Fluorine-19 MRI has emerged as a promising tool for in vivo cell tracking, yet low achievable signal-to-noise ratios remain a major challenge. Compressed sensing offers increased sensitivity at the cost of introducing signal intensity bias. We show that at low signal levels the quantification performance of compressed sensing is similar to conventional methods due to signal intensity distribution induced bias effects, which also affect the Fourier reconstruction. To improve quantification results, we propose an intensity correction scheme based on ex vivo reference data.
All animal experiments were carried out in accordance with local animal welfare guidelines (LaGeSo). EAE was induced in SJL/J mice and perfluoro-15-crown-5-ether rich nanoparticles were administered daily starting on the 5th day after EAE induction4. In vivo data was acquired from three mice on day 13 and 14 after induction and ex vivo data was acquired in tissue phantoms prepared from three different animals. MR experiments were performed on a 9.4T animal scanner (Bruker BioSpin, Ettlingen, Germany). A 3D-RARE protocol was employed for fluorine-19 MRI: TR=800ms, TE=4.4ms, ETL=40, FOV=(45 16 16)mm³, (140 40 50) matrix, 25 repetitions (in vivo), 40 repetitions (ex vivo), 3 averages per repetition. The average of all repetitions was used as reference. A cylindrical cap filled with 2% agarose and 20mM nanoparticles was included for quantification.
For five different measurement times, fully-sampled and 2 to 5-fold undersampled data were retrospectively sampled (fig.1A). CS and denoised reconstructions were computed using the accelerated alternating direction method of multipliers5 with isotropic total variation and image l1-norm regularization. The deviation of the reconstruction from the measured data was set 97% of the noise level by adjusting the regularization strength. 5 different datasets were generated and reconstructed for each measurement time and method. The Rician noise bias in the conventional Fourier reconstructed magnitude images was corrected as described by Henkelman6. Reconstructions were thresholded at 3.5 (Fourier) or 2 times (denoising and CS) the k-space data noise level. Groups of less than three connected voxels were removed as outliers (fig1B).
As signal bias depends on the noise level, it was computed for different levels of the measured signal scaled by the noise standard deviation of the Fourier reconstruction at equal scan time σF. Bias correction was performed based on a polynomial fit of the ex vivo data. Simulations testing the interpretation of the observed bias effects were performed as described in fig.2. Reconstructions, simulations and analyses were programmed in MATLAB 2017a (The MathWorks, USA).
1. Flögel, U., & Ahrens, E. (2016). Fluorine Magnetic Resonance Imaging. Pan Stanford
2. Starke, L., Waiczies, S., Niendorf, T., Pohlmann, A., (2018). Compressed Sensing Improves Detection of Fluorine-19 Nanoparticles in a Mouse Model of Neuroinflammation, presented at the annual meeting of the ISMRM, Paris, France
3. Hu, S., et al. (2010). 3D compressed sensing for highly accelerated hyperpolarized 13C MRSI with in vivo applications to transgenic mouse models of cancer. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, 63(2), 312-321.
4. Waiczies, S., Millward, J. M., Starke, L., Delgado, P. R., et al. (2017). Enhanced fluorine-19 MRI sensitivity using a cryogenic radiofrequency probe: technical developments and ex vivo demonstration in a mouse model of neuroinflammation. Scientific Reports, 7(1), 9808.
5. Goldstein, T., O'Donoghue, B., Setzer, S., Baraniuk, R. (2014). Fast alternating direction optimization methods. SIAM Journal on Imaging Sciences, 7(3), 1588-1623.
6. Henkelman, R. M. (1985). Measurement of signal intensities in the presence of noise in MR images. Medical physics, 12(2), 232-233.