hide
Free keywords:
-
Abstract:
Purpose/Introduction: Compressed sensing (CS) is an acceleration method that has been used in many MRI applications1, 2. CS randomly undersamples and reconstructs the data using denoising methods. This method allows high acceleration without structured aliasing artifacts. However, the applications in MR spectroscopicimaging are few3, 4.
In this work, we use compressed sensing to accelerate the acquisition of non-lipid suppressed high resolution 1H FID MRSI data of the human brain at 9.4T. Metabolite maps were evaluated for different acceleration factors. This is the first account of applying compressedsensing to accelerate non-lipid suppressed and high-resolution MRSI data.
Subjects and Methods: Fully sampled single-slice MRSI data was acquired with an FID-MRSI sequence5 on a 9.4T whole-body human scanner. The parameters were: FOV = 200 9 200 mm, resolution = 3.125 9 3.125 9 10 mm, TR = 220 ms from the brains of healthy volunteers. The data were retrospectively undersampled with a random variabledensity sampling scheme with effective acceleration factors of: 2, 4, 5, and 10. The reconstruction was performed using a 3D total totalvariation
and a 2D wavelet. Multiple coil channels were combined
during the reconstruction using ESPIRiT6.
Results: Figure 1 shows the spectra of different voxels for different acceleration factors overlaid on the fully sampled spectra. For R = 2, the accelerated data is very consistent with the fully sampled data, while R = 4 results in only slightly higher lipid contamination. However, the higher accelerations result in more lipid contamination. This is
supported by the lipid contamination maps shown in Figure 2 (for two volunteers). These maps show the absolute integral of the spectra between 0.3 and 1.8 ppm. The maps for four major metabolites for two volunteers are shown in Figure 3. No spatial smoothing or filtering was performed. For
R = 2, the maps look almost identical to the fully sampled. The maps get noisier for higher acceleration factors.
Discussion/Conclusion: The metabolite maps get patchier as the acceleration factor increases. This is likely due to the higher levels of lipid contamination that affects the quantification of the spectra. However, note that unlike conventional acceleration schemes such as GRAPPA and SENSE, the lipid artifacts show noise-like contamination as opposed to structured and strong lipid rings overshadowing the metabolite maps. We showed that compressed sensing for non-lipid suppressed MRSI is possible and that high acceleration factors (up to R = 4 or R = 5) is feasible using our proposed reconstruction scheme. Lipid suppressed
data could allow for higher acceleration factors, however lipid suppression at ultra-high fields is very challenging.