A comparison of lipid suppression by double inversion recovery, L1- and L2-regularisation for high resolution MRSI in the brain at 7 T
Gilbert Hangel1, Bernhard Strasser1, Michal Považan1, Martin Gajdošík1, Stephan Gruber1, Marek Chmelík1, Siegfried Trattnig1,2, and Wolfgang Bogner1

1MRCE, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna, Austria, 2Christian Doppler Laboratory for Clinical Molecular MR Imaging, Vienna, Austria

Synopsis

Reliable lipid suppression is essential for robust quantification of parallel imaging accelerated high- resolution MRSI. This work compared the performance of non-selective lipid suppression using double inversion recovery (DIR) with the application of L1- and L2-regularisation during data processing for single-slice MRSI with a 64×64 matrix and a GRAPPA-acceleration of nine in five volunteers. While DIR featured the best lipid suppression, it increased the measurement time and reduced metabolite SNR. L1 and L2 did not have these downsides, but twice as much lipid signal remained, with L1 increasing the data pre-processing time before spectral quantification by a factor of six.

Purpose

High-resolution MRSI in the brain at 7 T1,2 allows the non-invasive mapping of spatial metabolite distribution in detail, but suffers from long measurement times and contamination by macromolecules and trans-cranial lipids originating from a non-optimal point-spread function or B0-inhomgeneities. Parallel imaging3,4 solves the first problem but worsens the second one due to lipid fold-in. Different concepts for lipid suppression at 7 T were introduced so far, such as outer volume suppression2 or non-selective double inversion recovery5 (DIR) during the measurement itself. Another approach is the general removal of lipid signal projected into the brain from the trans-cranial region during post-processing like L1- and L2-regularisation6,7. This is based on the magnitudes stronger total lipid signal and the orthogonality of lipid and metabolite spectra. Using parallel imaging accelerated MRSI data5, we compared the performance of DIR, L1 and L2 in order to facilitate the right choice of methods for specific MRSI needs.

Methods

Five volunteers were previously measured5 with a Siemens 7 T Magnetom scanner and a 32-channel coil using an FID-MRSI sequence with phase encoding and elliptical weighting. The sequence with no lipid suppression (NLS) had a TR of 1038 ms, while the DIR sequence (Figure 1) had a total TR of 1300 ms which included a TI1/TI2 of 210/52 ms. Common parameters were an acquisition delay of 1.3 ms, a 64×64×1 matrix, an FOV of 220×200×10 mm³, 2048 sampling points with 6000 Hz receive bandwidth and a 3×3-GRAPPA-acceleration with an effective R of 8.3 (6:17 min for NLS, 6:51 for DIR). An anatomical MP2RAGE reference scan (4:39 min) was acquired. We used an in-house developed pipeline8 for data processing that included LCModel fitting. For the lipid suppression comparison, L1 with 5 and 10 iterations as well as L2 were applied to both NLS and DIR datasets. The performance of the methods was compared using the lipid signal remaining after suppression (signal sum of the 1.2 ppm lipid resonance after baseline subtraction), processing times for L1/L2 as well as NAA SNR, CRLB and FWHM. Further, individual spectra, total lipid maps and NAA maps were evaluated.

Results

In general, the application of L1 and L2 to the DIR datasets did not lead to reliable results, excluding these combinations from the comparison. Only negligible differences were found between the results of 5 and 10 iterations of L1. Figure 2 provides an overview of the performance: Considering lipid suppression, DIR was the most effective with around half as much remaining lipid signal as L1/L2, but lost more of the SNR due to the double inversion. The apparent SNR reduction of L1/L2 (and partially for DIR) was to some extent caused by the removal of contamination in the NAA region that would have been otherwise wrongly attributed to the NAA signal. The processing time before the LCModel fitting increased by a factor of 6 for L1, but did not significantly change for L2. Exemplary for these results are the spectra of Figure 3. Comparing the lipid maps (Figure 4) and NAA maps (Figure 5) of all methods shows that the lipid suppression performance of all is adequate. The higher SNR of L1 and L2 translates into a better metabolite map quality.

Discussion/Conclusions

Overall, DIR lipid suppression as well as L1 and L2 regularisation allow sufficient removal of lipid artefacts, minimising their impact on metabolite quantification. While DIR has the best suppression efficiency, it is affected by longer measurement times due to SAR limitations and metabolite SNR loss. L1’s and L2’s performance is similar, but L2 does not effectively increase processing times. If the absolute lipid suppression efficiency is less important than minimising measurement times, L2 regularisation appears to be the most attractive choice in lipid suppression for brain MRSI at 7 T.

Acknowledgements

This study was supported by the Austrian Science Fund (FWF): KLI-61 and the FFG Bridge Early Stage Grant #846505.

References

[1] Bogner et al., NMR Biomed 2012; 25(6):873-82
[2] Henning et al., NMR Biomed 2009; 22(7):683-96
[3] Kirchner et al., Magn Reson Med 2015; 73(2):469-80
[4] Strasser et al., Proc. Intl. Soc. MRM 21 (2013):2018
[5] Hangel et al., NMR in Biomed 2015; 28(11):1413-25
[6] Bilgic et al., MRM 2013; 69(6):1501-11
[7] Bilgic et al., JMRI 2014; 40(1):181-191
[8] Považan et al., Proc. Intl. Soc. MRM 23 (2015): 1973

Figures

Figure 1: A) DIR-FID-MRSI sequence scheme; B) The symmetric frequency sweep of the long IR-pulses leads to different inversion times for lipids and metabolites, reducing the metabolite SNR loss due to DIR.

Figure 2: Comparison of quality parameters for NLS and all lipid suppression methods. DIR features the least remaining lipid signal but also the least SNR.

Figure 3: Example of spectral fits with residua of a voxel in the occipital lobe in close proximity to the cranium for NLS and all lipid suppression methods. All three remove the lipid contamination in the metabolite region.

Figure 4: Lipid maps (sum of the signals in the 0-2 ppm range) of two volunteers for NLS and all lipid suppression methods, scaled to 1/100th of the maximum NLS lipid signal per volunteer. DIR shows the best performance in the brain while L2 reduces trans-cranial lipids the most.

Figure 5: NAA maps of all methods of a single volunteer. While all remove the lipid fold-in artefacts, the higher SNR of L1 and L2 allows to better discern the gyri structure on the maps.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
2354