Arthur Coste1, Nicolas Chauffert1,2, Alexandre Vignaud1, Philippe Ciuciu1,2, Fawzi Boumezbeur1, Pierre Weiss3, Sandro Romanzetti4, Denis Le Bihan1, and Cécile Lerman1
1MR Imaging and Spectroscopy Unit, NeuroSpin, Gif sur Yvette, France, 2Parietal, INRIA Saclay, Saclay, France, 3Institut des Technologies Avancées du Vivant, Toulouse, France, 4University Clinic RWTH Aachen, Aachen, Germany
Synopsis
This work illustrates the application of a wavelet based least square
regularized image reconstruction for non-Cartesian MR sampling trajectories in
the framework of low concentration Phosphorus Imaging.
We compared performances between 3D Cartesian Fast Imaging with Steady-state free
Precession (FISP) and the TPI (FISP)
sequence and show that in equivalent acquisition durations we are able to
produce better images with the TPI sequence. Prospective sub-sampling was also
performed and results open exciting possibilities to reduce acquisition time
without impacting image quality for non-proton MRI.Introduction
With the development of high field MRI, the in vivo study of non-proton nuclei becomes manageable in a clinical
research setting. These nuclei are particularly interesting to investigate
metabolism and cellular homeostasis in complement to
1H. However, some
of them presents quadrupolar moments leading to very short T
2*
decays creating additional constraints for the acquisition.
A solution is to resort to Compressed sensing which consists of
combining efficient k-space sampling trajectories in particular undersampled
Non-Cartesian sampling schemes, with dedicated reconstruction algorithm.
Purpose
This work demonstrates
in vitro
benefits of using the Twisted Projection Imaging sequence
[1] (TPI) FISP combined with a nonlinear reconstruction method.
Material and Methods
Acquisitions were performed on a 7T Magnetom MRI scanner (Siemens,
Erlangen, Germany) and using double-tuned (1H/31P) head
only phased array transceiver coil. (Resonance Research Inc, Billerica, USA)[2].
A home-made 2L spherical phantom, containing a Phosphate Buffered
Saline (PBS) solution at 40mmol/L was used for testing. Two galleries of 50mL tubes (containing
either pure water or PBS in various concentrations) were disposed around the
sphere as illustrated in Figure 1.
A reference image using a Cartesian 3D Fast Imaging with Steady-state free Precession (FISP) sequence was compared to acquisitions
performed using the TPI sequence. Acquisition parameters were the same for both
sequences: TR/TE = 100/4.5ms, FA = 10°, isotropic FOV of 320mm at 5mm isotropic
resolution, 1 average, dwell time was 20µs. TPI images were acquired with a
linear portion (p) of 0.75 and a number of spokes equivalent to the Cartesian
acquisition (4096).
Our goal was to compare reconstructed images acquired using
the two MRI sequences for identical scanning time.
Cartesian FISP image
was directly reconstructed on-line by Inverse Fourier Transform. TPI images
were reconstructed using a 3D wavelet-regularized least square reconstruction
using the Fast Iterative Shrinkage Thresholding Algorithm (FISTA)[3]
implemented in Matlab (The Mathworks, Natick, USA). The FISTA algorithm computes
the global minimizer of the following convex but non-smooth criterion:
$$argmin\left(\frac{1}{2} ||Ax-y||_{2}^{2} + \lambda||x||_1 \right), \qquad \text{with } A=F\Psi$$
We used an orthogonal wavelet basis ψ and the Nonequispaced Fourier Transform[4].
Neither the regularization parameter λ nor the number of iterations are
yet automatically determined and are set empirically.
Under-sampled TPI acquisitions were then explored by reducing the number
of acquired spokes and further compared to the equivalent Cartesian acquisition.
Results and Discussion
As illustrated in
Figure 2,
for equivalent acquisition parameters, the TPI image (
b), reconstructed with FISTA (75 iterations and λ = 10
-5),
exhibits a better SNR and enables a better delineation of the object structure
and compartments. This result is maintained for a 50% under-sampling factor offering
the potential to shorten acquisitions.
SNR comparisons are presented in
Figure
3 looking at a ROI centered on the 80mmol/L tube. This graph shows that
higher SNR is reachable in lower acquisition time with TPI and FISTA. Similar curves
were extracted from all visible tubes and lead to the same conclusion.
Reducing the number of iterations in FISTA leads to a smoother image
with a better SNR at the cost of a wider Point Spread Function (PSF). A higher
SNR can also be obtained by increasing the regularization parameter. However,
one must be cautious as this regularization can also create artifacts.
Further work will be dedicated to an optimal and unsupervised setting of
parameters to get the best reconstructed image for any given data set.
Conclusion
In this study, we have shown that the combination of non-Cartesian
sampling sequence with a non-linear reconstruction algorithm improves the
quality of X-nuclei MR images. The combination of 3D TPI and FISTA produces promising
results in a reasonable amount of time despite the low available signal. Such
acquisition protocol opens up exciting possibilities to apply X-nuclei MRI for
clinical research. As shown elsewhere, efficient X-nuclei MRI approach shall
help defining precise quantification pipeline lasting about an hour
[5].
Acknowledgements
No acknowledgement found.References
[1] F. E. Boada et
al., MRM 37, 706-715, 1997
[2] N. Avdievich, Appl Magn Reson. 2011 Dec; 41(2-4): 483–506
[3] A. Beck et al., FISTA, SIAM 2009
[4] J. Keiner et al., ACM Trans. Math.
Software, 36, 2009
[5] A. Coste et al.,
submitted ISMRM, 2016