Inter‐subject stability and regional concentration estimates of 3D‐FID‐MRSI in the human brain at 7 T

Recently, a 3D‐concentric ring trajectory (CRT)‐based free induction decay (FID)‐MRSI sequence was introduced for fast high‐resolution metabolic imaging at 7 T. This technique provides metabolic ratio maps of almost the entire brain within clinically feasible scan times, but its robustness has not yet been thoroughly investigated. Therefore, we have assessed quantitative concentration estimates and their variability in healthy volunteers using this approach.

Discussion: We defined the performance of 3D-CRT-based FID-MRSI for metabolite concentration estimate mapping, showing which metabolites could be robustly quantified in which ROIs with which inter-subject CVs expected. However, the basal brain regions and lesser-signal metabolites in particular remain as a challenge due susceptibility effects from the proximity to nasal and auditory cavities. Further improvement in quantification and the mitigation of B 0 /B 1 -field inhomogeneities will be necessary to achieve reliable whole-brain coverage.
K E Y W O R D S 7 T, healthy brain, high resolution, inter-subject reproducibility, MRS, MRSI

| INTRODUCTION
Proton magnetic resonance spectroscopic imaging (MRSI) in the brain based on the direct acquisition of the free induction decay (FID) signal has been introduced to overcome many technical challenges with ultra-high-field systems. 1,2 Such challenges include lower B 0 /B 1 + homogeneity, more restricted specific absorption rate (SAR), and shorter T 2 times. 3 This simple acquisition scheme further reduces SAR, eliminates signal loss due to T 2 relaxation and J coupling, improves spatial selection, and enables short repetition times (T R ). 1,2 Over the last several years, technical improvements have concentrated on faster data acquisition to reach shorter measurement times, increased brain coverage, and higher spatial resolutions, which are attractive for clinical metabolic brain mapping. 4 Starting with parallel imaging techniques, [5][6][7][8] echo-planar spectroscopic imaging (EPSI), 9 and T R reduction, [10][11][12] these recent innovations have concentrated on spatial-spectral encoding using spiral, 13,14 rosette, 15 and concentric ring trajectories (CRTs). [16][17][18][19][20] Research culminated in a CRT-based 3D-MRSI sequence at 7 T that can cover the whole brain with $3 mm isotropic resolution, acquired in 10-15 min. 21,22 7 T MRSI, with an increased signal-to-noise ratio (SNR) and spectral dispersion, compared with lower-field MR scanners, enables imaging of a wide range of metabolites, eg improving the separation of N-acetylaspartate (NAA) from N-acetylaspartylglutamate (NAAG), glutamate (Glu) from glutamine (Gln), or glycine (Gly) from myo-inositol (mIns). [22][23][24] Based on these benefits, 7 T MRSI has been successfully applied to research applications ranging from γ-aminobutyric acid (GABA) mapping 25,26 to the resolution of metabolism in tumors, 9,[22][23][24]27 multiple sclerosis, 28 and epilepsy. 29 While many technical milestones have been reached, open questions remain with respect to the stability of these whole-brain and/or highresolution MRSI methods. This includes inter-subject variations of metabolite concentrations in different brain regions, as well as intra-subject variations over time. So far, we had only investigated these variations for FID-MRSI in a small study at 3 T. 30 Historically, the results of MRSI quantification have varied greatly depending on study design, data evaluation, and investigated brain regions. [31][32][33][34][35][36][37] To facilitate the use of MRSI for clinical and research applications, we have to first establish the normal concentrations of metabolites within different brain regions of healthy volunteers instead of relative signals as previously.
To evaluate the performance and inter-subject stability of our 7 T 3D-CRT-FID-MRSI approach, we conducted a study with a larger subject cohort, detailed quantification estimation and regional evaluation.

| Purpose
The purpose was to acquire whole-brain, 3D-CRT-based FID-MRSI at 7 T in a wider cohort (24 volunteers) than in our previous volunteer studies and to derive for the first time concentration estimates for our method, and further to assess the local MRSI data quality in an array of different small and large brain regions in order to evaluate the quantification robustness and inter-subject variability for the concentration estimates of individual metabolites.
The results will define the performance limits of our 3D-CRT-based FID-MRSI method at 7 T in regard to which metabolites can be confidently and reliably mapped in which regions of the brain.

| Subject recruitment
This study was conducted with the approval of the local institutional review board. Subjects were included in this study when no contraindications for 7 T MRI (eg claustrophobia, ferromagnetic implants, non-ferromagnetic metal head implants >12 mm, or pregnancy) were reported. Written and informed consent was obtained from all 24 young healthy volunteers (12 male, 12 female, mean age 27 ± 6 years, Table 1). We chose a young cohort due to expected good compliance for motionlessness and easy reproducibility.

| 7 T MRSI measurement protocol
We performed the measurement protocol using a 7 T whole-body MR imager (Magnetom, Siemens Healthineers, Erlangen, Germany), readout duration/spectral bandwidth of 158 ms/606 Hz, and an Ernst angle of 27 , but with otherwise identical spatial coverage, and was acquired in 3 min 18 s as an internal water reference. This second scan was necessary as we required an unsuppressed water signal as reference.

| Data processing and quantification
For offline MRSI processing, we utilized our in-house-developed software pipeline 39  prior studies with metabolite-nulled measurements was included to improve quantification. 47,48 Water was quantified separately from the unsuppressed reference scan, using LCModel as well, with a water basis simulated as above. SNR (using the pseudo-replica method with receiver noise prescans acquired at the start of the MRSI sequence 6 ) and full width at half maximum (FWHM) were calculated voxel-wise from the LCModel fits of NAA and tCr at 3.02 ppm.
For the calculation of concentration estimates using the internal water concentration as a reference,  estimates for each spectrum were then calculated as the ratio of metabolite-to-water signal multiplied by the local water concentration and the correction factor calculated before influenced by the respective GM/WM/CSF fraction as well as T 1 relaxation.
We created 3D maps of concentration estimates for all metabolites and filtered these based on a spectral quality mask, motivated by recent consensus recommendations, 54 which excluded voxels with at least one of these parameter restrictions: tCr SNR < 5, tCr FWHM > 0.15 ppm, metabolite Cramér-Rao lower bound (CRLB) > 40%, and metabolite fit value > 13 median absolute deviations. For display purposes, these maps were interpolated tri-linearly to fourfold resolution in MINC's register tool.
FreeSurfer (6.0, Laboratory for Computational Neuroimaging at the Athinoula A. Martinos Center for Biomedical Imaging, Boston, MA) 55,56 was used for automated segmentation of structural images based on cortical and subcortical brain atlases. 56,57 In-house MATLAB codes were used for mask extraction within each region of interest (ROI). We defined 55 ROIs, as shown in Table 3, including small and large structures/cortices, GM and WM regions separated and merged, and ROIs per hemisphere, in order to investigate which brain regions and ROI sizes could be reliably imaged. Metabolic maps were interpolated to the 0.8 mm resolution of structural images using nearestneighbor interpolation. Interpolated maps were overlaid with derived masks and mean metabolite concentrations were calculated within each ROI. 58 2.4 | Data evaluation

| General overview of measurement quality
Metabolite maps of all subjects were controlled by a reader (G.H.) for the presence of lipid and movement artifacts. Quantification of all metabolites listed in the basis set was evaluated and metabolites that were not fit in at least 10% of brain voxels were discarded from further evaluation. Cysteine was discarded as well due to general doubts about its quantification. We compared concentration estimate maps with uncorrected metabolite maps for differences in data quality and contrast. Representative spectra and metabolite maps were selected for display.

| Quantification quality within ROIs
Regions Region-specific data quality was assessed by calculating the percentage of voxels within an ROI that had CRLBs less than 40% for all of NAA, tCr, tCho, and mIns. ROIs with more than 80% of voxels above that threshold were defined as good and those with 66-79% as acceptable, and those with less than 60% were rejected. Rejected ROIs were excluded from further evaluation. To also quantify the performance of individual metabolite fitting, the percentage of voxels with CRLBs less than 20% for NAA, tCr, tCho, Ins, and CRLBs less than 40% for all others were determined for every ROI.

| Metabolites
Only metabolites that were fit in more than 66% of voxels (mean of all regions) were considered as qualified for the main analysis, but the remaining ones are included in the Supporting Information (Supporting Tables 3 and 4).

| Quantification estimates
For all metabolites in all qualified ROIs, regional means per subject and inter-subject mean of means for all subjects not excluded in Section 2.4.1 were calculated. The range of observed ROI concentration estimates was compared with literature values. To facilitate the comparison with other studies that used ratios to tCr, we additionally calculated metabolite ratios to tCr.

| Inter-subject coefficients of variation
As a measure for the expected variability in metabolite estimates based on physiologic differences among subjects and the stability of our MRSI method, inter-subject coefficients of variation (CVs) of mean ROI-specific concentration estimates and ratios to tCr were calculated and compared for all qualified ROIs and metabolites based on the mean concentration estimates/ratios per subject for every ROI. This mean and its standard deviation were then used for CV calculation per ROI over all subjects.

| General overview of measurement quality
One dataset (Volunteer 1) was impaired by strong movement artifacts and had to be excluded from further analysis. In five subjects, some ROIs had to be excluded (Table 1), as the GM/WM classification based on T 1 -weighted imaging had failed in these ROIs. In these ROIs, mean concentration estimates were calculated for the remaining subjects. Total Cr SNR and FWHM over all volunteer brain voxels within the quality mask were 11 ± 5 and 0.06 ± 0.02 ppm. Asp and sIns were the only metabolites that were completely excluded from further analysis. Generally, fitting was of good quality, as illustrated by sample spectra in Figure 1 and Supporting Figure 2. Comparison of metabolite maps before T 1 correction and water referencing with concentration estimate maps showed that the concentration estimate maps showed a slight reduction of inhomogeneities and fewer outliers at the brain periphery ( Figure 2). The latter was related to the inclusion of GM/WM segmentation, which removed CSFdominant voxels from the maps. As our extensive 3D metabolite maps cannot satisfactorily be displayed with a limited number of figures, we supplied multiple complete datasets for review (see Section 4.5).

| Regions
Of the 55 segmented regions, 18 fulfill the criteria for "good," 26 for "acceptable," and 11 were rejected, as detailed in Table 3, including the mean ROI sizes. Concentration estimate standard deviations are generally higher for smaller regions. The rejected regions were mostly situated in F I G U R E 1 Sample spectra for pure GM and WM voxels in Volunteer 11. The spectra were first-order phased for viewing convenience. GM/WM differences are especially visible for Glu. Spectral phasing is due to the 1.3 ms acquisition delay.
In these examples, only small residual lipid signals remain visible the lower parts of the brain, were small, or were close to the nasal cavities/eyes. Of all metabolites, tCho, tCr, Glu, mIns, and NAA fulfilled the qualification criterion of being fit in more than 66% of voxels (Table 3). An overview over tCr SNR, FWHM, and metabolite CRLBs in relation to the resulting concentration estimate maps is given in Supporting Figure 3.

| Metabolites
Of all cortices, the parietal, motor, and cingulate cortices performed the best.

| Quantification estimates
Our concentration estimates for the five qualified metabolites over 44 ROIs are presented in Table 4 and graphically summarized in Figure 4.

| Inter-subject coefficients of variation
The inter-subject CVs in Table 6 show a good comparability of the regional analyses between subjects for most cases. These CVs were the lowest for tCr and NAA and, generally, in the range of 10-20%. In the majority of "good" ROIs, tCho, tCr, Glu, mIns, and NAA CVs were 10% or less. Table 4 and Supporting  Figure 3 illustrates these findings over multiple subjects. The CVs for ratios to tCr (Supporting Table 2) were very similar, with differences between the means over all ROIs not exceeding 4%.

| DISCUSSION
We have assessed the average and region-specific concentration estimates, as well as their inter-subject variability, for five neuro-metabolites that can be reliably mapped in the human brain using our 3D-CRT-based FID-MRSI sequence at 7 T. To improve upon previous work, 21 we have expanded the subject cohort, also evaluated less abundant metabolites, and assessed concentration estimates instead of ratios and data quality over a large number of automatically segmented brain ROIs to provide a more detailed understanding of the performance of 7 T 3D-CRT-based FID-MRSI. The additional inclusion of T 1 corrections and internal water referencing allowed a more quantitative assessment (ie of concentration estimates). In summary, we found our method to yield acceptable results in 23 of 24 volunteer subjects for five metabolites in 44 predefined ROIs.
Our concentration estimates are in the range of previous reports, while inter-subject CVs indicated a good level of stability in many of these metabolites and ROIs. Still, the quantification quality of GABA, Gln, Gly, GSH, NAAG, Ser, and Tau in healthy subjects cannot be considered sufficient for this MRSI application, necessitating different approaches or methodological improvements if these are required. These results are important to define the limits of stability, sensitivity, and regional reliability for our MRSI method for future research applications. As the sum of our generated imaging data is hard to convey within few figures, we invite the readers to look at the supplementary full datasets provided by us on Zenodo.
To contextualize our resultant metabolite distributions, we conducted an extensive comparison with previous research. Concentration estimates for many metabolites in many brain regions have been reported over the last decades, sometimes with contradicting results due to different processing methods, subject cohorts, partial volume effects, acquisition schemes (eg MRSI or single-voxel spectroscopy, SVS), and scanner types.
Different quantification algorithms are known to affect reported results. 60 Dhamala et al 37 Table 5, and mostly found in the middle or lower range of references. For metabolites separable in our study, such as Glu/Gln and NAA/NAAG, they individually agree with previous literature (Tables 5 and7), and further comparing their sums agrees well with studies that cannot separate them.
Going from overall reported values to specific ROIs, as compared in Table 7 34 and for the parietal cortex as well as the thalamus with Lecocq et al. 35 For metabolite ratios to tCr, we also compared our results with the 7 T MRSI results of Bhogal et al, 70 as seen in Table 8. Over the six compared ROIs, our ratios are consistently higher for all except GSH/tCr, which is mixed. The effect is most pronounced for Ins + Gly. The generally higher ratios could be sourced in lower quantification estimates of tCr or differences in the MRSI acquisition. Considering the multitude of applied methods and overall limited sample sizes, our results are in general agreement with the current state of knowledge in the field but are for the first time based on concentration estimation for high-resolution 7 T FID-MRSI. We see a further need to investigate metabolites such as GSH, Ser, and Tau, which are difficult to quantify and for which MRS-based concentration estimates are scarce.
Our inter-subject CVs were the smallest for metabolites with the highest SNR (ie, NAA, tCr, tCho, Glu, and mIns are in the <10% range) but approached 30% for other metabolites in some ROIs (see Table 6). Comparison with the literature is difficult, as most studies report intra-subject T A B L E 8 Comparison of this study's metabolite ratios and standard deviations of tCho, Glu, Glu + Gln, NAA + NAAG, Ins + Gly, and GSH to tCr in six ROIs compared with similar 7 T MRSI results of Reference. 70 The results of our study are consistently higher for all except GSH/tCr, with two higher/lower/same ratio regions each these results are similar to our findings, but our 7 T MRSI featured a higher resolution and more quantifiable metabolites and was also affected by additional physiologic inter-subject variation. Considering that the CV calculation did not account for diurnal effects, 36 or age [75][76][77] and sex differences, 78,79 these results seem convincing but still include methodological artifacts such as subject motion. Intra-subject CVs from a testretest study will be necessary to complete the picture. In another MRSI study at 3 T, we found that the application of motion correction improved the CVs of metabolite ratios to tCr by 30%. 30 Another source of local variability would be the combination of T 1 weighting with our short T R and a lack of knowledge of local tissue metabolite T 1 values. 52,80 Although we tried to correct our T 1 estimates and reference water concentrations for voxel-wise GM/WM fractions, we assume that additional variation exists based on this. 53 More precise concentration estimates could be obtained via the direct mapping of tissue water content. 81 Due to our echo-less acquisition approach with negligible acquisition delay, our results can be considered to be robust to T 2 effects. This is a potential advantage in the study of the aging brain or of pathologies that can cause local iron deposits, which affect metabolite and water T 2 values. 82

| Limitations
Our study has some limitations. T 1 values for multiple metabolites had to be estimated without previous reports. The assumption of a single metabolite T 1 over the whole brain is also a rough approximation. This could have biased the estimated concentrations.
Quantification is further limited by the precision of the internal water referencing, which relies on the assumption of tissue water relaxation times and concentrations, effective local excitation, and water signal quantification. Beyond the general variability of MRS quantification results based on the fit model parameters, L2 regularization is also known to influence metabolite signal estimation, especially NAA, 83 even if just by removing lipid signals. The comparison of quantification estimates between different field strengths, acquisition schemes, processing pipelines, resolutions, and brain segmentations limits the comparability to the even subset of MRS studies that report concentration estimates.
Our study is limited to the analysis of inter-subject variations and does not, therefore, report on intra-subject variations. Measuring intrasubject variation could also help to separate the more subject-specific (eg physiological) from the method-specific contributions to variability.
The exclusion of 11 ROIs, predominantly basal brain regions such as brain stem or cerebellum, from further analysis was necessary due to the lack of spectral fit confidence, limiting the mappable brain coverage. This shows that B 0 -and B 1 -field inhomogeneities (the first caused by the proximity to the nasal and auditory cavities, the second by the limitations of single channel transmit coils at 7 T) remain a significant ultra-highfield challenge, which will have to be resolved via hardware improvements [84][85][86] and/or further improved high-resolution approaches. 87 Filtering of outliers based on a set of rigid criteria is insufficient. In the future, automated quality assessment of voxels based on deep learning will be necessary to evaluate datasets of this scale adequately. The necessary interpolation between MRSI and reference imaging combined with MRSI partial volume effects, even at our high resolution, are further confounding factors for the regional analysis as relevant GM/WM fractions remain within the ROIs (Table 3), reducing the expected GM/WM contrast, eg for Glu. While we obtained results for difficult-to quantify metabolites in healthy tissue (ie GABA, Gln, Gly, GSH, NAAG, Ser, and Tau), the percentage of voxels within the quality criteria remained overall low (Table 1).
Subject motion is another factor not yet accounted for in our study, and significant improvements are expected in stability, 88,89 which is of particular importance for studies in children and elderly patients. 26 In particular, real-time correction has been shown to significantly enhance data quality in high-resolution 3D-CRT-based FID-MRSI at 3 T. 30

| Conclusions/outlook
We have stablished the brain region-specific concentration estimates and their variability in a large number of healthy young volunteers for our whole-brain MRSI-based metabolic maps at 7 T for the first time. While not all brain ROIs performed well enough to be considered, especially basal regions such as the cerebellum, we could successfully quantify five metabolites-tCho, tCr, NAA, Glu, and Ins-in 44 ROIs, with all others not being quantified with a sufficient quality in a substantial number of voxels. Our estimated concentrations are consistent with previous research.
This was a necessary first step to define the reliability and to guide future basic and clinical research of the brain metabolism and to show the capability of our high-resolution 3D-MRSI technique in the current discussion of MRSI standardization. 52,90,91 Our results can guide future study planning, targeting specific brain regions and metabolites of interest on the one hand and the eventual development of a 7 T-MRSI based metabolic brain atlas on the other. The next avenues of research could be the investigation of intra-subject variability, improved quantification by direct water concentration mapping, 81 also in pathologies, and better definition of specific relaxation times. With these improvements, the metabolites that currently lack reliability could be reevaluated and an in-depth study of sex and age differences could be carried out. In summary, we have shown new insights into the expectable results and stability of our fast high-resolution MRSI at 7 T, but this approach still requires more sophistication.