Antoine Klauser^{1}, Sebastien Courvoisier^{1}, Michel Kocher^{1}, Dimitri Van De Ville^{1,2}, and Francois Lazeyras^{1}

Proton FID-MRSI sequence was implemented in 3D to measure metabolite distributions over whole brain up to 5mm-isotropic resolution. MRSI dataset was reconstructed through a Low-Rank-TGV model enabling Compressed-sensing SENSE acceleration. Acceleration performance was assessed quantitatively on a-posteriori 6.6mm-isotropic dataset showing good agreement up to a factor 2. As proof of concept, metabolite images of a 5mm-isotropic dataset accelerated by a factor 2 were acquired and reconstructed.

**Introduction**

Proton
- Free induced decay - magnetic resonance spectroscopic imaging
(^{1}H-FID-MRSI)
method permits to detect metabolic profiles across the whole brain
slices in 2D acquisition at high resolution[1-3].
This state-of-the-art method is a promising candidate for the
realization of 3D isotropic metabolic profiles. The metabolite signal
acquired at ultra-short echo time is unaffected by
the
relaxation of strongly J-coupled metabolites. The technique is
therefore an excellent choice for unbiased quantification of the
metabolite distribution. TR and acquisition time can be reduced
considerably by using the Ernst flip angle. To allow coverage of the
whole brain parenchyma, no in-plane volume selection is included. We used Lipid-Suppression
by
spectral orthogonality[3-5]
to cope with
skull lipid leakage.
In
clinical setting (3T), metabolic signal-to-noise of high-resolution
(voxel~0.1ml) acquisition could be too low for reliable
quantification. The long acquisition time in 3D might be a strong
limitation too. However, using partial separability of the signal[6]
and TGV reconstruction[7],
MRSI dataset can be efficiently denoised. Also, following a random
k-space sampling combined with known coil-sensitivity profiles, the
acquisition can be accelerated with Compressed-sensing SENSE.
This
research aims to optimize both denoising and acceleration of
high-resolution 3D FID-MRSI covering the whole brain. The performance
of the technique is assessed on metabolite images and their
respective Cramer-Rao-lower-bound (CRLB) values.

**Method**

*
Sequence
*

A volume selective 3D FID-MRSI[1] sequence including WET water suppression was designed and implemented. The volume selective excitation pulse of 0.9ms was optimized with a Shinnar-Le Roux algorithm reaching a 9.5kHz bandwidth. Phase-encoding duration was shortened to reach a TE of 0.7ms and the FID signal was acquired with 1024 points at 4kHz sampling rate leading to a TR of 370ms. Considering a maximum metabolite T1 value of 1300ms, the Ernst angle for the excitation pulse was 40 degree. To determine the coils sensitivity profiles, water signal was acquired consecutively with the body coil and the head phased-array coil and with the same sequence but with a TR of 31ms, 48 FID points and a 10 degree flip angle.

*Experiment
*

Healthy
volunteers’ data were acquired at 3T
(Prisma/Siemens/Erlangen/Germany) with 64-channel receiver head coil.
Structural images were acquired with 3D-T1-weighted MPRAGE
sequence.
3D-FID-MRSI was planned on 210x210x105mm3
FOV and a 88mm thick slab excitation. Two encoding schemes were
experimented: 1.
full Low-Resolution scheme
consisting of 32x32x16 elliptical encoding
and 6.6x6.6x6.6mm^{3}
resolution requiring 39min acquisition time; 2. sparse
High-Resolution scheme
with a 40x40x20 k-space matrix
randomly
filled up
to 50% (acceleration factor 2) resulting in 5.3x5.3x5.3mm^{3}
resolution
that required 40min for acquisition

*Lipid
Suppression and Low-Rank TGV Reconstruction*

Based on a skull mask made on MPRAGE structural images, a projection onto the lipid temporal subspace, $$$P^L$$$ was defined and apply on acquired MRSI dataset, $$$S(\mathbf{k},t)$$$ to remove the lipid signal by spectral orthogonality [4,5]:$$S_{free}(\mathbf{k},t)=\int(\boldsymbol{1}-P^{L}(t,t'))S(\mathbf{k},t')dt'$$We apply a Low-Rank approximation[6] on the lipid-free signal in real-space:

$$S_{free}(\mathbf{r},t)=\sum_{c=1}^{N_C}U_c(\mathbf{r})V_c(t)$$The Low-Rank decomposition highlights the spatial-spectral features while efficiently denoising the MRSI dataset[6] through a TGV reconstruction model[7]: $$\mathbf{U}(\mathbf{r}),\mathbf{V}(t)=\arg\min_{\mathbf{U},\mathbf{V}}\:\:\left\|S_{free}( \mathbf{k},t)-\mathcal{FSB}\{\mathbf{U}(\mathbf{r})\mathbf{V}(t)\}\right\|^2_2+\lambda\:TGV^2(\mathbf{U}(\mathbf{r}))$$

where $$$\mathcal{B}$$$, is the B0 inhomogeneity operator, $$$\mathcal{S}$$$, the coil-sensitivity profiles and $$$\mathcal{F}$$$, the Fourier encoding. The presented model allows for Compressed-sensing SENSE acceleration[8,9] if the k-space is sparsely and randomly sampled.

*Quantification and
comparison
*

The reconstructed
MRSI dataset was quantified using LCModel[10]
.
The **full
Low-Resolution**
dataset was a-posteriori
undersampled at various acceleration factors (AF):
{1.5;2;2.5;3;3.5;4} for comparison using
CRLB
of tNAA, tCre, Cho, Ins and Glx, SNR, relative
root mean square error (RRMSE)
and voxelwise correlation (VC). Eventually, an example of metabolite
profiles of **sparse
High-Resolution** acquisition
are shown.

**Results**

On Fig.1 acceleration of

**Discussion**

[1] Henning, A., Fuchs, A., Murdoch, J.B. and Boesiger, P. (2009). Slice-selective FID acquisition, localized by outer volume suppression (FIDLOVS) for (1)H-MRSI of the human brain at 7 T with minimal signal loss. NMR Biomed 22, 683-696.

[2] Nassirpour, S., Chang, P., Henning, A. (2017). High and ultra-high resolution metabolite mapping of the human brain using 1H FID MRSI at 9.4T. NeuroImage

[3] Hangel, G., et al. (2016). Ultra-high resolution brain metabolite mapping at 7 T by short-TR Hadamard-encoded FID-MRSI. NeuroImage

[4] Bilgic, B., Gagoski, B., Kok, T. and Adalsteinsson, E. (2013) Magn Reson Med 69, 1501-1511

[5] Klauser, A., Van De Ville, D. ,Lazeyras, F. . Low-Rank TGV Reconstruction of High-Resolution 1H-FID-MRSI of Whole Brain Slices. (2017) ISMRM 25th Annual Meeting, preceeding #5517

[6] Nguyen, H.M., Peng, X., Do, M.N. and Liang, Z. (2013). Denoising MR spectroscopic imaging data with low-rank approximations. IEEE Trans Biomed Eng 60, 78-89.

[7] Kasten, J., Lazeyras, F., & Van De Ville, D. (2013). Data-Driven MRSI Spectral Localization Via Low-Rank Component Analysis. Medical Imaging, IEEE Transactions on, 32(10), 1853–1863.

[8] Liang, D., Liu, B., Wang, J., & Ying, L. (2009). Accelerating SENSE using compressed sensing. Magnetic Resonance in Medicine, 62(6), 1574–84. https://doi.org/10.1002/mrm.22161

[9] Otazo, R., Kim, D., Axel, L., & Sodickson, D. K. (2010). Combination of compressed sensing and parallel imaging for highly accelerated first-pass cardiac perfusion MRI. Magnetic Resonance in Medicine, 64(3), 767–776. https://doi.org/10.1002/mrm.22463

[10] Provencher, S.W. (1993). Estimation of metabolite concentrations from localized in vivo proton NMR spectra. Magn Reson Med 30, 672-679.

[11] Lecocq, A., Le Fur, Y., Maudsley, A. A., Le Troter, A., Sheriff, S., Sabati, M., … Ranjeva, J.-P. (2015). Whole-brain quantitative mapping of metabolites using short echo three-dimensional proton MRSI. Journal of Magnetic Resonance Imaging, 42(2), 280–289.

Fig. 1: 3D
Metabolite images of total creatine, choline and glutamate/glutamine
of the **full
Low-Resolution** acquisition.
Data were undersampled a-posteriori by a factor two (2x) and four
(4x) to simulate acceleration. 2x preserves most of the features
present in the full dataset (1x) while 4x shows strong distortions of
the images.

Fig. 2: Taking the **full
Low-Resolution**
as reference dataset,
the
relative root mean square error and the voxelwise correlation were
computed for each metabolite estimated by LCModel at several
acceleration factors.
In addition, the mean Cramer-Rao
lower
bound
of the brain for each metabolite and the signal-to-noise
ratio
are shown.

Fig. 3: Metabolite images
reconstructed from the **sparse
High-Resolution** 5mm
isotropic MRSI dataset that were accelerated by a factor 2 during
acquisition.