To quantitatively examine the SNR implication of multiple-voxel averaging from high-resolution CSI data, resulting in a significant SNR loss compared to a single voxel from low-resolution CSI data with matched spatial resolution. High-resolution chemical shift imaging (hrCSI) is highly desired for basic and clinical research due to less partial volume effect, thus better specificity for identifying and studying diseased tissue. However, lower signal-to-noise ratio (SNR) caused by a small voxel size impedes reliable quantification of metabolites, particularly challenging for low-γ X-nuclear MRS, such ¹³C and ¹⁷O spins. As one of solutions to overcome the SNR issue, spatial averaging of multiple hrCSI voxels taken from the region of interest has been frequently employed. In this context, it is interesting to examine if the averaged signal from multiple hrCSI voxels could substantially sacrifice a SNR compared to a large single voxel acquired from low-resolution CSI (lrCSI) with matched spatial resolution and position, and if yes, how much? To address this critical question, we conducted a thorough SNR comparison study of X-nuclear ¹⁷O CSI of natural abundance water in phantom and human occipital lobe in vivo with two spatial resolutions. All experiments were performed at 7.0T/90-cm bore scanner (Siemens) using RF surface coil and 8 healthy volunteers participated this study. The Fourier Series Window (FSW) technique ^[1-2] was used to acquire all 3D ¹⁷O CSI data under full relaxation condition (¹⁷O T1 = 4.5ms, TR = 100ms, 1610 k-space scans; spectral width = 30kHz; 9x9x7 phase encodes using two field of views (FOV): (8×8×6 cm³) hrCSI and (24×24×18 cm³) for lrCSI, resulting in 27 times of volume difference. A total of 27 hrCSI voxels’ FIDs (see Figs.1 and 2) were summed in time domain. No line broadening was applied to measure true noise levels and SNR. The RF transmission field (B₁⁺) of ¹⁷O RF coil was mapped (Fig. 2) for calculating flip angle (B₁⁺) and its effect on the ¹⁷O water intensity. The noise correlation among inter-voxels in the hrCSI was measured. Finally, the effect of point-spread function (PSF) was estimated based on the cylindrical voxel shape from the FSW-CSI method ^[1-2].

Figure 1 shows a typical in vivo ¹⁷O CSI distribution of H₂¹⁷O signal and the selection of target voxels of interest used for SNR comparison. Noise and linewidth (FWHM) of single voxel spectra were not significantly different between hrCSI and lrCSI data, and independent on the voxel position. However, the summed spectrum from 27 hrCSI voxels showed a much high noise level as well as higher signal intensity (Fig. 1D) compared to the single voxel spectrum from the lrCSI with the similar sample volume and position (Fig. 1E). The signal level difference indicates that the true hrCSI voxel size was larger than the apparent voxel size due to the spatial overlapping associated with PSF (Fig. 4). The estimated PSF from the FSW-CSI method accounted ~41% signal increase. The B₁⁺ maps shown in Fig. 2 were highly inhomogeneous among hrCSI voxels and had effects on the ¹⁷O signal, for example, counting for ~35% and ~33% signal loss for phantom and in vivo, respectively. In addition, PSF induced the noise coherence among adjacent hrCSI voxels as illustrated in Fig. 3, resulting a high noise level ratio (phantom ~7.8, in vivo ~ 9.1) compared to low resolution CSI. With consideration and correction of the PSF and B₁⁺ effects on each hrCSI voxel, phantom and in vivo brain data consistently suggest that multiple-voxel averaging of hrCSI data resulted in a huge SNR reduction (~5.4 times) compared to that of the lrCSI. This experimental result was close to the theoretical prediction of 5.2 based on the relationship between the multiple-voxel averaged SNR (SNR_m) and single voxel SNR (SNR_s): $$$SNR_m = (\sqrt{NumberofMeasurement} \times SNR_s $$$ if inter-voxel noise is completely random and independent (i.e., PSF=0), and B₁ distribution is homogenous in space. As we demonstrated, this theory might be invalid for practical CSI application commonly with significant PSF and B₁ inhomogeneity, especially for a surface coil used in this study, thus, the SNR_m calculation should be modified accordingly: $$Signal_m =(\sum_{i=1}^{27}Signal_i\times sin(\alpha_i))\times PSF$$ $$Noise_m = NoiseCorrelation\times \sqrt{{Numberof Voxels}}\times \sigma_n$$ where σ_n is the mean voxel noise.

This study demonstrates that the spatial voxel averaging of hrCSI data results in a huge SNR trade-off, which should be useful for selecting an appropriate voxel size of high-resolution CSI for in vivo application. A better strategy is to optimize the CSI voxel size for achieving the desired SNR, rather to averaging multiple hrCSI voxels for matching the same SNR that requires a much long scanning time.