In Vivo Glx Measurements From GABA-Edited HERMES at 3 T Are Not Consistent With Those From Short-TE PRESS Across Scanners, Brain Regions, Diagnostic and Age Groups

Abstract

1H-Magnetic resonance spectroscopy (1H-MRS) is a noninvasive technique for quantifying brain metabolites, including glutamate, glutathione (GSH), and γ-aminobutyric acid (GABA), which are essential for brain function and implicated in various neurodevelopmental conditions. As such, 1H-MRS methods that enable reliable and accurate measurement of these metabolites are of considerable clinical value. Hadamard Encoding and Reconstruction of MEGA-Edited Spectroscopy (HERMES; echo time [TE] = 80 ms) is a spectral editing technique that allows for the simultaneous quantification of GABA and GSH, using subtraction approaches to resolve these metabolites in a difference spectrum. Additionally, glutamate plus glutamine resonances (Glx) can be resolved either from the HERMES GABA-edited difference spectrum (GABA-DIFF) or from the sum of all HERMES transients (SUM spectrum). However, the reliability of 80-ms HERMES for quantification of Glx has not been systematically assessed. Here, we evaluate the agreement between Glx obtained from HERMES GABA-DIFF and SUM spectra with Glx derived from short-TE PRESS (TE = 35 ms), which is conventionally used for Glx estimation and has demonstrated reproducibility. Data were acquired from 139 participants across two brain regions (ACC and Thalamus voxels), three scanners, two diagnostic groups (autism and neurotypical development) and two age groups (adolescent/adult and preschooler). Comparisons were made using both creatine-scaled and tissue-corrected Glx estimates. Our findings demonstrate significant systematic and proportional bias between Glx estimates from HERMES (SUM and GABA-DIFF) and short-TE PRESS, consistent across scanners, voxels, age groups and diagnostic categories. These findings indicate that Glx estimates derived from HERMES are not directly comparable to those from short-TE PRESS, and this discrepancy is consistent across a multisite study setting. This underscores the importance of sequence selection and careful methodological consideration when integrating and interpreting data from 1H-MRS across different acquisition protocols.

Keywords: 1H‐MRS; ACC; Glx; HERMES; PRESS; Thalamus; agreement; edited 1H‐MRS.

FIGURE 1

Voxel placement and mean HERMES and PRESS spectra. (A) Mean HERMES GABA‐DIFF, HERMES SUM and PRESS spectra showing model fit, model baseline and median residual fits (error bars). (B) Standard T1‐weighed image displaying heat plots of the MRS voxels on the thalamus (blue) and ACC (red), created by normalising T1‐weighted scans and then voxels to a standard space (MNI52 T1 1 mm brain) and calculating the overlap. Heat plots are shown on a standard anatomical image (MNI52 T1 1 mm brain). Greater intensity (brighter) indicates increased overlap and thus consistency in MRS voxel placement between participants.

FIGURE 2

Summary of the acquisition and analysis workflow used. Statistical analysis methods and results are labelled accordingly.

FIGURE 3

Thalamus Glx concentrations estimated from paired HERMES GABA‐DIFF, HERMES SUM and PRESS spectra. Tissue corrected (i.u) and creatine‐scaled (/tCr) thalamus Glx concentrations estimated from the HERMES GABA‐DIFF (DIFF), HERMES SUM (SUM) and PRESS spectra are shown per scanner. Significant differences in paired Glx estimates across acquisition approaches after multiple comparison correction are indicated per scanner; *p _adjusted < 0.05, **p _adjusted < 0.01, ***p _adjusted < 0.001, ****p _adjusted < 0.0001. Also shown are spaghetti plots, which show paired thalamic Glx measurements from the same participant (HERMES GABA‐DIFF, HERMES SUM and PRESS).

FIGURE 4

ACC Glx estimated from paired HERMES GABA‐DIFF, HERMES SUM and PRESS spectra. Tissue corrected (i.u) and creatine‐scaled (/tCr) ACC Glx concentrations estimated from the HERMES GABA‐DIFF (DIFF), HERMES SUM (SUM) and PRESS spectra are shown per scanner. Significant differences in paired Glx estimates across acquisition approaches after multiple comparison correction are indicated per scanner; *p _adjusted < 0.05, **p _adjusted < 0.01, ***p _adjusted < 0.001, ****p _adjusted < 0.0001. Also shown are spaghetti plots, which show paired ACC Glx measurements from the same participant (HERMES GABA‐DIFF, HERMES SUM and PRESS).

FIGURE 5

Correlations between HERMES (SUM & GABA‐DIFF) and PRESS Glx estimates from the thalamus and ACC voxels. Spearman's rank partial correlation coefficients (rho) were calculated between Glx estimated from HERMES (GABA‐DIFF & SUM) and PRESS from the thalamus and ACC voxels while controlling for FWHM, per scanner and for tissue‐corrected (i.u.) and creatine‐scaled (/tCr) data. Where significant after Bonferroni correction, partial correlation coefficients are displayed on the graph. Yellow points = correlation between HERMES GABA‐DIFF and PRESS Glx estimates, light blue points = correlation between HERMES SUM and PRESS Glx estimates. Grey line = linear association.

FIGURE 6

(A) Bland–Altman (Giavarina version) plots comparing ACC Glx concentrations estimated from paired HERMES GABA‐DIFF and PRESS spectra per scanner. For each plot, the percentage difference between paired measures is shown on the y‐axis, while the mean of the paired measures is shown the x‐axis. Solid blue lines represent the overall mean percentage difference (estimated bias) while dashed black lines represent the upper and lower limits of agreement (overall mean difference ± 1.96 standard deviation). Confidence intervals (95%) for limits of agreement are also shown. Linear regressions (model = percentage difference of paired measured ~ mean of paired measures) were used to identify proportional bias between paired measures, the resulting beta‐coefficient and corresponding p value is shown in the right‐hand corner of each plot; *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. (B) Bland–Altman (Giavarina version) plots comparing thalamus Glx concentrations estimated from paired HERMES GABA‐DIFF and PRESS spectra per scanner. For each plot, the percentage difference between paired measures is shown on the y‐axis, while the mean of the paired measures is shown the x‐axis. Solid yellow lines represent the overall mean percentage difference (estimated bias) while dashed black lines represent the upper and lower limits of agreement (overall mean difference ± 1.96 standard deviation). Confidence intervals (95%) for limits of agreement are also shown. Linear regressions (model = percentage difference of paired measured ~ mean of paired measures) were used to identify proportional bias between paired measures, the resulting beta‐coefficient and corresponding p value is shown in the right‐hand corner of each plot; *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

FIGURE 7

(A) Bland–Altman (Giavarina version) plots comparing ACC Glx concentrations estimated from paired HERMES SUM and PRESS spectra per scanner. For each plot, the percentage difference between paired measures is shown on the y‐axis, while the mean of the paired measures is shown the × axis. Solid blue lines represent the overall mean percentage difference (estimated bias) while dashed black lines represent the upper and lower limits of agreement (overall mean difference ± 1.96 standard deviation). Confidence intervals (95%) for limits of agreement are also shown. Linear regressions (model = percentage difference of paired measured ~ mean of paired measures) were used to identify proportional bias between paired measures, the resulting beta‐coefficient and corresponding p value is shown in the right‐hand corner of each plot; *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. (B) Bland–Altman (Giavarina version) plots comparing thalamus Glx concentrations estimated from paired HERMES SUM and PRESS spectra per scanner. For each plot, the percentage difference between paired measures is shown on the y‐axis, while the mean of the paired measures is shown the × axis. Solid yellow lines represent the overall mean percentage difference (estimated bias) while dashed black lines represent the upper and lower limits of agreement (overall mean difference ± 1.96 standard deviation). Confidence intervals (95%) for limits of agreement are also shown. Linear regressions (model = percentage difference of paired measured ~ mean of paired measures) were used to identify proportional bias between paired measures, the resulting beta‐coefficient and corresponding p value is shown in the right‐hand corner of each plot; *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.