In-vivo measurements were performed at 9.4T with the approval of the local ethics board. A custom-built head coil [1] consisting of 16 transmit and 31 receive channels was used.

Sequence optimization: the diagram of the proposed sequence is presented in fig. 1. Flip angles (FA) of water saturation pulses were numerically optimized for a T₁ range of 800-2800 ms and B₁⁺ inhomogeneities of ±50%. The slice-selection gradient shape was optimized to minimize sidebands, induced by mechanical gradient vibrations, which may hinder spectral quantification [2]. Specifically, the gradient frequency spectrum was numerically optimized to minimize mechanical resonances (≈550 and ≈1500 Hz, fig. 2a, dark gray bands). Additionally, acquisition duration, TR and FA were set to achieve optimal SNR [3]

GRAPPA accelerated MRSI: TE=1.6 ms, TR=138 ms, FA=25°, spectral bandwidth=6000 Hz, acquisition duration=85 ms, for high-resolution spectra. TE=1.6, TR=138 and 102ms, FA=25°, spectral bandwidth=6000Hz, acquisition duration=85 and 42ms for low-resolution spectra. Total acquisition time (TA) without GRAPPA acceleration was 5 min 8 sec (high-resolution) and 2 min 8 sec (low-resolution).

Functional MRSI: matrix of 32×32 voxels, 2 weighted averages, nominal voxel size of 6×6×10 mm³ and otherwise identical sequence parameters. TA was 2 min 8 sec. For comparison, BOLD images were collected with GRE-EPI (TR=500 ms, TE=20 ms, nominal FA=50°, GRAPPA R=3, voxel size 1mm isotropic). The visual stimulus consisted of a flickering (7 Hz) radial checkerboard: 6 interleaved blocks (off-on), 2 min 8 sec each. TA was 12 min 28 sec.

Post-processing and quantification: An additional non-water-suppressed MRSI-FID data set was acquired for data reconstruction with adaptive combine coil combination [4], GRAPPA calibration and eddy current correction [5]. High-resolution MRSI data were retrospectively undersampled to simulate different GRAPPA acceleration factors. In order to reduce the g-factor loss, CAIPIRINHA patterns [6] were used for simulating factor 2 and 4 acceleration. Spectra were evaluated with LCModel [7] using a basis-set simulated with VeSPA [8, 9]. GRE-EPI functional images were post-processed using FSL-FEAT [10].

Fig. 2 shows the results of the numerical optimization for sidebands reduction. The best and worst (blue and red) performing gradient shapes together with their frequency spectra are displayed in fig. 2a (left and right side). The frequency spectrum of the best performing gradient has the local minima overlapping with the two forbidden frequency ranges, whereas the worst performing gradient shows exactly the opposite. Phantom and in-vivo spectra (2b and c) confirmed the results of the optimization procedure. In both cases the sidebands at ≈550 and ≈1100 Hz were greatly reduced. Moreover, even the sideband at ≈1500 Hz (ignored in the optimization) was significantly smaller.

The results of GRAPPA accelerated high- and low-resolution MRSI are presented in fig. 3. Apart from some noise enhancement, the quality of all presented spectra is similar.

Figure 4 shows the results of functional MRSI of the visual cortex. Here the spectra were displayed without 1st order phase correction. Instead, the same phase error was introduced to the basis-set used for quantification with LCModel. Both single (b) and mean ‘rest’ vs. ‘active’ spectra (c) show differences in the region between 1.8 and 2.5 ppm, which is associated with two major neurotransmitters: gamma-aminobutyric acid (GABA) and glutamate (Glu). This is confirmed by the average time courses (d) calculated for both metabolites over five voxel region associated with positive BOLD response. A clear correlation between the changes in GABA/tCr and Glu/tCr concentration ratios and the stimulation periods can be seen. The average differences between the stimulus on- and off-set were ≈13% and ≈11%, for GABA /tCr and Glu/tCr, respectively.

The proposed MRSI-FID sequence enables, reliable acquisition of proton spectra with reduced sideband artifacts, high spatial resolution, minimized sensitivity to B₀ and B₁ inhomogeneities, very short acquisition delay and SNR optimized acquisition duration at 9.4T. The high SNR makes it theoretically possible to reduce the acquisition time even further by utilizing parallel imaging techniques, whereas high temporal resolution enabled the assessment of functional related changes in metabolite concentrations during visual stimulation.

Our observations regarding the functional MRSI are in accordance with the results published previously [11-13]. However, strong contaminations with lipid signal currently still hinders the analysis of the spectra from the regions close to the scalp. Further studies, with a large number of participants will be necessary to elucidate the observed changes in concentrations of Glu and GABA in the regions associated with positive BOLD response.