Introduction

Magnetic resonance fingerprinting (MRF) is a framework that allows for simultaneous quantification of tissue properties.¹ However, there are no consistent implementations of MRF to evaluate accuracy and repeatability across different sites and vendors. Recently, Pulseq was established as an open-source tool that enables fast prototyping of pulse sequence in MATLAB and Python.^2,3 In this work, we leverage Pulseq to develop an end to end open source MRF package named OMEGA in MATLAB and validate its functionalities.

Method

The gold standard (GS) measurements of T₁ and T₂ in corresponding arrays of the ISMRM/NIST phantom were performed on a Siemens 3T Prisma scanner (Siemens Healthineers, Erlangen, Germany) using the same protocols in previous study.⁴ The last four spheres of T₂ array were fitted separately using TE=12, 15, 18, and 21 ms to better characterize their values.
OMEGA contains five modules. The sequence design module used repetition time (TR), echo time (TE), and flip angles (FA) reported in literature⁴ and generated a .seq file that was played on the scanner. The sequence was implemented using Pulseq² with the following parameters: maximum gradient amplitude 32 mT/m, maximum slew rate 130 T/m/s, field of view 225 mm, and matrix size 256x256, with a total of 1000 time points. An inversion pulse was added at the beginning to enhance contrast and improve dictionary matching performance.
In vitro data was acquired using the generated sequence on ISMRM/NIST phantom at the same scanner. The scans were performed using a 20-channel head coil. The collected data from each coil was inputted into the image reconstruction module and sliding window reconstructed using Michigan Image Reconstruction Toolbox (MIRT) for Non-Uniform Fast Fourier reconstruction of spiral k-space data. The reconstructed data of each channel was corrected for phase and combined to form the final image.
The dictionary simulation module generated a dictionary using Extended Phase Graph (EPG)⁵ simulations with the same parameters as in the sequence. The dictionary contains 2627 entries covering T₁ and T₂ of all spheres in ISMRM/NIST phantom. The reconstructed image data was matched with EPG-simulated dictionary using vector dot product in dictionary matching module to generate T₁ and T₂ maps. Inversion time (TI) was calibrated to optimize dictionary matching performance.
The automated ROI analysis module was designed to extract mean and standard deviation (SD) of all spheres in the phantom. The tool assumed that: 1) the relative distance from each sphere to Sphere 1 are constants, 2) the experiment estimations are +/-15% of NIST values. The tool returns an error if the difference is greater than 15%, and 3) the phantom is at the isocenter of the magnet. 100 pixels per sphere were selected for OMEGA data and 50 pixels were selected for GS data due to different matrix sizes.
Previous studies suggested that T₂ values are more susceptible to B₁ variations than T₁.⁸ A Bloch-Siegert based B₁ mapping was performed to investigate the issue.

Results/Discussion

The GS measurements of T₁ and T₂ values are presented in Figure 1. All values are within 10% of NIST values.⁶ The extra four TEs helped improve the curve fitting performance with small T₂ values because longer TEs mostly contribute noise and compromise the curve fitting performance.
The TR and FA design of OMEGA are described in Figure 2. The design is identical to previous study.⁴ With the latest version of Pulseq interpreter, we were able to implement 1000 different radiofrequency (RF) pulses waveforms that correspond to the literature sequence.
Figure 3a shows the 500th time point of MRF acquisition and 3b compares the signal evolution of one pixel with its corresponding dictionary entry. The MRF images were relatively artifact free. The signal evolution and dictionary entry showed divergence in the first 200 points due to an inaccurate inversion pulse. A series of pulse calibrations were performed to fix this issue.
The T₁ and T₂ maps generated using OMEGA and boxplot with min and max are shown in Figure 4. Spheres 6-13 are within 15% of GS values. The first five T₁ values underestimate GS values by approximately 20%. This agrees with our previous study⁷ and can be corrected using a constant bias term. Sphere 14 was matched incorrectly because of limited dictionary resolution in short T₁ range. Deep learning method is a potential solution to this problem. T₂ values have a larger SD comparing to T₁. Figure 5 presents the B₁ map acquired on T₂ array with a flip angle of 8 degrees. The last two spheres with shortest T₂ values experience larger B₁ variations compared to other spheres. Sphere 1-8 received larger flip angles. B₁ map can be incorporated in EPG dictionary simulation and improve the dictionary matching performance.
The complete OMEGA package is open source and accessible on Github.⁹ This enables consistent implementations across multiple vendor platforms for meaningful comparisons of MRF sequences.

Conclusion

In this study, we demonstrated OMEGA, the first end to end, open source, vendor neutral MRF package using Pulseq and MATLAB reconstruction. Our results suggested that OMEGA is accurate for 17 spheres of T₁ and T₂ arrays in ISMRM/NIST phantom. Future work includes incorporating B₁ map to EPG dictionary simulation, deep learning based image reconstruction, and diffusion imaging.

Acknowledgements

This study was funded [in part] by the Seed Grant Program for MR Studies and the Technical Development Grant Program for MR Studies of the Zuckerman Mind Brain Behavior Institute at Columbia University and Columbia MR Research Center site.