T1ρ quantification of undersampled MRI data via low-rank reconstruction: towards multiparametric myocardial fingerprinting in small animals at 7T

Background
Compared to human cMRI, quantitative assessment in small animals is challenging for several reasons. Firstly, the ultra-high field strengths of preclinical scanners pose high demands on calibration steps and artifact correction. Furthermore, the acceleration of MR sequences is hampered as parallel imaging is limited due to the reduced multi-channel-coil options and advanced readouts such as spiral are hardly feasible¹. Therefore, it is necessary to optimize reconstruction pipelines of undersampled MRI data for maximum efficiency.
In this study, we transferred a novel low-rank-reconstruction technique developed for human MR fingerprinting² to highly undersampled MR data obtained from cMRI experiments in mice. The aim of this work was to realize myocardial T_1ρ quantification with 8-fold acceleration.

Methods
The measurements were performed on a 7T preclinical MRI system (Bruker PharmaScan) with a single-channel quadrature transmit-receive birdcage coil. Data of an 8-fold accelerated T_1ρ mapping experiment with radial sampling pattern were analyzed, including one healthy animal and one with myocardial infarction. The results of undersampled reconstruction (UR) were compared with results of the low-rank reconstruction (LRR). The algorithm of LRR iterates a Dai-Yuan conjugate gradient descent² and finally leads to a model-based solution, which implies the exponential T_1ρ decay as a physical constraint. The thus quantified T_1ρ and R² values were compared with UR in the left ventricle.
 
Results
Fig. 1 depicts the direct comparison of UR and LRR for exemplary T_1ρ weighted images. A significant improvement in image quality (signal-to-noise-ratio: LRR 21.2, UR 8.4) is evident and LRR is nearly free of undersampling artefacts. Fig. 2 shows the T_1ρ map of the healthy animal and its evaluation in the left ventricle. The mean T_1ρ deviation compared to the values obtained with UR is 1.8±3.3ms. However, significantly increased R² values were obtained for LR (0.994±0.005 vs. 0.976±0.010). The results of the MI study (Fig. 3) demonstrate that LRR can clearly differentiate infarcted tissue from remote tissue, whereby increased T_1ρ values were observed.
 
Discussion
In this work, 8-fold accelerated quantification of the T_1ρ relaxation time was successfully achieved using LRR. Compared to UR, a diagnostic image quality was generated without the need for k-space view-sharing approaches³, which are known for both miss-quantification and edge blurring/sharpening effects. The key to this efficient method is that low-rank implies a physical model for the reconstruction process. Our presented work is the first step to implement highly efficient multiparametric acquisitions for preclinical cMRI. In future studies, we plan to use LRR for fingerprinting sequences to achieve simultaneous and co-registered T₁, T₂ and T_1ρ maps in short acquisition times. Another important step will be the upgrade from radial to spiral acquisitions to achieve further efficiency gains⁴.

[1] Gu Y, et al. doi: 10.1002/mrm.27345
[2] Hamilton JI, et al. doi: 10.1002/nbm.4041
[3] Gram M, et al. doi: 10.1007/s10334-021-00951-y
[4] Scholten H, et al. doi: 10.1002/nbm.5249

Acknowledgements
The authors would like to thank Tom Griesler and Jesse Hamilton (Seiberlich Lab, University of Michigan) for the initial support with the LRR, which was adapted from human cMRI studies². We would also like to thank the Ramos Lab (University Hospital Würzburg) for the preparation of the MI model.