0135

Accelerated Magnetic Resonance Fingerprinting using Soft-weighted key-Hole (MRF-SOHO)

Gastao Cruz¹, Andreia S. Gaspar¹, Tom Bruijnen², René Botnar¹, and Claudia Prieto¹

¹Division of Imaging Sciences and Biomedical Engineering, King's College London, London, United Kingdom, ²Center for image sciences, University Medical Center Utrecht, Utrecht, Netherlands

Synopsis

Magnetic Resonance Fingerprinting estimates multi-parametric maps from a series of highly undersampled time-point images. However, MRF scan times are still long due to the large amount of time-point images (~1000) required to produce accurate multi-parametric maps. Here we propose to exploit redundant information in time-point images with similar contrast to accelerate the MRF scan by further undersampling each time-point image and/or significantly reducing the number of required images in the series. The proposed approach achieved an acceleration factor of 5.7× compared to conventional undersampled MRF while maintaining parametric map quality.

Introduction

Magnetic Resonance Fingerprinting (MRF)¹ relies on temporal incoherence of the signal evolution of different tissues to enable pixelwise matching of a measured signal to a dictionary of simulated signals. The parametric space is typically explored by varying the repetition time (TR) and flip angle (FA) between time-point images to generate unique signal evolutions. Traditionally used slowly varying FA produces a smooth change in contrast, resulting in redundant high frequency information in the series of time-point images. To further accelerate the MRF scan, this information is exploited using a SOft-weighted key-HOle (MRF-SOHO) reconstruction that enables data sharing of high frequency information between different time-point images. The proposed approach was validated on a standardized T₁/T₂ phantom and in-vivo brain acquisitions.

Methods

MRF-SOHO combines parallel imaging², soft-gating³ and key-hole⁴ methods to accelerate the MRF scan. In MRF-SOHO, each time-point image is reconstructed using high frequency data from its temporal neighbourhood by solving I_t = arg min ||W_tⁿ(FSI_t- K_t)||². F is the Fourier transform, S are the coil sensitivities and K_t are the acquired k-space data for time-point t. The soft-weights W_tⁿ for time-point t and neighbour n ∈ {t-Δt,...,t+Δt} are given by W_tⁿ(k_r) = 1 - {exp[(k_r- α(n))/(β(n)]+1}^-1 for n≠t and W_tⁿ(kr)=1 for n=t, where k_r is the distance from the k-space center, and α(n) and β(n) control the amount of shared data by neighbour n. MRF-SOHO shares high frequency information, reconstructing alias free time-point images without compromising image contrast (Figure 1). The MRF dictionary was built with an Extended Phase Graph method based on⁵ for T₁ (ranges [0, 6000] and [0, 1600]) and T₂ (ranges [0, 2600] and [0, 260]) values (for brain and phantom, respectively). Template matching was performed as described in¹.

Experiments

A standardized T₁/T₂ phantom⁶ was scanned in a 1.5T Philips system using an 18-channel coil. Data was obtained with a rewound gradient echo acquisition after an initial inversion pulse using a tiny golden radial (~23.6º) trajectory⁷. Relevant scan parameters include: resolution 2×2 mm², 10 mm slice thickness, field-of-view 320×320 mm², 8 radial spokes per time-point, 1000 time points and TR (varying between 6.0 and 7.4 ms) and FA (varying between 0º and 70º) patterns similar to the ones described in⁸. One healthy subject underwent a brain scan in the same system using a 12-channel coil and the same protocol as above, (except field-of-view=440×440 mm²). In both acquisitions, the neighborhood size n for the MRF-SOHO reconstruction was set to 17, α(n) varied linearly between 0 and N_FE/6, and β(n) varied linearly between N_FE/60 and N_FE/40, where N_FE is the number of frequency encodings. Both datasets were reconstructed with a zero fill gridding reconstruction as in conventional MRF¹ and the proposed MRF-SOHO using 8, 4 and 2 radial spokes (corresponding to angular undersampling factors of ~31×, ~63×, ~126× and ~43×, ~86×, ~173× for the phantom and brain datasets, respectively). Template matching of these reconstructions was performed using different number of time-point images, from 50 to 1000 (as in¹) time points.

Results

Plots of the measured T₁ and T₂ values compared with ground-truth values in the phantom are presented in Figure 2. Conventional MRF consistently underestimated T₁ and T₂; the proposed approach achieved higher accuracy and precision in all cases. T₁ and T₂ in-vivo brain maps for 4 radial spokes and 1000, 500, 350 and 100 time-points are presented in Figure 3 for conventional MRF and MRF-SOHO. MRF-SOHO is robust to all cases, where conventional MRF fails at reduced time points. T₁ and T₂ in-vivo brain maps for 500 time points and 8, 4 and 2 radial spokes are shown in Figure 4 for MRF and MRF-SOHO approaches. Aliasing artefacts are observed for conventional MRF with 4 and 2 radial spokes, whereas a significant reduction of these artefacts is observed with the proposed MRF-SOHO approach. The optimal MRF-SOHO accelerated dataset was experimentally determined to be ~350 time-points, 4 spokes. Parametric maps for the standard MRF (1000 time-points, 8 spokes) and the MRF-SOHO (350 time-points, 4 spokes; ~5.7× accelerated relative to standard MRF) are shown in Figure 5, alongside T₁ and T₂ measurements for white matter, grey matter and cerebrospinal fluid. All estimated values were in agreement with literature, despite a slight underestimation of T₁ values in conventional MRF.

Conclusion

A novel approach for highly accelerated MRF was proposed, based on data sharing of high frequency k-space. The proposed MRF-SOHO achieved acceleration factors of ~5.7× relative to conventional undersampled MRF while maintaining accurate T₁ and T₂ map estimations. Future work will further study the acceleration limit of MRF-SOHO with additional in-vivo experiments.

Acknowledgements

This work was supported by EPSRC EP/P001009/1 and FONDECYT 1161055.

References

1. Ma D et al. Magnetic Resonance Fingerprinting. Nature. 2013; 495:187-192

2. Pruessman et al. Advances in sensitivity encoding with arbitrary k-space trajectories. MRM. 2001; 46:638-651

3. Johnson et al. Improved Least Squares MR Image Reconstruction Using Estimates of k-Space Data Consistency. MRM. 2012; 67:1600-1608

4. Song et al. Non-Contrast Enhanced 4-D Dynamic MRA with Golden Angle Radial Acquisition and K-space Weighted Image Contrast (KWIC) Reconstruction. MRM. 2004; 52:815-824

5. Weigel. Extended phase graphs: Dephasing, RF pulses, and echoes - pure and simple. JMRI. 2015; 41:266-295

6. Captur et al. A T1 and ECV phantom for global T1 mmaping quality assurance: the T1 mapping and ECV standardisation in CMR (T1MES) program. J Cardiov Magn Reson. 2016; 18:1-3

7. Wundrak et al. Golden ratio sparse MRI using tiny golden angles. MRM. 2016; 75: 2372–2378.

8. Jiang et al. MR Fingerprinting Using Fast Imaging with Steady State Precession (FISP) with Spiral Readout. MRM. 2015; 74:1621-1631

Figures

Figure 1: a) Gridding reconstruction used in conventional MRF at different time points t in the acquisition (8 spokes per time-point). b) MRF-SOHO reconstruction at different time points t in the acquisition (8 spokes per time-point). Diagnostic quality images are reconstructed at each time point with MRF-SOHO without affecting the contrast of the time-point images. c) Flip angle pattern used presents smooth evolution. Coloured time-points correspond to the images in a) and b). d) Soft-weights (W_tⁿ) for different neighbouring time-points (n) as a function of k-space radius (k_r). The further away the time-point, the less low frequency data is shared.

Figure 2: T₁ and T₂ measurement plots (in ms) with conventional MRF and MRF-SOHO reconstructions, using 8 spokes, 1000 time points; 4 spokes, 500 time points; and 2 spokes, 500 time points. An estimation bias of high T₁ values was observed with the conventional MRF reconstruction in comparison with the ground-truth phantom values. MRF-SOHO correctly estimated T₁, even at increased undersampling factors and reduced number of time points (4 spokes, 500 time points and 2 spokes, 500 time points). MRF-SOHO produced more accurate and precise measurements than conventional MRF, particularly when less data (time points and/or radial spokes) was used.

Figure 3: T₁ and T₂ maps (in ms) for conventional MRF reconstruction and the proposed MRF-SOHO reconstruction using 4 radial spokes and different number of time points (1000, 500, 350 and 100). The conventional MRF reconstruction produces good quality parametric maps with 1000 and 500 time points, but starts failing with reduced number of time-point images. The proposed MRF-SOHO reconstruction produces good quality and accurate parametric maps in all cases, however the signal-to-noise ratio is reduced in cases of high acceleration.

Figure 4: T₁ and T₂ maps (in ms) for conventional MRF reconstruction and the proposed MRF-SOHO reconstruction using 500 time points and different number of radial spokes (8, 4 and 2). The conventional MRF reconstruction produces good quality parametric maps with 8 radial spokes, but quickly deteriorates with increased undersampling factor per time-point image. The proposed MRF-SOHO reconstruction produces good quality and accurate parametric maps in all cases, however the signal-to-noise ratio is reduced in cases of high acceleration.

Figure 5: T₁ (top) and T₂ (bottom) maps (in ms) for the conventional MRF (left) reconstruction using 1000 time points, 8 spokes and the proposed MRF-SOHO (right) using 350 time points, 4 spokes (~5.7× accelerated relative to conventional MRF). Literature T₁ and T₂ values are show in parenthesis. The conventional MRF reconstruction produces overall accurate parametric maps, however a slight underestimation of T₁ for white and grey matter is observed. The proposed MRF-SOHO achieved accurate estimation of T₁ and T₂ for all tissues despite the high acceleration factor, however a slight overall loss in signal-to-noise ratio was also observed.

Proc. Intl. Soc. Mag. Reson. Med. 25 (2017)

0135