Introduction

Quantitative MRI provides valuable information regarding cardiovascular pathologies as myocardial T₁ changes may reflect infarction, diffuse fibrosis or inflammation¹. The clinical approach consists of acquiring multiple T₁-weighted images with a Cartesian MOLLI (MOdified Look-Locker) sequence and fitting a 3-parameter model: S(TI)=A–B.exp(-TI/T₁*); where T₁~T₁^*(B/A-1). However, other factors that affect the magnetization state and lead to T₁ bias are not explicitly taken into account (e.g. effective flip angle, heart rate variability and others). Another issue of conventional cartesian MOLLI is the lengthy breath-hold required. To address this, we propose coupling an accelerated golden-angle (GA) radial MOLLI with a model-based regularized reconstruction (SALSA).

Methods

An in-house open-source radial GA (with alternating spoke direction) trajectory with a MOLLI 5(3)3 scheme was considered (flip angle = 35º, 2 inversions with inversion time (TI) = 104 and 186ms, TR=3.03ms, FOV=200mm² and trigger delay=400ms). The pulse sequence was implemented with Pulseq² and data acquired on a Siemens Aera 1.5T scanner. Preliminary testing was performed on the ISMRM/NIST T₁/T₂ phantom and in vivo on skeletal leg muscle with a simulated ECG (50 beats per minute). Radial GA data was acquired at different acceleration rates from fully sampled (R =1) to undersampled (R = 4). T₁-weighted radial images were then reconstructed offline with NUFFT³ and SALSA⁴. Zeroth-order phase corrections were applied to k-space before reconstruction; i.e. the phase of k-space centre in each radial spoke was calculated and subtracted to all other spoke readout points. An Extended Phase Graph (EPG) dictionary was computed considering the acquisition parameters and a T₁ range of [5:5:2400]ms and T₂ of [10:10:1000]ms. MR image reconstruction can be described as y=Bx+n, where y is the acquired noisy k-space data, x the reconstructed image and B is the direct operator (here including the Fourier transform). SALSA solves the constrained optimization problem min 1/2 ||y-M(EXF)||+||d X W|| where M is the mask which defines the undersampling of the acquired k-space y, E is the compressed dictionary computed with EPG, F is the Fourier operator (which in the non-Cartesian case corresponds to NUFFT), W is the 2D discrete wavelet operator of the rows of X (matrix of coefficients combining the compressed dictionary components) and d is a diagonal matrix controlling the weight of W. This can then be solved by using an alternating direction method (ADMM). T₁ maps were obtained by 3-parameter fit to the accelerated T₁-w images, and by matching the NUFFT and SALSA T₁w-images to the dictionary. To serve as a comparison, T₁ maps were also acquired with the commercial Cartesian MOLLI sequence (TI = 231/331ms) for both phantom and in vivo. A schematic representation of the methodology is seen in Fig 1.

Results

Reduction of aliasing artefacts (accelerated radial R = 4) can be seen in T₁-w images reconstructed with SALSA (for phantom studies, λ=1e-5 was chosen) compared to standard NUFFT (Fig 2). Mean T₁ values measured in each phantom vial were compared to the T₁ ground truth from the NIST phantom manual (Fig 2 j - dark blue bar) and the commercial Cartesian MOLLI sequence (light blue bar). Accelerated radial (allowing to reduce breath-hold time compared to Cartesian MOLLI) demonstrated comparable T₁ measurements to the commercial sequence for vials 2-4 (corresponding to the range of more relevant T₁ = 1489 to 780ms). However, an underestimation was observed for the dictionary methods (NUFFT and SALSA) for vial 1 (T₁=2033 ms). SALSA also resulted in less T₁ spatial dispersion within each vial (lower variation coefficient - Fig 2 k) compared to NUFFT. T₁ maps were also computed for leg muscle data sets (Fig 3), where a λ= 3e-6 was chosen. Similarly to phantom results, reduction of the aliasing artefacts can be seen in the accelerated radial SALSA T₁-w images (Fig 3 f) compared to NUFFT. As expected, the SALSA method leads to improved precision compared to the NUFFT T₁ maps (lower variation coefficient - Fig 3 k). However, mean T₁ values of the SALSA maps appear to be slightly lower compared to the commercial MOLLI sequence.

Discussion and conclusions

The proposed radial GA MOLLI with SALSA-EPG appears to be a promising method to obtain fast open-source T₁ mapping. Accelerations up to 4-fold were tested in the NIST phantom and leg muscle, demonstrating the potential of the proposed method for reducing the overall breath-hold duration of standard MOLLI. Good T₁ accuracy and precision was obtained with the SALSA method compared to the ground truth and the commercial sequence for the T₁ range most relevant for a later cardiac application (1489-780ms). However, estimated T₁ values in vivo were lower with the proposed method than the commercial MOLLI. Potential reasons for this could be the choice of TI used in the radial sequences (shorter than the commercial), and differences between off-line and in-line estimation procedures (the MOLLI Cartesian map was obtained using Siemens reconstruction and estimation pipeline). Future work will attempt to improve the accuracy of the SALSA-EPG by optimizing the chosen TI, including coil sensitivity information in the SALSA optimization and correcting for first-order gradient-induced phase errors.

Acknowledgements

Fundação para a Ciência e a Tecnologia (SFRH/BD/120006/2016, PTDC/EMD-EMD/29686/2017) e Programa Operacional Regional de Lisboa 2020 (LISBOA-01-0145-FEDER-029686)

References

[1]Taylor, JACC: Card Img(2016)9:67-81 [2] Layton, MRM (2017) 77:1544-1552 [3]Fessler, JMR(2007)188:191-195 [4]Afonso, IEEE Trans Img Process(2010)19:2345-56