In single voxel proton Magnetic Resonance Spectroscopy (MRS), low concentrations of metabolites of interest typically require the acquisition of a significant number of identical scans to improve the signal to noise ratio (SNR). Combining these together, the true signal adds coherently while the noise is incoherent, causing the SNR to increase as the square root of the number of averages. The crucial assumption of this technique is that the signal does not vary significantly between averages. Variation in the central frequency of the scanner causes displacements of the averages relative to each other, broadening and distorting the peaks.

One of the largest causes of frequency shift in MR scanners is heating of the gradient coils and passive shims¹. Although most spectroscopy sequences make only minimal use of gradients, it is not uncommon for MRS exams to be combined with echo planar imaging sequences such as for diffusion tensor imaging (DTI) or functional MRI. The fluctuations set up by these sequences gradually dissipate during the acquisitions that follow. In this study we explored the effects of gradient-heavy sequences on subsequent MRS acquisitions.

Datasets from 53 subjects were evaluated. The MRS exam was a standard PRESS localized single 20x20x20mm3 voxel in the posterior cingulate gyrus with TE/TR = 30ms/2s and 128 averages. The acquisitions were performed on a Siemens Magnetom Verio 3T scanner using a 32 channel head coil. An acquisition without water suppression (with only 16 averages) was also performed to allow eddy current correction and water reference quantification. During the same session each subject also underwent a DTI imaging protocol. In 31 cases the DTI was performed before spectroscopy, in the remaining 22 cases the spectroscopy was performed first.

Data was extracted from the console in both the Siemens .rda format and the more complete ‘twix’ format which contains all channel and average data. The .rda spectrum is created by simply averaging the repeat scans. For the twix data, channel weightings were determined from the water reference signal using a singular value decomposition based method designed to maximize SNR, then applied to the main data. The individual scans were frequency corrected using spectral registration². The rate of drift was calculated by a linear regression of frequency shifts across the averages. The resulting signals from each format were analyzed using LCModel³ to obtain metabolite concentrations and peak widths.

The MRS exams performed after DTI were substantially more affected by drift than those performed before, with a mean drift more than four times greater (1.52 Hz / min. compared to 0.35 Hz / min). Figure 1 shows a histogram of the measured drift values for the two groups.

When analysing the metabolite concentrations quantified by LCModel, for 19 of the 22 cases where spectroscopy was acquired first, no difference was found between the .rda spectra and those from the twix files. By contrast, for all 31 exams acquired after DTI, significant differences were observed between the standard .rda data and the frequency corrected data from twix. Figure 2 shows two example spectra, in one the peak distortion caused by the drift seriously compromises the phasing of the spectrum, while in the second an apparent lipid peak in the standard data is clearly resolved into the ethanol triplet when correction is performed. As shown in figure 3, the mean FWHM reported by LCModel for the corrected data was half of that for the normal spectra.

In addition, the metabolite concentrations were in some cases significantly altered. Figure 4 shows the shifts in concentration of some of the principal metabolites, which are largely increased, particularly for choline (σ=15.7%, max=54.4%), glutamate/glutamine (σ=13.0%, max=48.6%) and GSH (σ=15.3%, max = 109.5%). Perhaps most importantly, these changes altered the study results. As a result of frequency drift, significant differences in choline were observed between patients and controls which disappeared after correction.

MR spectroscopy is highly sensitive to the B0 drift produced by prior DTI imaging sequences. This can interfere with processing, either causing data to be discarded as low quality or leading to mis-quantification of metabolites. Spectral registration on the raw data allowed the spectral quality to be largely recovered for this dataset and this should always be applied in cases with significant drift. However, it is not a universal solution, B0 drift can cause water suppression to go off-resonance, leading to much larger water signals, and the method can also perform poorly for spectra with low SNR. A greatly preferable solution is simply to ensure that spectroscopy sequences are performed before gradient-heavy EPI sequences such as DTI, which will immediately greatly reduce the degree of drift.