Reproducibility of Metabolic Mapping Using 2D Multi‐Slice Short‐TE and GABA‐Edited Spin‐Echo MRSI at 3T in a Single Protocol

ABSTRACT

MRSI is a non‐invasive tool for mapping metabolic distributions in multi‐focal or other diseases where the location of abnormalities may be uncertain. High‐concentration metabolites can be investigated at 3T using non‐edited MRSI, whereas lower‐concentration metabolites, such as GABA, typically require specialized editing techniques because of spectral overlap. This study reports on the reproducibility of a protocol containing both co‐localized short‐TE and GABA‐edited multi‐slice spin‐echo 2D MRSI. Multi‐slice, short‐TE (TE 20 ms) and GABA‐edited (TE 68 ms) MRSI at a nominal spatial resolution of 2.2 cm³ was performed twice (7 to 14 days apart) at 3T on 11 healthy volunteers (age range 7 to 43 years). Data analysis was performed in the “Osprey” software package, including retrospective motion compensation, consensus‐recommended processing, and linear‐combination modeling. Metabolite estimates for six metabolites were quantified relative to total creatine (tCr) and water in 14 regions of interest. Reproducibility was assessed using intra‐ and inter‐subject coefficients of variation. Short‐TE MRSI metabolite estimates for total N‐acetylaspartate (tNAA), tCr, total choline (tCho), myo‐inositol (mI), and the sum of glutamate and glutamine (Glx) were found to be highly reproducible for both creatine‐ and water‐referenced concentration estimates, with 77% of the regions of interest meeting the quality‐control criteria for both visits and 96% for at least one visit. Average intra‐subject CVs were 5.8% and 4.8%, and inter‐subject CVs were 11.1% and 9.7% for water‐referenced and tCr‐referenced estimates, respectively. For GABA+ (GABA + macromolecules) estimates, 46% of the voxels of interest met quality‐control criteria for both visits, and 82% for at least one visit. In the remaining datasets, the average intra‐subject CVs were 13.5% for both quantification methods, and the inter‐subject CVs were 13.5% and 16.9% for water‐referenced and creatine‐referenced estimates, respectively. 3T‐MRSI sequences can achieve reproducible mapping with extended brain coverage of five major metabolites (tNAA, tCr, tCho, mI, and Glx). Reproducibility assessment for GABA+ mapping remains challenging, with 18% of the data being rejected in at least one visit, but it yielded acceptable reproducibility in datasets that met quality control criteria in both visits (46%).

Short‐TE MRSI at 3T enables reproducible metabolic mapping for tNAA, tCr, tCho, mI, and Glx; however, GABA+ mapping with edited MRSI remains challenging due to susceptibility to intermittent subject motion. Although 96% of regions from the short‐TE MRSI met quality control for major metabolites for at least one visit, only 82.5% of voxels of the GABA‐edited MRSI yielded acceptable data. The remaining regions yielded acceptable test–retest coefficients of variation for all metabolites.

Abbreviations

Cho

choline

Cr

creatine

−CrCH₂

negative creatine methylene

CV

coefficient of variation

GABA

gamma‐aminobutyric acid

Glx

glutamate and glutamine

GPC

glycerophosphocholine

GSH

glutathione

HGDB

hypergeometric dual‐band

LCM

linear‐combination modeling

Lip

lipids

mI

myo‐inositol

MM

macromolecules

NAA

N‐acetylaspartate

OVS

outer‐volume suppression

PE

phosphoethanolamine

sI

scyllo‐inositol

tNAA

total N‐acetylaspartate

TSC

tuberous sclerosis complex

1. Introduction

Magnetic resonance spectroscopic imaging (MRSI) has the potential to map neuro‐metabolic profiles [1], which is particularly valuable for studying multifocal, diffuse, or heterogeneous diseases such as brain tumors [2, 3, 4] and tuberous sclerosis complex [5, 6]. Multi‐slice, spin‐echo 2D‐MRSI with outer‐volume lipid suppression [7] has been shown to be useful in many clinical applications [8, 9, 10, 11]. This sequence allows for the quantification of several neuro‐metabolites, including N‐acetylaspartate (NAA), creatine (Cr), and choline (Cho) as well as myo‐inositol (mI) and the combined signal of glutamate and glutamine (Glx) [1]. Limiting factors for the detection of low‐concentration compounds, for example, γ‐aminobutyric acid (GABA), the brain's primary inhibitory neurotransmitter, are the detection sensitivity due to SNR and spectral overlap with other metabolites [1, 12, 13]. For single‐voxel localized spectroscopy at 3T, the spectral overlap can be resolved by using spectral‐editing techniques (e.g., MEGA‐PRESS) [14]. Spatial mapping of GABA would be a valuable addition to the metabolic profile measured with MRSI; however, combining MRSI and spectral‐editing is technically challenging.

Compared with conventional spin‐echo MRSI, GABA‐edited MRSI requires a compromise between spatial resolution and acquisition time to achieve sufficient SNR for reliable quantification [15]. Further, SNR improvement can be achieved by acquiring multiple averages; however, this increases the susceptibility to subject motion during the acquisition, which occurs even in cooperative subjects and is more likely in clinical populations. A potential solution is prospective motion correction, which interleaves a navigator to identify head movements, update the frequency, B₀‐shimming, and slice‐selection, and reacquire the affected voxel [15]. Although prospective motion correction is valuable, MRSI sequences with such features are not widely available and require substantial effort to be implemented for different vendors and scanners. Retrospective correction offers another solution to compensate for occasional, intermittent subject motion. For single‐voxel localization techniques, retrospective, shot‐by‐shot, phase‐ and frequency‐correction (and outlier rejection) are commonly used in the processing of GABA‐edited spectra [16, 17]. When spectral editing is combined with MRSI, retrospective motion correction is more difficult because of the varying phase‐encoding gradients. Recently, a retrospective motion‐compensation method was proposed for single‐slice 2D GABA‐MRSI at 3T, which compares multiple excitations for each point in k‐space [18]. In contrast to prospective motion correction, it requires no modification of the MRSI sequence and can be applied to any multi‐average MRSI k‐space data.

Establishing the reproducibility of new methods is a common step in biomedical imaging. Despite the wide potential application of MRSI, only a few studies have reported the reproducibility of metabolite estimates from short‐TE conventional MRSI [19, 20, 21, 22, 23, 24], and none of them used a multi‐slice 2D‐MRSI sequence. Even fewer studies investigated the reproducibility of GABA‐edited MRSI. Current literature includes studies of the reproducibility of 3D GABA‐edited MEGA‐LASER with prospective motion correction applied in the basal ganglia region [25]. However, for the clinical application of GABA‐edited MRSI in multifocal disease, the spatial coverage of cortical and subcortical areas is particularly of interest. In addition, reproducibility studies are commonly acquired in healthy, young, and cooperative subjects regardless of the prospective cohort in the potential application study [19, 20, 21, 22, 23, 24, 25]. Therefore, it is imperative to establish the reproducibility of more widely available MRSI sequences without prospective motion correction in healthy volunteers.

The purpose of this study was to evaluate the reproducibility of multi‐slice 2D short‐TE and GABA‐edited spin‐echo MRSI measurements after retrospective motion compensation in healthy volunteers at 3T. The protocol was designed for subsequent application in tuberous sclerosis complex (TSC), a genetic disorder characterized by multifocal brain lesions and disease onset early in life [5]. To this end, 9 healthy adults and 2 children were enrolled to establish the reproducibility of the protocol. Metabolite estimation was performed using expert consensus‐recommended linear‐combination modeling [12, 26] after retrospective motion correction [18] in the “Osprey” software package [27]. The test–retest coefficients of variation (CVs) were assessed in 14 brain regions typical for the occurrence of brain lesions in patients with TSC.

2. Methods

2.1. Subjects

Eleven healthy volunteers: 9 adults (3 M/6F, age 28 ± 6.7 years, min 22, max 43) and 2 children (1 M/1F, age 7/11 years) were scanned twice between 7 and 14 days apart (mean ± standard deviation = 10 ± 3 days). Written informed consent was obtained from all volunteers or their legal guardians prior to participation in the study.

2.2. MR Protocol

A 45‐min MR protocol including anatomical MRI, B₀ field maps for second‐order shim correction, short‐TE [7] and GABA‐edited [28] spin‐echo multi‐slice MRSI (Figure 1A), and a water reference MRSI was implemented (Philips 3T “Ingenia Elition,” 32‐channel head coil). Hypergeometric dual‐band (HGDB) water and lipid suppression pulses and eight outer‐volume suppression (OVS) pulses arranged in an octagonal pattern (Figure 1B) were used for both MRSI sequences [28, 29]. All MRSI scans were performed with three 15‐mm oblique‐axial slices (2.5‐mm gap), 14 × 17 matrix (elliptical k‐space sampling), nominal voxel size 12 × 12 × 15 mm (≈2200 mm³), 512 datapoints, 2000‐Hz spectral width, and T_acq = 0.256 s. For GABA‐editing, sequence parameters were TR/TE = 1.8 s/68 ms, 4 excitations (2 edit ON, 2 edit OFF), edit ON/OFF frequencies 1.9/0.7 ppm, editing pulse bandwidth 150 Hz (Sinc‐Gaussian pulse shape), scan time 22 m 21 s. For the edit OFF experiment, the edit frequency of 0.7 ppm was symmetric around the 1.3‐ppm lipid signal to reduce residual lipid signals in the difference spectrum [28]. The short‐TE spin‐echo MRSI parameters were TR/TE = 1.75 s/20 ms, 1 excitation, scan time 5 m 50 s. Parameters for water MRSI were TR/TE = 0.85 s/20 ms, 1 excitation, scan time 2 m 39 s without the OVS and HGDB pulses. All MRSI scans were acquired with the same slice locations and prescribed on a T₁‐weighted anatomical scan by an experienced MRS researcher (> 30 years of experience). Brain coverage was from the lateral ventricles to the vertex (Figure 1C). The volume prescription for the retest MRSI scans was performed using screenshots from the initial MRSI scan to ensure consistent positioning.

FIGURE 1.

(A) Spin‐echo GABA‐edited MRSI sequence diagram including HGDB + OVS preparation module for lipid‐ and water‐suppression. (B) Localization of the OVS pulses. (C) Localization of the three‐slice MRSI volume upwards from the lateral ventricles.

2.3. MRSI Processing

Reconstruction of MRSI data was performed using Osprey [27]. All MRSI datasets were reconstructed from the raw k‐space data (exported in Philips .data/.list format). For the GABA‐edited MRSI, retrospective motion compensation was applied to the raw k‐space data [18]. After spatial transformation, coil combination, and phase correction were applied using information from the water reference scan. Next, the GABA‐edited MRSI difference spectra were calculated, and residual water was removed using a Hankel singular value decomposition filter [30] for both the short‐TE and GABA‐edited MRSI data. For correct frequency referencing, the data were zero‐filled by a factor of 4, and a wavelet baseline correction was employed to remove residual lipids (frequency region between −2 and 1.85 ppm; regularization parameter λ = 10) and baseline signals (frequency region between −2 and 4.2 ppm; regularization parameter λ = 10,000) [31], before performing a frequency‐shift estimation by using cross‐correlation with delta functions at 2.01, 3.03, and 3.22 ppm. Subsequently, the frequency estimate was applied to the original data. The sum spectra were used for frequency estimation for the GABA‐edited MRSI data, and the frequency shift was then applied to all sub‐spectra.

The MRSI data (reconstructed from Philips .data/.list format in Osprey and exported as NIfTI‐MRS) and anatomical images can be found online [32].

2.4. Linear‐Combination Modeling (LCM)

Basis sets were simulated using the cloud‐based MATLAB toolbox “MRSCloud” derived from FID‐A [33, 34]. A generalized LCM algorithm recently implemented in Osprey was used for modeling [35]. A detailed description of the LCM algorithm modeling steps can be found in the supplementary materials.

For the short‐TE MRSI, 18 metabolites (ascorbate, aspartate, creatine, negative creatine methylene to account for water suppression/relaxation (−CrCH₂), GABA, glycerophosphocholine (GPC), glutathione (GSH), Gln, Glu, mI, lactate, NAA, NAAG, phosphocholine, phosphocreatine, phosphoethanolamine (PE), Scyllo (sI), and taurine) were included. For the macromolecules (MM) and lipids (Lip), six Gaussian parameterized macromolecular/lipid (MM₂₀, MM₃₁, MM₃₇, MM₃₈, MM₄₀, Lip₂₀) basis functions were included to parameterize all MM signals in the model range. The frequency and linewidth expectation and standard deviation values for MM₂₀ and Lip₂₀ were defined as described in the LCModel manual [36]. For MM₃₁, MM₃₇, MM₃₈, and MM₄₀ the parametrization was as previously described [37, 38]. The frequency and linewidth expectation and standard deviation values for those macromolecules were derived from the same studies by calculating the distributions across the whole cohort. A detailed description of the parametrization can be found in Data S1. Note that the amplitude soft constraints for MM₂₀ were defined with respect to tNAA instead of the commonly used MM₀₉ signal, which was affected by the HGDB pulses. The expectation values and standard deviations are derived from a previous in vivo short‐TE single‐voxel MRS study [38]. A detailed description of the soft constraints for the MMs can be found in Data S2. For the metabolite basis functions, amplitude soft constraints were defined as described in the LCModel manual [36].

For the GABA‐edited MRSI basis functions, 6 metabolites (GABA, GSH, Glu, Gln, NAA, and NAAG) were included, assuming perfect subtraction of the other metabolites like Cr and Cho. The co‐edited macromolecules at 3 ppm were parameterized by adding a Gaussian peak (2 proton amplitude, 14‐Hz FWHM) to the GABA basis function with a 1:1 amplitude ratio with the empirical assumption of 50% signal contribution of MM to the 3‐ppm signal [39, 40], effectively generating a GABA+ basis function [26]. An amplitude soft constraint (0.15/1.0 as defined in the LCModel manual [36]) for the NAAG to NAA ratios was applied.

2.5. Quantification

For the conventional spectra, tNAA, tCr, tCho, Glx, and mI were investigated while GABA+ was reported from the GABA‐edited difference spectra. The estimates were referenced to the water signal and multiplied by the concentration of pure water (55.5 M) without further correction for tissue relaxation times or water content to generate institutional units (i.u.). This approach was used to be consistent in the planned study in subjects with TSC, in which literature reference values for water content and relaxation times are not available, particularly in TSC‐related brain lesions. Additionally, creatine‐referenced estimates were calculated for all metabolites using the tCr estimates from the conventional spectra.

2.6. Regions of Interest

A total of 14 regions of interest were selected in the three MRSI slices based on the common occurrence of brain lesions in patients with TSC. The following regions were selected: frontal/posterior gray matter and left/right motor cortex in the superior slice, anterior/posterior cingulate cortex and left/right centrum semiovale in the middle slice, and corpus callosum genu/splenium, left/right insular cortex, and left/right thalamus in the inferior slice. All brain regions were identified in FSLeyes [41] using the anatomical MRI scan and the MRSI model results and were identified by an experienced clinician (D.L.). Spectra with obvious lipid or other artifacts were excluded from further analysis.

2.7. Quality Metrics

SNR and FWHM were used to assess the data quality of each region following the consensus‐recommended thresholds [13] and were calculated for the total NAA peak at 2.0 ppm for the short‐TE MRSI. FWHM was also assessed for the edit‐OFF spectra of the GABA‐edited MRSI. Additionally, CRLB estimates of each metabolite of interest, inspection of the residual, and overall model quality were used to assess the quality of the modeling process. Metabolite estimates where the CRLB exceeded the minimum CRLB across all subjects by a factor of 3 were excluded. Metabolite estimates that did not fall between 1% and 300% of the median estimate across all subjects were automatically identified as outliers. The exclusion was performed separately for creatine‐ and water‐referenced estimates of each metabolite. The methods are summarized according to the minimum reporting standards in MRS [42] in Data S3.

2.8. Effects of Motion Compensation

A secondary analysis of the effect of motion compensation on 2D multi‐slice GABA‐MRSI was performed by comparing regions of interest that passed the quality control with and without retrospective motion compensation.

2.9. Statistical Analysis

Descriptive statistics of all metabolite estimates and quality metrics were calculated across all eligible brain regions. Test–retest reproducibility was assessed with the intra‐ and inter‐subject CV for each metabolite estimate and reference method. ANOVA was employed to assess differences in contralateral brain regions and between sessions. One‐way ANOVA was employed to analyze differences between brain regions, followed by Bonferroni correction of the post hoc comparisons if significant effects of brain region or session were found. Additionally, a paired t‐test was employed to assess differences in the FWHM between both MRSI scans.

3. Results

All datasets were successfully processed and modeled. After visual inspection, two of the adult subjects were removed from further analysis due to poor overall data quality resulting from motion.

Figure 2 provides an overview of the spectral quality of the short‐TE and GABA‐edited MRSI scans after filtering the data according to the described quality criteria. Both visits and the mean spectrum across all subjects are visualized. After exclusion, the visual spectral quality was good for the short‐TE MRSI across all three slices. Fewer voxels were included for the GABA‐edited MRSI due to severe subtraction artifacts even after applying prospective motion compensation. Further visualization of the data quality can be found in Data S4, showing a broader frequency range to evaluate residual water and lipid signals for all subjects. Data S5 and S6 show example LCM results for the short‐TE and GABA MRSI from both visits of an example subject. The mean tNAA SNR values for the short‐TE MRSI for the superior slice ranged between 187 and 244, between 205 and 255 for the middle slice, and between 112 and 181 for the inferior slice. The mean tNAA FWHM values for the short‐TE MRSI for the superior slice ranged between 15.0 and 20.8 Hz, between 9.3 and 10.2 Hz for the middle slice, and between 10.9 and 18.8 Hz for the inferior slice. The mean tNAA FWHM values for the edit OFF spectra of the GABA MRSI for the superior slice ranged between 12.1 and 19.5 Hz, between 7.9 and 9.0 Hz for the middle slice, and between 10.7 and 18.3 Hz for the inferior slice. No statistically significant differences in the linewidth were found between the short‐TE and the GABA‐edited MRSI. Detailed values for each region of interest can be found in Data S3. The FWHM for both MRSI scans is visualized in Data S7.

FIGURE 2.

MRSI spectra for both the short‐TE and GABA‐edited MRSI from the 14 regions of interest after quality control. Regions of interest are indicated on T₁‐weighted brain MRI. Spectra from Visit 1 and Visit 2 are plotted in the top and bottom part of each panel. Individual spectra are shown as black lines, and the average spectra across all subjects are color‐coded for each region of interest.

Figure 3 shows the water‐referenced metabolite estimates of both visits. Water‐referenced metabolite estimates for short‐TE (Figure 3A) indicate good performance for the short‐TE MRSI in the middle and inferior slices and poorer performance in the superior slice. More metabolite estimates were excluded in regions that are susceptible to lipid artifacts (motor cortex) and usually exhibit poorer spectral quality (frontal regions and regions close to the sinus). The statistical analysis indicated that there were neither differences between the contralateral brain regions nor between the two sessions for any of the water‐referenced metabolite estimates.

FIGURE 3.

Water‐referenced metabolite estimates for the short‐TE (A) and GABA‐edited (B) MRSI for all regions of interest and metabolites from both visits. Color coding is consistent with the previous figures. Mean and standard deviations are shown in color with closed circles for Visit 1 and squares for Visit 2. The gray lines connect individual data points from each visit. The total number of included spectra is added in the top row. F, frontal; P, posterior; R, right; L, left; GM, gray matter; MC, motor cortex; ACC, anterior cingulate cortex; PCC, posterior cingulate cortex; CSO, centrum semiovale; genu, genu of corpus callosum; IC, internal capsule; tha, thalamus; splenium, splenium of corpus callosum.

The one‐way ANOVA indicated no differences in water‐referenced metabolite estimates between brain regions.

Figure 3B shows water‐referenced GABA+ estimates. As noted before, considerably fewer spectra were included, and the reproducibility is worse than for the short‐TE MRSI. Again, no statistically significant differences were found between contralateral brain regions and between sessions as well as for the between‐region one‐way ANOVA.

The relative CRLBs for each metabolite after exclusion are summarized in Data S8. Overall, relative CRLBs were < 15% across all metabolites and slices. This is mainly due to the comparably large nominal voxel size, which yielded high SNR spectra.

The success rates after applying the data quality criteria are summarized in Table 1. The overall success rate was higher for the short‐TE MRSI compared with the GABA‐edited MRSI for water‐referenced metabolite estimates. Separated by slice, the middle slice had the highest success rate. Applying retrospective motion correction more than doubled the success rate for the GABA‐edited MRSI when both visits were considered and resulted in at least a 20‐percentage point increase for at least one successful visit.

TABLE 1.

Success rates for the short‐TE and GABA‐edited MRSI scans after applying the quality criteria. The combined success rates for all slices, as well as those separated by slice, are reported. For the GABA‐edited MRSI scans, the success rates with and without retrospective motion compensation are reported.

		Superior slice	Middle slice	Inferior slice	All slices
Water‐referenced	Short‐TE
	Both visits	54%	93%	82%	77%
	At least one visit	88%	100%	94%	96%
	GABA‐edited
	Motion compensation
	Both visits	46%	67%	39%	46%
	At least one visit	77%	97%	77%	83%
	No motion compensation
	Both visits	14%	25%	15%	18%
	At least one visit	61%	73%	58%	63%
Creatine‐referenced	Short‐TE
	Both visits	53%	95%	85%	77%
	At least one visit	87%	100%	95%	96%
	GABA‐edited
	Motion compensation
	Both visits	33%	72%	37%	46%
	At least one visit	72%	97%	79%	82%
	No motion compensation
	Both visits	7%	39%	20%	23%
	At least one visit	61%	73%	58%	63%

The creatine‐referenced metabolite estimates are shown in Figure 4. The reproducibility of the metabolite estimates is similar to the water‐referenced estimates. Statistical results were also comparable with the water‐referenced estimates; no statistically significant differences were found between contralateral brain regions or between visits, as well as between brain regions in the one‐way ANOVA analysis.

FIGURE 4.

Creatine‐referenced metabolite estimates for the short‐TE (A) and GABA‐edited (B) MRSI for all regions of interest and metabolites from both visits. Color coding is consistent with the previous figures. Mean and standard deviations are shown in color with closed circles for Visit 1 and squares for Visit 2. The gray lines connect individual data points from each visit. The total number of included spectra is indicated in the top row of each panel.

The overall success rate showed similar behavior for the creatine‐referenced metabolite estimates as for the water‐referenced estimates (see Table 1).

Interestingly, we found significantly higher water‐referenced Glx estimates for the youngest participant (age 7 years) compared with three older subjects (ages 22 to 31 years) with p < 0.01 after Bonferroni correction. For the second child (age 11 years), significantly higher water‐referenced Glx estimates were found than in the same 22‐year‐old subject, with p < 0.05 after Bonferroni correction. Note that a post hoc comparison was performed for the subject factor, including both visits and all brain regions at once, indicating a global increase of Glx for the younger subjects. For creatine‐referenced estimates, significantly higher Glx estimates were only found for the youngest subjects (age 7 years) compared with one older subject (age 31 years), with p < 0.05 after Bonferroni correction. Again, the post hoc comparison was performed on the subject factor, indicating global changes in Glx.

Figure 5 summarizes the intra‐subject test–retest variability for both referencing methods. As indicated in the previous figures, the middle slice shows the lowest intra‐subject test–retest CVs (on average 5.8% across all metabolites) for water‐referenced estimates (Figure 5A) and creatine‐referenced estimates (4.8% across all metabolites Figure 5C). Similarly, for GABA‐edited MRSI the intra‐subject test–retest CVs of the middle slice were lowest with 13.5% for both water‐referenced (Figure 5B) and creatine‐referenced estimates (Figure 5D).

FIGURE 5.

Intra‐subject coefficient of variation estimates for all regions of interest. Water‐referenced estimates of the short‐TE (A) and GABA‐edited MRSI (B), as well as creatine‐referenced estimates of the short‐TE (C) and GABA‐edited MRSI (D). The total number of included spectra is indicated in the top row of each panel.

Figure 6 summarizes the inter‐subject test–retest variability for both referencing methods. The middle slice had the lowest inter‐subject test–retest CVs with an average of 11.1% across all metabolites for water‐referenced estimates (Figure 6A) and 9.7% across all metabolites for creatine‐referenced estimates (Figure 6C). Similarly, for GABA‐edited MRSI the inter‐subject test–retest CVs of the middle slice were lowest for water‐referenced estimates (Figure 6B) with 13.5% and 16.9% for creatine‐referenced estimates (Figure 6D).

FIGURE 6.

Inter‐subject coefficient of variation estimates for all regions of interest. Water‐referenced estimates of the short‐TE (A) and GABA‐edited MRSI (B), as well as creatine‐referenced estimates of the short‐TE (C) and GABA‐edited MRSI (D). The total number of included spectra is indicated in the top row of each panel.

4. Discussion

In this study, multi‐slice 2D short‐TE and GABA‐edited spin‐echo MRSI at 3 T were performed in healthy subjects at two timepoints between 7 and 14 days apart. Retrospective motion compensation and expert‐consensus linear‐combination modeling were employed during the analysis. The main aim of the study was to investigate the test–retest reproducibility of the protocol MRSI for subsequent application in TSC. To the authors' best knowledge, this is the largest publicly available test–retest MRSI study at 3T, offering an interesting resource for further development of different processing and analysis methods.

Several previous studies assess the test–retest repeatability of short‐TE MRSI at 3T [19, 20, 21, 22, 23, 24]; these cover a wide range of sequence implementations, ranging from fast EPSI sequences with full brain coverage to high‐resolution sLASER‐MRSI in the basal ganglia region. However, these methods are not readily available on most MR scanners, limiting their clinical utility and generalizability. In contrast, the short‐TE and edited sequences used in this study are derived from the Philips product multi‐slice 2D‐MRSI sequence, with the addition of the HGDB lipid‐ and water‐suppression pulse, and editing pulses.

Similar to previous studies [25], we found no session effect or differences when comparing the regions from both brain hemispheres in the statistical analysis, indicating highly reproducible metabolite estimates for both MRSI acquisitions. In contrast to previous reports showing differences in cortical and subcortical GABA+ estimates for GABA‐edited MRSI [25], and a gray‐to‐white matter contrast for several metabolite estimates for short‐TE MRSI [19, 20, 43], no statistically significant differences between cortical and subcortical GABA+ estimates or between gray and white matter were found in our study. This indicates relatively homogeneous distributions of metabolites across the included MRSI volume in our study. Possible reasons for this are the nominal voxel size, which was comparably large to achieve sufficient SNR for the GABA‐edited MRSI and matched between both MRSI acquisitions, and the limited number of included regions of interest and subjects.

Globally increased Glx values (both water‐ and creatine‐referenced) were found for the 7‐year‐old subject compared with older subjects, and also increased water‐referenced Glx for the 11‐year‐old subject. These findings are in line with previous literature reporting age‐related decreases of creatine‐referenced glutamate estimates from adolescence to adulthood using a J‐refocused MRSI sequence at 7T [44], which the authors attribute to underlying brain development.

4.1. Intra‐Subject Variability

The test–retest CVs for the repeated scans in the nine healthy volunteers included in the final analysis were on average 5.8% across all metabolites for water‐referenced estimates and 4.8% for creatine‐referenced estimates for the short‐TE MRSI, suggesting high reproducibility. However, test–retest CVs increased in the superior slice and in frontal regions, mainly due to increased linewidth due to B₀‐inhomogeneities and lipid contamination from the scalp. The results align well with previously reported MRSI reproducibility studies, which reported test–retest CVs ranging from 5% to 30% [19, 20, 21, 22, 23, 24]. These values are comparable with those of SVS reproducibility studies, indicating high data and modeling quality [24, 45, 46, 47].

For the GABA+ estimates from the GABA‐edited MRSI, poorer reproducibility has been found with intra‐subject CVs of 13.5% for both water‐referenced and creatine‐referenced estimates. A previous 3D GABA‐edited MRSI study using prospective motion correction with shim updates reported a median test–retest CV of 8% for GABA+ [25]. Although the CV values are comparable, about 50% of our datasets had to be excluded during the analysis; note that this is also because two successful acquisitions are required for the test–retest analysis. The most likely explanation is that the retrospective motion compensation scheme employed here is not able to compensate for more than minimal and intermittent head motion and that other approaches (including prospective updates of shim and slice‐selection parameters) are likely necessary for robust, multi‐slice edited MRSI acquisitions. Combining both prospective and retrospective motion correction should be considered to achieve the highest possible success rate.

4.2. Inter‐Subject Variability

The inter‐subject variability measured in this study was higher than the intra‐subject variability. A possible interpretation of this is that variability introduced by the MRI itself (scanner drift stability, gradient coil heating, and B₀‐inhomogeneities), practical factors (positioning of the subject in the scanner, reproducibility in placing the MRSI volume), and individual factors (subject motion, individual variability of metabolite concentrations between sessions) is comparably low for our specific study design. Still, the reported values of 11.1% for water‐referenced estimates and 9.7% for creatine‐referenced estimates for the short‐TE MRSI are comparable with previous studies [48]. For the GABA‐edited MRSI, CVs of 13.5% and 16.9% for water‐ and creatine‐referenced estimates were found, respectively. These values are slightly higher than the previously reported inter‐subject CVs (~8% for GABA+/tCr for a 3D GABA‐edited MRSI) [25], likely again due to improved motion robustness achieved by the prospective motion correction in the prior study. However, our study has increased coverage and is not limited to the basal ganglia region.

The overall findings of this study reflect recommendations of the advanced MRSI consensus paper with highly reproducible metabolites with extended coverage for the metabolites with major singlets (tNAA, tCho, and tCr) as well as mI and Glx¹.

5. Limitations

This study used retrospective motion compensation to account for possible subject motion during the multi‐slice 2D GABA‐edited MRSI acquisition (22 m 21 s). While this approach has shown effectiveness in accounting for intermittent motion in single‐slice 2D GABA‐edited MRSI, it cannot account for gross subject motion, leading to significant data loss after quality control. The data quality and success rate for the multi‐slice GABA‐MRSI have also improved after retrospective motion compensation, but the overall success rate for two consecutive acquisitions is only 46%. This problem could be addressed by improving the retrospective motion‐compensation algorithm. For example, the automated motion outlier identification could be improved by using data‐driven metrics instead of hard thresholding. Alternatively, one might apply a more conservative outlier removal in combination with low‐rank data reconstruction approaches [49]. Additionally, prospective motion correction is effective for edited MRSI [25]; however, it requires substantial modification of the MRSI sequence.

Although the HGDB pulses effectively suppressed unwanted lipid and water signals, they also suppressed the MM₀₉ peak commonly used as a non‐overlapped reference peak to constrain the amplitudes of the other macromolecular signals during LCM. To account for this, the model range was limited between 1.95 and 4 ppm, and the MMs were constrained relative to tNAA during modeling the short‐TE MRSI. However, recently developed lipid suppression and removal approaches [43, 50] could offer an effective way of removing lipid contamination without affecting the MM₀₉ peak. The HGDB pulses use both frequency‐selective lipid (and water) suppression pulses, as well as spatial OVS pulses, which suppress signal from peri‐cranial tissue [28, 29]. Because the cutoff frequency for the lipid suppression is 1.8 ppm, lipid signals resonating above 1.8 ppm are not suppressed: These may include methylene protons α to C=C double bonds (2.03 ppm) and COO groups (2.25 ppm) [51], which may also complicate spectral fitting, as it usually relies on amplitude soft constraints between the non‐overlapped lipid at 1.3 ppm to regularize the lipid signals at 2.03 and 2.25 ppm [36]. For the GABA‐edited MRSI, instead of using the MM₀₉ peak to constrain MM₃₀, this was addressed by adding a 2‐proton MM peak to the GABA basis function; therefore, the HGDB pulses had no impact on the modeling of those datasets.

6. Conclusion

In conclusion, this study shows that fast and reproducible metabolite mapping of five major brain metabolites is possible with 2D multi‐slice MRSI sequences at 3T for application to assess multi‐focal brain disease. Mapping of GABA+ with the same spatial coverage remains challenging, with data acceptance of 82.5% for at least one visit and acceptable reproducibility after conservative quality control in 46% of the voxels of interest.

Author Contributions

Helge J. Zöllner: data curation, software, formal analysis, methodology, investigation, writing the original draft, writing review and editing, and visualization. Dillip K. Senapati: investigation and writing review and editing. İpek Özdemir: investigation and writing review and editing. Kimberly L. Chan: software, investigation, and writing review and editing. Georg Oeltzschner: software, writing review and editing, supervision, and funding acquisition. Doris D. M. Lin: writing review and editing, supervision, and funding acquisition. Peter B. Barker: conceptualization, methodology, investigation, writing review and editing, supervision, project administration, and funding acquisition.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1 Properties of the Gaussian functions of the broad macromolecule and lipid resonances included in the basis sets, MM20 and Lip20 are taken from section 11.7 of the LCModel manual while the other peaks are parameterized according to Hui et al., MRM, 2021. The amplitude values are scaled relative to the CH₃ singlet of creatine with amplitude 3.

Data S2 Amplitude soft constraints for the MM basis functions defined relative to the total NAA amplitude (NAA + NAAG) included in the short‐TE MRSI data. The expectation values and standard deviations are derived from Zöllner et al., NMR Biomed, 2023.

Data S3 MRSinMRS summary generated in Osprey

Zöllner H., Senapati D., Özdemir İ., et al., “Reproducibility of Metabolic Mapping Using 2D Multi‐Slice Short‐TE and GABA‐Edited Spin‐Echo MRSI at 3T in a Single Protocol,” NMR in Biomedicine 39, no. 1 (2026): e70175, 10.1002/nbm.70175.

Data Availability Statement

The data that support the findings of this study are openly available in the NITRC repository 3T test‐retest 2D MRSI online (https://www.nitrc.org/projects/repro_2d_mrsi/).