Comparison of Multivendor Single-Voxel MR Spectroscopy Data Acquired in Healthy Brain at 26 Sites

Abstract

Background

The hardware and software differences between MR vendors and individual sites influence the quantification of MR spectroscopy data. An analysis of a large data set may help to better understand sources of the total variance in quantified metabolite levels.

Purpose

To compare multisite quantitative brain MR spectroscopy data acquired in healthy participants at 26 sites by using the vendor-supplied single-voxel point-resolved spectroscopy (PRESS) sequence.

Materials and Methods

An MR spectroscopy protocol to acquire short-echo-time PRESS data from the midparietal region of the brain was disseminated to 26 research sites operating 3.0-T MR scanners from three different vendors. In this prospective study, healthy participants were scanned between July 2016 and December 2017. Data were analyzed by using software with simulated basis sets customized for each vendor implementation. The proportion of total variance attributed to vendor-, site-, and participant-related effects was estimated by using a linear mixed-effects model. P values were derived through parametric bootstrapping of the linear mixed-effects models (denoted P_boot).

Results

In total, 296 participants (mean age, 26 years ± 4.6; 155 women and 141 men) were scanned. Good-quality data were recorded from all sites, as evidenced by a consistent linewidth of N-acetylaspartate (range, 4.4–5.0 Hz), signal-to-noise ratio (range, 174–289), and low Cramér-Rao lower bounds (≤5%) for all of the major metabolites. Among the major metabolites, no vendor effects were found for levels of myo-inositol (P_boot > .90), N-acetylaspartate and N-acetylaspartylglutamate (P_boot = .13), or glutamate and glutamine (P_boot = .11). Among the smaller resonances, no vendor effects were found for ascorbate (P_boot = .08), aspartate (P_boot > .90), glutathione (P_boot > .90), or lactate (P_boot = .28).

Conclusion

Multisite multivendor single-voxel MR spectroscopy studies performed at 3.0 T can yield results that are coherent across vendors, provided that vendor differences in pulse sequence implementation are accounted for in data analysis. However, the site-related effects on variability were more profound and suggest the need for further standardization of spectroscopic protocols.

© RSNA, 2020

Online supplemental material is available for this article.

Summary

In a multisite study of brain MR spectroscopy performed at 26 sites, interindividual differences were found as the main contributor to variability, demonstrating that appropriately analyzed MR spectroscopy data acquired at different sites and with different scanners can be compared.

Key Results

• ■ Short-echo-time MR spectroscopy of the human brain performed at 3.0 T quantifies eight metabolites (N-acetylaspartate and N-acetylaspartylglutamate [total NAA], glycerophosphocholine and phosphocholine, creatine and phosphocreatine, myo-inositol [mI], glutamate and glutamine [Glx], glucose and taurine, gluthathione, aspartate) in all participants based on the defined quality exclusion criteria in a 27-mL voxel and approximately 2-minute acquisition time.

• ■ All MR spectra showed a consistent linewidth of N-acetylaspartate (range, 4.4–5.0 Hz), signal-to-noise ratio (range, 174–289), and low Cramér-Rao lower bounds (≤5%) for total NAA, glycerophosphocholine and phosphocholine, creatine and phosphocreatine, mI, and Glx.

• ■ The vendor contribution to total variance of metabolite levels ranged from 0% to 27%. P values were derived through parametric bootstrapping of the linear mixed-effects models (denoted as P_boot). No vendor effects were found for levels of mI (P_boot > .90), total NAA (P_boot = .13), Glx (P_boot = .11), ascorbate (P_boot = .08), aspartate (P_boot > .90), glutathione (P_boot > .90), and lactate (P_boot = .28). The site-related effects ranged from 3% up to 44% and the remaining 50%–83% of the total variance was accounted for by interindividual differences, which included both actual biologic differences as well as within-participant measurement error.

Introduction

Single-voxel proton (hydrogen 1 [¹H]) MR spectroscopy gives useful information for both clinical diagnosis and research studies (1,2) and is commercially available with most MRI scanners at 1.5 T and 3.0 T. The point-resolved spectroscopy (PRESS) sequence (3) is most widely used for brain MR spectroscopy because it provides robust results from well-defined regions of interest. However, implementations of the sequence are diverse among manufacturers and research sites. They differ in terms of timing and waveforms of the radiofrequency pulses, as well as water suppression and preparation methods, all of which may influence quantification. While recent efforts have focused on standardization of single-voxel MR spectroscopy methods between vendors, including the use of advanced localization sequences such as semi-localized by adiabatic selective refocusing (4,5), currently most clinical MR spectroscopy studies continue to use the manufacturer-supplied PRESS sequence.

At 3.0 T, conventional MR spectroscopy can reliably detect the signals of N-acetylaspartate (NAA) and N-acetylaspartylglutamate (total NAA), total choline (glycerophosphocholine and phosphocholine), total creatine (tCr) (creatine and phosphocreatine), myo-inositol (mI), and glutamate and glutamine (Glx). These metabolites are commonly known as major metabolites. In addition, a variable number of less abundant minor metabolites are often reported, based on the spectral quality and the defined exclusion criteria.

Recently, a large study involving almost 300 healthy participants at multiple sites was performed, investigating the reliability of the measurement of brain γ-aminobutyric acid levels by using the Mescher-Garwood (MEGA)-PRESS (6) spectral editing technique (7,8). As a second part of that study, conventional short-echo-time MR spectroscopy data were also collected by using the vendor-supplied PRESS sequence. The current study reports on the results of the analysis of these single-voxel PRESS data, which could have important implications for the design of multisite studies using MR spectroscopy as an outcome measure, as well as for comparing clinical or research results in individual participants performed with different scanners.

The purpose of our study, therefore, was to compare multisite quantitative brain MR spectroscopy data acquired in healthy participants at 26 sites by using the vendor-supplied single-voxel PRESS sequence.

Materials and Methods

Study Participants

All data in our prospective study were acquired between July 2016 and December 2017. Institutional review board approval and participant written informed consent were obtained at each site. Participants were recruited with the following inclusion criteria: age between 18–35 years, no known neurologic illness, and medication free without usage of over-the-counter pain relievers or allergy medication in the preceding 36 hours. Anonymized data were analyzed at the Johns Hopkins University (Baltimore, Md). As stated previously, participants’ brain γ-aminobutyric acid levels recorded by using a spectral editing sequence were previously reported in the literature (7,8). This study reports levels of metabolites collected by using conventional PRESS sequence.

Data Acquisition

Each site acquired vendor-specific ¹H MR spectroscopy PRESS data with a 3.0-T scanner using phased-array radiofrequency head coils. Eight sites were equipped with GE scanners (GE Healthcare, Chicago, Ill), 10 with Philips scanners (Philips Healthcare, Best, the Netherlands), and eight with Siemens scanners (Siemens Healthineers, Erlangen, Germany) (Table E1 [online]). Three-dimensional T1-weighted images were acquired to position a 30 × 30 × 30 mm³ MR spectroscopy voxel in the medial parietal lobe (Fig 1). The ¹H MR spectroscopy data were measured with the following parameters: repetition time, 2000 msec; echo time, 35 msec; number of averages, 64; spectral bandwidth, 2 kHz or 4 kHz or 5 kHz; data points, 2048 or 4096; scan time, 2 minutes 8 seconds. Water reference scans were acquired for coil combination and eddy current correction with similar parameters and 8 or 16 averages. No other MR acquisition parameters were specified to participating sites. Therefore, additional acquisition-related methodology including the methods for main magnetic field (B₀) homogeneity adjustment, phase cycling scheme, water suppression, and outer volume suppression varied from site to site according to local preferences and availability. The ¹H MR spectroscopy data were saved as P-files (GE), SDAT/SPAR files (Philips), or TWIX files (Siemens). An overview of hardware and software parameters is provided in Table E1 (online).

Figure 1:

Image shows example of 30 × 30 × 30 mm³ spectroscopic voxel placement on T1-weighted image. Voxel was rotated in sagittal plane to be parallel with line connecting genu and splenium of corpus callosum. L = left, R = right.

Data Quantification

Data were processed by a medical physicist (M.P., with 7 years of experience) by using automated in-house Matlab-based (MathWorks, Natick, Mass) software (9) with the FID-A toolkit (10). Results were evaluated in consensus with all coauthors. Raw data (GE and Siemens) were coil combined using water reference (11) and eddy current corrected, frequency corrected, and phase corrected (12). Subsequently, spectra were quantified by using LCModel (version 6.3–0D; Stephen Provencher, Oakville, Canada) (13) over the frequency range from 0.5 ppm to 4.0 ppm. Basis sets consisted of 19 metabolites: alanine, ascorbate, aspartate, creatine, γ-aminobutyric acid, glucose, glutamine, glutamate, glutathione, glycerophosphocholine, lactate, mI, NAA, N-acetylaspartylglutamate, phosphocholine, phosphocreatine, phosphoethanolamine, scyllo-inositol, and taurine. Metabolite basis spectra were generated by using custom-built fully localized two-dimensional density matrix simulations and vendor-native shaped refocusing pulse information and phase cycling (14). Chemical shifts and scalar coupling (J) constants from Tkáč et al (15) were used for all metabolites except for NAA, N-acetylaspartylglutamate, glutamate, and glutamine, which used values from Govind and colleagues (16). Macromolecules were modeled in LCModel (Appendix E1 [online]). Four vendor-specific basis sets were used to fit the data (two for GE [due to different sequence timings], one for Philips, and one for Siemens).

All spectra were visually inspected to detect any spurious artifacts or lipid contamination. A spectrum was excluded from further analysis if the lipid peak amplitude was higher than the amplitude of the NAA singlet (-CH₃ group at 2.008 ppm).

Quantified values were processed based on quality criteria as described by Tkáč and colleagues (17). Metabolite values with Cramér-Rao lower bounds (CRLB) above 100% were excluded. If a specific metabolite did not have CRLBs less than 100% in at least 50% (147 of 293) of the quantified spectra, then it was excluded from further analysis. Finally, a metabolite was also excluded if it had average CRLBs higher than 50%. Signal-to-noise ratio (SNR) was calculated as a ratio of NAA signal intensity divided by a standard deviation of noise. Linewidth was calculated as a full width at half maximum of the NAA peak.

The relative concentration of metabolites was calculated as a ratio to tCr.

Statistical Analysis

Linear mixed-effects models were fit to the metabolite data in R (version 3.5.3; R Foundation, Vienna, Austria [18]) (M.M. [neuroscientist with 12 years of experience with statistics]). An unconditional model (Equation 1 in Mikkelsen et al [7]) was fitted to calculate variance partition coefficients to estimate the proportion of total variance attributed to vendor-, site-, and participant-related effects. Secondary, conditional linear mixed-effects models (Equation 5 in Mikkelsen et al [7]) were also fitted to assess the impact of SNR, linewidth, age, and sex on quantified measurements. Goodness of fit was calculated as a log-likelihood statistic. Parametric bootstrapping was then used whereby the null distributions of each test was simulated by bootstrapping the fitted distributions (2000 simulations) to evaluate the probability of observing the test statistics given the distribution of the null hypotheses (19). P values are denoted P_boot.

Effects that were tested included vendor and site, SNR and linewidth, and age and sex. Posthoc pairwise comparisons were corrected for multiple comparisons by using the Holm-Bonferroni method (20). P values are denoted P_Holm. An adjusted P value less than .05 was considered to indicate statistical significance.

Results

Study Participants

Twenty-six research sites participated in our study, with each site contributing up to 12 data sets. In total, data from 296 healthy participants were collected (155 women and 141 men; mean age ± standard deviation, 26 years ± 4.6; age inclusion range, 18–35 years). After visual inspection, three participants (examined by using Siemens scanners) were removed from further analysis due to lipid contamination. There were no differences in sex (P = .69) of participants between the groups examined by using scanners from the different vendors. Participants examined by using Siemens scanners were older by approximately 2 years compared with those examined at GE (P_Holm = .03) and Philips (P_Holm = .005) sites. Demographic information on participants is given in the Table.

Participant Demographic Data by Site

Data Quantification

Mean PRESS spectra are displayed in Figure 2, A–C, for each vendor. Qualitatively, spectral profiles were in very good agreement across vendors. The creatine and phosphocreatine peak of 3.9 ppm was consistently lower for data acquired by using GE scanners compared with the other two vendors due to higher bandwidth of the water suppression pulses (Figure 2, A–C; dashed line). Figure 2, D–F, shows vendor-specific mean fits and residuals from the LCModel fitting routine. An example of a quantified spectrum with fitted basis spectra of metabolites and macromolecules is shown in Figure 3.

Figure 2:

Graphs show average spectra from, A, GE Healthcare (Chicago, IL), B, Philips Healthcare (Best, the Netherlands), and, C, Siemens Healthineers (Erlangen, Germany) scanners with gray areas representing ±1 standard deviation. Average LCModel (version 6.3–0D; Stephen Provencher, Oakville, Canada) fit results (red) and average residuals are shown for, D, the GE, E, the Philips, and, F, the Siemens scanners. Dashed line illustrates difference in methylene peak of creatine and phosphocreatine at 3.9 ppm that was partially suppressed by water suppression. Glx = glutamate and glutamine, NAA = N-acetylaspartate, mI = myo-inositol, MM = macromolecules, tCho = total choline, tCr = total creatine.

Figure 3:

Image shows example LCModel (version 6.3–0D; Stephen Provencher, Oakville, Canada) analysis with one spectrum, LCModel fit (red), spline baseline, residual, and individual macromolecule and metabolite contributions. Gray band highlights spectral region approximately 3.2 ppm, where signals of phosphoethanolamine (PE), myo-inositol (mI), glucose (Glc), taurine (Tau), glycerophosphocholine (GPC), and phosphocholine (PC) substantially overlap. Spectral fitting algorithm based on linear combination of model spectra as used in this study may be affected by this spectral overlap. This may introduce additional variability that is reflected on vendor level. Asc = ascorbate, Asp = aspartate, Cr = creatine, GABA = γ-aminobutyric acid, Gln = glutamine, Glu = glutamate, GSH = glutathione, Lac = lactate, NAA = N-acetylaspartate, NAAG = N-acetylaspartylglutamate, PCr = phosphocreatine, sI = scyllo-inositol, CrCH2 = correction term for Cr, MM09 = macromolecule peak at 0.9 ppm, MM12 = macromolecule peak at 1.2 ppm, MM14 = macromolecule peak at 1.4 ppm, MM17 = macromolecule peak at 1.7 ppm, MM20 = macromolecule peaks at 2.0 ppm and 3.0 ppm, Lip13 = lipid peak at 1.3 ppm, Lip20 = lipid peaks at 2.0 ppm and 2.2 ppm and 2.8 ppm.

The overall high quality of the data is reflected in high SNR and narrow linewidth values. Vendor-specific mean SNR ± standard deviation for GE, Philips, and Siemens scanners were 174.3 ± 37.4, 196.3 ± 52.9, and 288.8 ± 80.6, respectively. Vendor-specific mean linewidth of NAA for GE, Philips, and Siemens was 5.0 Hz ± 0.8, 4.6 Hz ± 1.0, and 4.4 Hz ± 0.6, respectively. The individual site and vendor SNR and linewidth values are shown in Figure E1 (online). SNR values of data acquired by using Siemens scanners was higher than were SNR values acquired by using GE and Philips scanners (both P_Holm < .001). SNR values were not different between data acquired by using GE and Philips scanners (P_Holm = .35). There were no differences in linewidth between the data acquired by using scanners from the different vendors: GE versus Philips, P_Holm = .29; GE versus Siemens, P_Holm = .06; Philips versus Siemens, P_Holm = .32.

Based on the correlation matrix provided by LCModel (ie, mean correlation coefficient for a metabolite pair less than −0.3), sums of metabolites are reported for total NAA, tCr (creatine and phosphocreatine), glycerophosphocholine and phosphocholine (total choline), and taurine and glucose. Also, due to substantial spectral overlap of glutamate and glutamine (Glx) at 3.0 T, only their summed values (ie, Glx) are reported (21). Thus, 14 metabolites or metabolite sums were quantified. Mean CRLB of alanine was above 50% and was therefore excluded from further analysis. Eight metabolites were quantified in all participants, scyllo-inositol in 99.0% (290 of 293), phosphoethanolamine in 98.0% (287 of 293), γ-aminobutyric acid in 95.9% (281 of 293), lactate in 94.9% (278 of 293), and ascorbate in 81.9% (240 of 293) of all participants (Table E2 [online]).

Metabolite ratios relative to tCr and their relative CRLBs are listed in Tables E3 and E4 (online), and vendor-specific mean metabolite ratios compared by site are shown in Figure 4. CRLBs were below 2% for tCr and total NAA and below 5% for total choline, mI, and Glx across all sites and vendors (Table E4 [online]). Eleven of 13 metabolites acquired with Siemens scanners, 10 of 13 metabolites acquired with Philips scanners, and eight of 13 metabolites acquired with GE scanners had site-specific mean and vendor-specific mean CRLBs below 20%.

Figure 4:

Graphs show metabolite ratios to total creatine (tCr) shown by site and vendor. Black lines represent mean, darker areas represent 95% confidence intervals, and tinted areas represent ±1 standard deviation. Color-coded horizontal lines represent vendor-specific mean of plotted metabolite. A, Major metabolite signals: total N-acetylaspartate (NAA), total choline (tCho), myo-inositol (mI), and glutamate and glutamine (Glx) (plotted on range of ±50% of their grand mean). B, Smaller signals: ascorbate (Asc), aspartate (Asp), γ-aminobutyric acid (GABA), glucose (Glc) and taurine (Tau), glutathione (GSH), lactate (Lac), phosphoethanolamine (PE), and scyllo-inositol (sI) (plotted on range of ±100% of their grand mean).

Mean between-participant coefficients of variation (Table E5 [online]) and intersite coefficients of variation (Table E2 [online]) were below 10% for total choline/tCr, total NAA/tCr, Glx/tCr, and mI/tCr. Between-participant coefficients of variation are compared against site-mean CRLBs in Figure 5. Four major metabolite ratios (total choline/tCr, Glx/tCr, mI/tCr, and total NAA/tCr) showed substantially lower CRLBs (in the range of 1%–5%) compared with their corresponding between-participant coefficients of variation (range of 4%–15%).

Figure 5:

Graphs show between-participant coefficient of variation plotted against site-mean Cramér-Rao lower bounds (CRLBs) (error of fitting) for reported metabolite ratios to total creatine (tCr) displayed for, A, four major metabolites with lower CRLBs: total N-acetylaspartate (tNAA), total choline (tCho), myo-inositol (mI), and glutamate and glutamine (Glx). B, Smaller signals with higher CRLBs. Black line is identity line. Color-coded by vendors. Asc = ascorbate, Asp = aspartate, GABA = γ-aminobutyric acid, Glc = glucose, GSH = glutathione, Lac = lactate, PE = phosphoethanolamine, sI = scyllo-inositol, Tau = taurine.

Statistical Analysis

The vendor contribution to the total variance of the metabolite concentrations ranged from 0% to 27%. The site-related effects were more pronounced and found for all but one metabolite (lactate; P_boot = .20) with site-level variance ranging from 3% up to 44% of total data variance. The remaining 50%–83% of the total variance was accounted for by interindividual differences, which included both actual biologic differences and participant-to-participant measurement error.

Results of the linear mixed-effects modeling are summarized in Table E6 (online). The linear mixed-effects modeling showed the effects of vendor on total choline/tCr (P_boot < .001), γ-aminobutyric acid/tCr (P_boot < .001), phosphoethanolamine/tCr (P_boot < .001), scyllo-inositol/tCr (P_boot = .03), and glucose and taurine/tCr (P_boot = .003) but not on mI/tCr (P_boot > .90), total NAA/tCr (P_boot = .13), Glx/tCr (P_boot = .11), ascorbate/tCr (P_boot = .08), aspartate/tCr (P_boot > .90), glutathione/tCr (P_boot > .90), or lactate/tCr (P_boot = .28). There was an effect of site on all metabolites (all P_boot < .02) except for lactate/tCr (P_boot = .20).

The conditional linear mixed-effects modeling showed an effect of SNR on ascorbate/tCr (P_boot < .001), glutathione/tCr (P_boot < .001), and mI/tCr (P_boot = .02) measures. Linewidth was related to aspartate/tCr (P_boot < .001), γ-aminobutyric acid/tCr (P_boot < .001), glutathione/tCr (P_boot = .05), phosphoethanolamine/tCr (P_boot = .002), and Glx/tCr (P_boot = .007). No metabolites showed an age effect except for lactate/tCr (P_boot = .007). Sex differences were observed only for ascorbate/tCr (P_boot = .03) and total choline/tCr (P_boot = .003).

Discussion

We demonstrated that short-echo-time point-resolved spectroscopy (PRESS) MR spectroscopy of the human brain performed at 3.0 T can quantify eight different metabolites in all 293 participants based on the defined quality exclusion criteria in a 27-mL voxel and approximately 2-minute scan time. Our reported metabolite ratios are in agreement with previously published values (21–23). Uncertainty estimates (ie, Cramér-Rao lower bounds [CRLBs]) of the major metabolites (ie, glutamate and glutamine [Glx], myo-inositol [mI], total choline, total creatine [tCr], and total N-acetylaspartate [NAA]) were under 5% in all data sets, consistent with prior studies that have also reported less than 5% CRLBs for these compounds at 3.0 T (24). Furthermore, despite no effort being made to standardize the acquisition beyond harmonization of relatively basic parameters (repetition time, echo time, voxel size, and scan time), there was no vendor effect observed for seven of the 12 metabolite-to-tCr ratios that were reported, including the major metabolites total NAA, Glx, and mI, as well as ascorbate, aspartate, lactate, and glutathione among the smaller resonances.

Interestingly, there was a between-vendor difference of 16% for total choline, which is somewhat surprising given the lack of difference for the other major metabolite resonances. We found a vendor effect for four other metabolites (scyllo-inositol, phosphoethanolamine, glucose, and taurine) and one macromolecular signal (MM17). All these metabolites occur in a region of the spectrum (approximately 3.2 ppm to approximately 3.4 ppm) that is heavily overlapped with a number of multiplet signals. This vendor-level variability is, therefore, most likely driven by the ability of the spectral fitting algorithm to distinguish between these heavily overlapped resonances (25) and is influenced by the vendor-specific prior knowledge available, rather than by intrinsic between-vendor spectral differences per se. Consistent with this, resonances without substantial overlap (eg, macromolecules or lactate) or those that have at least one dominant peak with minimal overlap (total NAA at 2.01 ppm, creatine and phosphocreatine at 3.03 ppm, Glx at 3.75 ppm, and mI at 3.52 ppm) were less affected by vendor-specific effects. Site-specific differences in methodology were reflected in higher contribution of site effect in total variance of most metabolites, including the major metabolites that ranged from approximately 15% for total choline and mI up to approximately 44% for Glx. This effect needs to be considered in future multisite studies; however, it may be substantially minimized by standardization of spectroscopic sequences.

Multiple technical factors influence the appearance and quality of brain MR spectra. Linewidths reflect the success of the shimming routine used, which was not standardized between sites in our study. Our relatively narrow and consistent linewidths are similar to those reported in the literature (21,24). Our data suggest that the choice of shimming method does not influence results for the midparietal brain region (which is relatively distant from magnetic susceptibility effects) (26). SNRs) depend on many factors, including radiofrequency coils used, voxel size, and pulse sequence efficiency. In our study, SNR was higher for Siemens scanners, which can be attributed to differences in how the width of the slice-selective pulses are defined by the vendors (27) (ie, the Siemens data were collected with larger actual voxel volumes compared with the other scanners, due to differences in the profiles of the slice selective pulses used for spatial localization). Although linewidth and SNR have an impact on quantification accuracy and precision (28), SNR and linewidth were not major factors influencing the variance of our data because of the large voxel size and favorable brain region used.

Importantly, we customized our analyses for each vendor. LCModel basis sets were customized to account for between-vendor differences in the PRESS sequence (ie, timing of the radiofrequency pulses and full spatial simulations of the refocusing pulse profiles). Four sites equipped with GE scanners used PRESS sequences with slightly different interpulse timings, spectral bandwidth, and number of acquired points, so we created a separate basis set for these sites.

As was the case for the shimming routine, the method of water suppression was not standardized. A variety of water suppression modules are available (29), and different scanners used different approaches. In addition, the bandwidth of the water suppression pulses was not standardized, which influenced spectral appearance near the water peak. The mI peak at 4.05 ppm was suppressed by the water suppression pulses in data sets acquired by using GE scanners and partially in data sets acquired by using Philips scanners. To a smaller extent, this effect was also observed with creatine and phosphocreatine at 3.9 ppm. To avoid this impacting spectral quantitation, the 4.1-ppm mI peak was excluded from the LCModel analysis by limiting the spectral range to 0.5–4.0 ppm, and the correction term (CrCH2) was included in the basis set to account for effects on creatine and phosphocreatine at 3.9 ppm.

Outer-volume suppression pulses were implemented in combination with PRESS localization on some scanners but not others. Outer-volume suppression helps to improve spatial localization by suppressing unwanted signals from outside the voxel, especially those from pericranial lipids that can complicate spectral analysis. We had to exclude three data sets acquired by using Siemens scanners due to lipid contamination, two of which did not include use of such pulses.

Finally, CRLBs of the major metabolites were substantially lower compared with the between-participant coefficients of variation. This has previously been suggested to reflect the dominance of interparticipant physiologic or measurement variance over the error in spectral fitting (17,30). Between-participant coefficients of variation and CRLBs were more similar for the lower-concentration metabolites. Thus, the contribution of the modeling error to between-participant differences cannot be neglected for these compounds.

Our study had several limitations. It is likely that voxel position varied slightly from participant to participant, because voxels were placed manually by different operators at each site. In addition, the introduction of a measured macromolecular spectrum into a spectral quantification (31,32) would potentially remove some of the fitting variability most notably within the range from 3–4 ppm, as the chosen macromolecular model does not cover this region.

In conclusion, our results are encouraging for the implementation of MR spectroscopy in large multisite studies of major metabolites using 3.0-T MR scanners from different vendors and also for allowing the quantitative comparison of spectra recorded at different sites (eg, for diagnostic purposes) to be more readily compared. Although single-site and single-vendor studies are ideal in minimizing variance of MR spectroscopy data, in many instances these are not feasible because it is impossible to recruit sufficient numbers of patients at individual sites. An ideal future study would image participants at all sites and might include imaging of a subset of volunteers across multiple sites to minimize between-participant variance while leaving between-site and between-vendor variance to be investigated.