boost the reconstruction performance of the original GRASP and XD-GRASP reconstruction. Recently, magnetization-prepared GRASP (MP-GRASP) and MP-Dixon-GRASP have been developed for rapid free-breathing 3D T1 mapping. We have also shown that MP-Dixon-GRASP enables fat-water separated T1 mapping to remove the influence of fat (and potentially iron as well).
The GRASP project represents a decade of innovation by our team consisting of imaging scientists, clinicians and our industry partners. The GRASP paper was announced as the third most-cited MRM paper at the 2017 ISMRM annual meeting, and the XD-GRASP paper was announced as the top most-cited MRM paper at the 2019 ISMRM annual meeting. With the rise of Artificial Intelligence in recent years, we are now aiming to integrate GRASP MRI with deep learning approaches to enable further improvement in reconstruction quality and speed, as well as new uses of this imaging framework. The initial feasibility of deep-learning-enabled golden-angle radial MRI has been demonstrated by us with a technique called SANTIS (see below), and we are in the process of developing new quantitative imaging methods based on a combination of GRASP with deep learning.
Deep Learning for Rapid MRI
I have been actively exploring the direction of deep learning for rapid MRI, particularly rapid quantitative MRI, together with my collaborator Dr. Fang Liu, who has pioneered the use of Artificial Intelligence for many aspects of MRI research. We recently proposed a sampling-augmented deep image reconstruction strategy for improved reconstruction performance and robustness against sampling discrepancy. This is achieved by extensively varying undersampling scheme during network training, so that the trained network can recognized various undersampling artifact patterns and thus enables better removal of undersampling artifact during the reconstruction process. It turns out that this framework is well-suited for golden-angle radial image reconstruction since golden-angle imaging features non-repeating sampling, which can be treated as a nature augmentation strategy. We also proposed a framework for model-based MR parameter mapping based on deep-learning, which enables efficient, highly-accelerated and accurate parameter mapping as demonstrated in our first feasibility study. This framework is currently under extension for rapid quantification of other MR parameters with the incorporation of corresponding signal models. We have recently written a paper together to review the application of deep learning for rapid quantitative MRI.
-
Liu F, Samsonov A, Chen L, Kijowski R, Feng L. “SANTIS: Sampling-Augmented Neural neTwork with Incoherent Structure for MR image reconstruction"
Magn Reson Med. 2019 Nov;82(5):1890-1904. doi: 10.1002/mrm.27827
-
Liu F, Feng L, Kijowski R. “MANTIS: Model-Augmented Neural neTwork with Incoherent k-space Sampling for efficient MR parameter mapping”
Magn Reson Med. 2019 Jul;82(1):174-188. doi: 10.1002/mrm.27707
-
Feng L, Ma D, Liu F. "Rapid MR Relaxometry using Deep Learning: An Overview of Current Techniques and Emerging Trends"
NMR Biomed. 2020 Oct 15;e4416. doi: 10.1002/nbm.4416. Online ahead of print
-
Liu F, Kijowski R, Feng L, Fakhri GE. "High-Performance Rapid MR Parameter Mapping Using Model-Based Deep Adversarial Learning"
Magn Reson Imaging. 2020 Sep 25;74:152-160