Feasibility and Reproducibility of Neurochemical Profile Quantification in the Human Hippocampus at 3T

summary

Hippocampal dysfunction is known to be associated with several neurological and neuropsychiatric disorders such as Alzheimer's disease, epilepsy, schizophrenia and depression, therefore there has been significant clinical interest to study hippocampal neurochemistry. However the hippocampus is a challenging region to study using ¹H MRS, hence the use of MRS for clinical research in this region has been limited. Therefore, our goal was to investigate the feasibility of obtaining high quality hippocampal spectra that allow reliable quantification of a neurochemical profile and to establish inter-session reproducibility of hippocampal MRS, including reproducibility of voxel placement, spectral quality and neurochemical concentrations. Ten healthy volunteers were scanned in two consequent sessions using a standard clinical 3T MR scanner. Neurochemical profiles were obtained with a short-echo (TE=28ms) semi-LASER localization sequence from a relatively small (~4mL) voxel that covered ~62% of the hippocampal volume as calculated from segmentation of T₁-weighted images. Voxel composition was highly reproducible between sessions, with test-retest coefficients-of-variance (CV) of 3.5% and 7.5% for gray and white matter volume fraction, respectively. Excellent signal-to-noise ratio (~54 based on the N-acetylaspartate (NAA)-methyl peak in non-apodized spectra) and linewidths (~9 Hz for water) were achieved reproducibly in all subjects. The spectral quality allowed quantification of NAA, total choline, total creatine, myo-inositol and glutamate with high scan-rescan reproducibility (CV ≤ 6%) and quantification precision (Cramér-Rao lower bounds, CRLB < 9%). Four other metabolites, including glutathione and glucose, were quantified with scan-rescan CV below 20%. Therefore, the highly optimized, short echo semi-LASER sequence together with FASTMAP shimming substantially improved the reproducibility and number of quantifiable metabolites relative to prior reports. In addition, the between-session variation in metabolite concentrations, as well as CRLB were lower than between-subject variation of the concentrations for most metabolites indicating that the method has the sensitivity to detect inter-individual differences in the healthy brain.

Graphical Abstract

INTRODUCTION

The hippocampus, a brain structure located deep in the temporal lobe, plays an essential role in learning, memory formation (1) and stress regulation (2). Hippocampus dysfunction is thought to be involved in several neurological, psychiatric and metabolic diseases, including Alzheimer's disease (3), temporal lobe epilepsy (4), schizophrenia (5), and diabetes (6). Importantly, hippocampal function can be affected both by pharmacological and non-pharmacological treatments, such as exercise training (7,8).

Therefore there is great interest in studying the hippocampus using proton magnetic resonance spectroscopy (¹H MRS) to elucidate disease-related neurochemical alterations and their reversal with therapies. As a complementary modality to conventional structural MRI, MRS can be used to investigate changes in cell density, cell type or biochemical composition of neural tissue. Thereby spectroscopy can differentiate pathologies indistinguishable by structural MRI (9) and is sensitive to early cellular changes in central nervous system (CNS) diseases. For instance, MRS may help guide critical presurgical decisions in pharmacologically intractable epilepsy (10), and provide surrogate markers to monitor disease progression or treatment effects in trials (11). Despite this potential, the role of hippocampal spectroscopy has been limited in clinical research due to challenges associated with obtaining consistently high quality MRS data from this region.

Prior ¹H MRS studies reported hippocampal concentrations of 3-5 neurochemicals only (3,12), even at 3T with the short echo approach (13). Several reports of hippocampal MRS presented quantification results only as ratios to total creatine (tCr) (14,15). Furthermore, scan-rescan reproducibility (16,17) was shown to be lower than in other brain regions at 1.5T (16,18,19). The complicating factors for obtaining high quality MRS data from the hippocampus include the small size of the structure, which necessitates a small voxel to minimize partial volume effects, thereby compromising the signal to noise ratio (SNR), and large susceptibility effects that lead to broad linewidths and loss of spectral resolution (12). In addition, small inconsistencies in voxel placement may lead to large variability in spectral quality between scanning sessions because of the vicinity of the structure to a region with large susceptibility changes (20).

The hippocampus has also been investigated by spectroscopic imaging (MRSI) (21,22). With the higher spatial resolution achievable with MRSI, distinct neurochemical profiles have been demonstrated recently at 7T for the anterior and posterior regions of the hippocampus (21). These studies still only reported metabolite ratios (total N-acetylaspartate-to-total creatine, tNAA/tCr and tNAA-to-total choline, tNAA/tCho) (21) or absolute concentrations of 3 metabolites (tNAA, myo-inositol, glutamate) (22). Therefore, complementary use of single voxel MRS (SVS) and MRSI is expected to provide the most comprehensive evaluation of hippocampal neurochemistry, with SVS providing extended neurochemical profiles and MRSI characterizing the spatial heterogeneity of lesions (9).

The aim of the present study was 1) to investigate the feasibility of obtaining high quality hippocampal spectra that allow reliable quantification of a neurochemical profile using standard clinical 3T hardware and the semi-LASER sequence (23,24), which is gaining wider acceptance at high fields, and 2) to establish inter-session reproducibility of hippocampal MRS, including reproducibility of voxel placement, spectral quality and neurochemical concentrations. We utilized a single-voxel, short-echo semi-LASER sequence because it minimizes chemical shift displacement errors (CSD) that are particularly large with vendor-provided protocols at 3T, and provides longer apparent T₂ relaxation times and attenuated J-evolution relative to conventional Hahn spin echo sequences, excellent water suppression and a gradient scheme optimized to generate artifact-free single shots (23,25). The sequence further allows retrospective correction of motion artifacts by saving each FID separately and correcting for frequency and phase shifts that result from subject motion and eliminating those shots that show evidence of substantial motion.

EXPERIMENTAL DETAILS

Participants

Ten healthy volunteers (3 males, 7 females, mean age 37 ± 10 years) were enrolled in the study. All subjects signed an informed consent form according to the procedures approved by the Institutional Review Board at the University of Minnesota. Each subject was scanned in two sessions performed 34±30 days (mean ± SD) apart. Data acquisition and all post-processing steps were performed by a single operator (P.B.).

Data acquisition

The study was performed at a 3T Siemens Tim Trio scanner (Siemens Medical Solutions, Erlangen, Germany) with body coil excitation and the standard 32-channel receive-array Siemens head coil. The shoulders of the subjects were supported by a 3cm-thick pad with the head being slightly hyperextended in order to get the long axis of the hippocampus aligned closer to the transversal plane as suggested previously (12). High-resolution sagittal MPRAGE images (T_R = 2530 ms, T_E = 3.65 ms, flip angle = 7°, slice thickness = 1 mm, 224 slices, field-of-view = 256×176 mm², matrix size = 256×256) were obtained. The spectroscopic volume of interest (VOI) 13 × 26 × 12 mm³ (~ 4mL) was placed in the left hippocampus. The VOI was inclined by 20-35° in one dimension on the sagittal images (Fig. 1), to align it with the long axis of the hippocampal body. Due to known susceptibility differences between tissues in the anterior part of the hippocampus (26), the most anterior part of hippocampus and amygdala were not included in the VOI. Left-right voxel position was adjusted on transversal images aligning the lateral border of the VOI and lateral part of the hippocampus while minimizing the inclusion of the ventricular temporal horn. The VOI placement in the second scanning session was independent of the first session and was based on these anatomical landmarks rather than the images from the first session.

Fig. 1. Fitting of a sample spectrum with LCModel.

Spectrum acquired in 5.2 min from the left hippocampus (VOI 13 × 12 × 26 mm³) with the LCModel fit overlaid in red (a). Residual (b) and baseline spline function (c) resulting from LCModel analysis are presented. No apodization filter was applied on the displayed data. Spectra acquired with semi-LASER (TR = 5 s, TE = 28 ms, 64 transients). VOI position is shown on sagittal and axial MPRAGE images.

The first and second order shims were adjusted with the echo-planar version of FASTMAP (27) (number of echoes = 8, bar thickness =5 mm, bar FOV = 300 mm, excite pulse duration = 5.12 ms, refocusing pulse duration = 7.68 ms). MRS data were acquired with the optimized semi-LASER localization sequence (23,25) (T_E = 28 ms, T_R = 5 s) with integrated water suppression (VAPOR) and outer-volume-suppression (OVS) modules (28). The thickness of the OVS bands were set to 80 mm in all axes (x, y, z) and their distance from the VOI was 7 mm. B₁ levels required for the slice selective, asymmetric 90° pulse were adjusted for each VOI by monitoring the signal intensity whilst increasing the RF power and choosing the RF power setting that produced the maximum signal. The B₁ for the adiabatic full passage and OVS pulses was automatically set relative to the 90° pulse. In addition, the power for the VAPOR pulses was calibrated for optimum water suppression in each VOI. The signals from the 32 receive channels were combined after adjustment for the phase shifts and scaling of the signal amplitudes based on the coil sensitivities (29). The metabolite spectra were acquired in ~5 min (64 transients). Single shot data were saved separately for further post-processing. Two unsuppressed water signals were acquired to correct for residual eddy currents and as an internal reference for metabolite quantification (25).

Segmentation

In order to assess the consistency in the voxel positioning between sessions, the within-VOI brain tissue was classified with segmentation of the MPRAGE images. Probabilistic maps of the gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF) were derived by segmenting the 3D MPRAGE image using SPM8 software package (19). An in-house written MATLAB script was used to determine within-VOI fraction of GM, WM and CSF by using the iterative method of threshold selection (30) and mask of the VOI obtained for each session. Moreover, in order to assess the fraction of the hippocampus that was included in the VOI, the volume of the whole hippocampus was calculated by the standard segmentation algorithm implemented in Freesurfer software package (31).

MRS – postprocessing and quantification

Prior to summation, the single scans of metabolite spectra were corrected for small frequency and phase fluctuations and residual eddy currents were removed (32). No water removal or baseline correction was applied. All post-processing steps were performed automatically with in-house developed software. Summed spectra from each session were analyzed with LCModel (version 6.3-0G) (33). The basis set for LCModel contained 18 brain metabolites simulated with the density matrix approach (34): Ascorbate (Asc), aspartate (Asp), creatine (Cr), phosphocreatine (PCr), γ-aminobutyric acid (GABA), glucose (Glc), glutathione (GSH), glutamine (Gln), glutamate (Glu), myo-inositol (myo-Ins), N-acetylaspartate (NAA), N-acetylaspartyl-glutamate (NAAG), glycerophosphocholine (GPC), phosphoethanolamine (PE), phosphocholine (PCho), lactate (Lac), scyllo-inositol (scyllo-Ins) and taurine (Tau), as well as the spectrum of the macromolecules measured from occipital cortex as described previously (25). The validity of using a general macromolecule spectrum obtained in the occipital cortex was recently demonstrated (35). Spectra were analyzed in the frequency range 0.5-4.2 ppm and the zero- and first- order phases were fixed to zero during the analysis (for LCModel parameters see our previous work (25)).

The water scaling option in LCModel was used to quantify metabolite concentrations, utilizing the unsuppressed water spectrum as an internal reference. Since the transverse relaxation is known to differ across brain regions (36), the T₂ of the tissue water was measured in the hippocampus of three subjects by acquiring a series of unsuppressed water signals at different echo times (T_E=28-4000 ms, T_R=15 s) and fitting the integrals of the water peaks by a biexponential function. For the biexponential fit, T₂ of CSF was fixed at 740 ms, which was measured with the same semi-LASER sequence from a small voxel (0.125-0.360 mL) from the lateral ventricles in 4 healthy subjects (T_E = 28-4000 ms, T_R = 15 s). The T₂ relaxation of tissue water obtained in hippocampus (T₂ = 74.1 ± 1.1 ms, mean ± SD) was taken into account in LCModel fitting, assuming that the T₂ of water under Carr-Purcell conditions is 1.5× longer than the measured free precession T₂ (25). The smaller effect of T₂ relaxation of metabolite signals was neglected (25). The brain water content was derived assuming 81% of water in GM, 71% in WM and 100% in CSF (37). Average GM (63.8%), WM (32.7%) and CSF (3.4%) within-VOI fraction was calculated for our group of subjects (from all 20 sessions) and the average brain water content 78% (0.81 × 0.64 + 0.71 × 0.33 + 1.00 × 0.03) was used to correct metabolite concentrations. The CSF within-VOI fraction obtained for each session (as described above) was used to correct concentrations from each session.

Previously published reliability criteria were used for reporting neurochemical concentrations (25). Namely, metabolites were considered reliably quantified if the estimated error of the quantification i.e. Cramér-Rao Lower Bounds (CRLBs) were < 50% in at least 5 of 10 scans in both sessions (visit #1, visit #2). In addition, metabolite concentrations were reported only as a sum when the pair of metabolites correlated strongly to each other (r < −0.7), as in the case of Cr and PCr. They were reported both separately and as a sum when correlation coefficients were in the range −0.5 > r > −0.7, as in the case of Glc and Tau, as recommended in the LCModel manual (33).

In order to assess spectral quality, SNR of the metabolite spectra and linewidth (LW) of the unsuppressed water spectrum were evaluated. SNR was measured as a ratio of the NAA methyl resonance at 2.02 ppm and root mean square of the noise on the summed spectrum from each session (64 averages). LW was measured as full-width-at-half-maximum of the water spectrum.

Statistics

CRLBs, SNR, LW and the within-VOI hippocampal volumes were compared between the two sessions with the two-tailed paired t-test. The pair-wise analysis of within-VOI fractions of the GM, WM and CSF was performed using the Wilcoxon Signed Rank test. The threshold of statistical significance was set at p = 0.05. Pearson's correlation coefficients were obtained to compare within-VOI CSF fraction, LW and SNR between sessions.

The between-session reproducibility of metabolite concentrations was assessed by calculating coefficients-of-variance (CV = SD/mean) for each subject (n = 10) and then averaging to obtain mean between-session CVs for each metabolite. The between-session CVs for the within-VOI fractions of GM, WM and CSF were also obtained. In order to describe variability of metabolite concentrations and voxel composition (GM, WM, CSF fraction) between subjects, between-subject CVs were calculated as SD/mean of the between-session averaged values. To assess whether variability in concentrations is partly due to between-person variability in voxel composition (GM, WM fraction) or spectral quality (LW, SNR), we also fit a linear mixed model for each metabolite with those four characteristics as predictors. The linear mixed model included a random effect for person to account for the within-person correlation in measured concentrations.

RESULTS

Brain tissue segmentation confirmed consistency in the VOI placement between sessions (see Fig. in Supplementary Materials). Pairwise comparison of the within-VOI fractions of GM, WM and CSF did not reveal significant differences in GM and WM fractions between sessions, but showed a trend in the CSF fraction (p = 0.02). Therefore the CSF fraction from each session was used to correct the metabolite concentrations. The between-session CVs for the within-VOI fractions of GM, WM and CSF are shown in Table 1 and demonstrate excellent test-retest reproducibility of voxel composition. Mean volume of the whole hippocampus obtained from Freesurfer was 4261 ± 640 mm³ in session #1 and 4184 ± 582 mm³ in session #2 (mean ± SD) and the mean fraction of the hippocampal volume that was included in the VOI was 61.9 ± 6.9% and 62.0 ± 7.5% for session #1 and session #2, respectively. We did not observe a significant difference in within-VOI hippocampal volume between sessions.

Table 1.

Mean values and test-retest reproducibility of the voxel composition (white matter – WM, gray matter - GM , and cerebrospinal fluid – CSF fraction), linewidth (LW) and signal-to-noise ratio (SNR).

	Mean ± SD	Mean between – session CV (%)
WM (%)	32.7 ± 4.6	7.5
GM (%)	63.8 ± 4.7	3.5
CSF (%)	3.4 ± 2.3	17.2
LW (Hz)	8.7 ± 1.1	4.8
SNR	53.6 ± 8.1	6.3

Means are calculated from between-session averages obtained for each subject. Between-session CVs averaged across subjects are shown.

High reproducibility of the spectral pattern and quality (SNR, spectral resolution, water and unwanted coherence suppression) between subjects and between the two sessions is demonstrated in Fig. 2. Notably, spectral patterns appeared subject-specific (note the ratios of the major metabolite peaks in each subject in Fig. 2a). An example of the LCModel analysis of a spectrum acquired in ~5 min, demonstrating the quality of spectral fitting routinely achieved in this study, is presented in Fig. 1. The mean concentrations, CRLBs and between-session CVs of metabolites that met our abovementioned reliability criteria are presented in Fig. 3. The metabolite concentrations and CRLBs were not significantly different between-sessions and the correction for the small CSF fraction (3-4%) in the hippocampus VOI did not affect the test-retest reproducibility of neurochemical concentrations, as expected.

Fig. 2. Between-session and between-subject reproducibility of the hippocampal spectra.

(a) Spectra acquired in 5 selected healthy subjects in two consequent sessions from the hippocampus are overlaid to demonstrate between-session reproducibility of the MRS data. (b) In order to demonstrate between-subject reproducibility, the spectra from the two sessions are averaged for each subject and overlaid (n=10). NAA-methyl resonance at 2 ppm is used to scale the spectra. Gaussian apodization function (σ = 0.13 s) was used for display purposes.

Fig. 3. Group analysis of LCModel quantification.

Concentrations were averaged between the two sessions for each subject and mean values calculated across 10 subjects. Concentrations corrected for CSF and tissue water T₂ relaxation are shown (a). The corresponding mean CRLBs and mean between-session coefficients of variance (CVs, SD/mean) of metabolite concentrations are shown in percent (b) and in absolute units (c). Error bars represent between-subject SDs.

The spectral quality allowed quantification of tNAA, NAA, Glu+Gln (Glx), tCho, tCr, Glu and myo-Ins with CRLB below 10% and between-session CVs of 6% or less (Fig. 3b). Asc+GSH, Glc, GSH, Glc+Tau, scyllo-Ins, and Tau were quantified with mean CVs < 20% (Fig. 3b). Test-retest CVs and CRLBs were comparable per metabolite. While they showed great variance across metabolites when expressed in percent (Fig. 3b), they were more uniform across metabolites when expressed in absolute units (Fig. 3c).

Metabolite concentrations and three factors that are known to affect metabolite levels (within-VOI CSF fraction, LW and SNR, see Table 1) were also compared using Pearson's correlations between sessions (Fig. 4). Fig. 4a demonstrates the reproducibility for the most reliably quantified metabolite concentrations (CRLB < 9%). LW and SNR were highly reproducible between sessions, with mean test-retest CVs of 5-6% (Table 1). Consistently, the water LW and SNR correlated significantly between session #1 and session #2 (R = 0.76, p = 0.01 for LW and R = 0.80, p < 0.0006 for SNR; Fig. 4b,c). Similarly, CSF fractions per subject were highly correlated between sessions, indicating the subject specificity of these (R = 0.98, p < 0.0001, Fig. 4d).

Fig. 4. Reproducibility of metabolite concentrations, linewidths and SNR between sessions.

Only metabolites with CRLBs < 9% are presented and concentrations of selected metabolites are distributed along identity line (a). Correlation plots demonstrate the subject-specific nature of linewidths (measured on the unsupressed water spectra) (b), SNR (c) and within-VOI CSF fraction (d).

Given consistent voxel positioning between sessions and minimal day-to-day biological variation in the quantified neurochemicals in the healthy brain, the between-session CVs of metabolite concentrations represent the error of the measurement and/or quantification. The between-subject CVs encompass both the measurement error and inter-individual differences, which can be detected only if the measurement error (between-session CVs and CRLBs) is smaller. In the current study, the between-session CVs appeared to be well below between-subject CVs for most metabolites (Fig. 5a). Similarly, the CRLBs tended to be lower than between-subject CVs (Fig. 5b), at least for the metabolites quantified with CRLB < 25%, Fig. 5c. Importantly, for almost all metabolites, the voxel composition (GM, WM fraction) and spectral quality (SNR, LW) characteristics were not predictive of concentration in the linear mixed models. For Glu, spectral quality was significantly negatively associated with concentration (p=0.03 after Bonferroni correction for testing LW for each of the 16 metabolites). Therefore, except for Glu, the variance in the voxel composition and spectral quality did not explain the variance in metabolite concentrations we observed in the healthy human hippocampus.

Fig. 5. Comparison of between-subject and between-session variance of MRS data.

Comparison of mean between-session coefficients of variation (CVs) and mean between-subject CVs (calculated from concentrations averaged between sessions per subject) (a). Comparison of mean CRLBs and mean between-subject CVs (both calculated from between-session averages) (b). (c) is a zoomed version of (b) and shows only the metabolites measured with CRLBs < 25%. Each data point represents a metabolite.

DISCUSSION

This study demonstrated the feasibility of obtaining spectra of high quality from the hippocampus, a region particularly challenging for MRS (12). Our study benefited from utilizing an in-house developed semi-LASER localization sequence implemented at high field (3T) in combination with FASTMAP B₀ shimming. The same methodology optimized for 3T was recently introduced in a 2-site study focusing on brain regions other than the hippocampus (25). Here we were able to quantify a neurochemical profile in the hippocampus with a reproducibility and quantification precision comparable to other brain regions (18,25).

Short TE semi-LASER localization provides a similar spectral pattern to that achieved with ultra-short echo STEAM, because the apparent T₂ and J evolution is reduced by the Carl-Purcell train of adiabatic pulses (38). In addition, semi-LASER provides almost two-fold increase of the SNR relative to STEAM (23). Broadband adiabatic pulses improve the slice selection profile and substantially reduce the CSD relative to the standard full signal intensity non-adiabatic localization approach (PRESS) at 3T. For example, with the vendor-provided PRESS sequence the CSD is 12-13%/ppm for the slices selected by the 180° pulses on the Tim Trio system we utilized, whereas it is 2%/ppm for the adiabatic 180° pulses in semi-LASER. Finally, the precise localization together with robust OVS and VAPOR techniques allow acquiring of spectra free of unwanted coherences such as the out of phase signals of lipids. Note that both the semi-LASER sequence we have used in this study (23) and the FASTMAP shimming tool are now available as work-in-progress (WIP) package on the Siemens platform.

The iterative FASTMAP protocol we used to adjust 1^st and 2^nd order shims led to excellent LW (below 9 Hz for water on average). The hippocampus is located close to the skull base and nasal sinuses, which have magnetic properties different from the brain causing distortions in B₀. Since broad LWs decrease spectral dispersion and can compromise quantification (16), optimal adjustment of B₀ field within the spectroscopy voxel was crucial in the hippocampal region. Interestingly, the LW achieved in our study ranged from 7-11.4 Hz among the 10 subjects and correlated significantly between sessions (Fig. 4b). We have demonstrated a similar correlation between LW achieved at 4T and 7T in individual subjects (39). These findings indicate that between-subject variance (Table 1) results to a large extent from inter-individual differences in the (microscopic) magnetic susceptibility of the brain tissue within the VOI. In addition, the location of the VOI within the brain, such as its distance from the skull base may affect the achieved LW. The variation of LW affected the variation of SNR (Fig. 4b,c, Table 1), which is further influenced by other subject-specific factors such as coil loading. The high reproducibility of LW and SNR within subjects likely contributed to the high between-session reproducibility of metabolite concentrations.

The high SNR achieved in our study thanks to the full intensity localization sequence and the high-field system, allowed us to use a small VOI (~4mL). The relatively small voxel reduced partial volume effects and avoided the structures surrounding the hippocampus (parahippocampal gyrus, thalamus), as well as areas with increasing susceptibility effects (amygdala) (12). We were able to include ~64 % of the hippocampal GM within our VOI, consistent with other studies (37) and we covered ~62% of the total hippocampal volume. We further found that the correction for the small CSF fraction in the VOI (3-4 %) had a negligible effect on metabolite levels and their reproducibility in the healthy cohort. Nonetheless, this correction will likely be critical when clinical populations with hippocampal atrophy are studied (37). Similarly, the correction of metabolite concentrations for the effect of T₂ relaxation might be valuable when healthy volunteers are compared with clinical populations, since the T₂ of brain water and metabolites may differ between groups (40).

The CRLBs, a measure of quantification precision provided by LCModel, observed in our study were improved and allowed reliable quantification of more metabolites than prior studies (8,12,17). We were able to quantify 5-7 metabolites with CRLBs below 9% (Table 2) and another five metabolites with CRLBs below 30%. Therefore, the current study provides the most detailed in vivo characterization of the human hippocampal neurochemistry obtained with widely available clinical hardware. The neurochemical profile was consistent with other studies, e.g. we observed higher myo-Ins and lower NAA than in neocortical tissues due to a higher glial and lower neuronal density (16,17). The CRLB obtained from the hippocampus with our approach were comparable to CRLB reported from other brain regions (25).

Table 2.

Metabolites quantified with high reproducibility and quantification precision.

	CRLB (%)	Between-session CV (%)	Between-subject CV (%)
Glu	8.7	6.0	12.1
Glx	6.6	5.6	13.3
myo-Ins	3.5	3.1	13.0
tCho	5.6	3.3	13.0
tCr	3.5	2.0	3.8
NAA	4.0	3.0	7.5
tNAA	3.4	4.3	7.3

CRLBs, between-session CVs and between-subject CVs of the metabolites that were quantified with CRLB < 9 % and between-session CVs ≤ 6% are presented.

Notably, the SNR and spectral resolution achieved in our study was sufficient to separately quantify Glu and Gln resonances. Even though these metabolites correlated moderately (r ~ - 0.5), they could be quantified reproducibly between sessions (Fig. 3). Hippocampal Glu is a metabolite of particular clinical interest as demonstrated in studies seeking surrogate markers for diagnostics (4,37) or therapy monitoring in Alzheimer's disease (8,41) and schizophrenia (11).

The between-session CVs reported in prior test-retest studies (16,17,19,42) of the hippocampus were substantially higher than our findings. We observed between-session CVs ≤ 6% for Glu, Glx, myo-Ins, tCho, tCr, NAA and tNAA (Table 2), whereas CVs previously reported at 1.5T were in the range 7-20% (for tNAA, tCho, myo-Ins, tCr, Glx) (16,17). In a study conducted at 3T (42), between-session CVs were 5% for tNAA, above 7% for tCr, tCho, myo-Ins and 22% for Glx when using LASER with T_E = 65 ms. In another study the average between-session CV for tNAA, tCho and tCr was 13.9%, despite using higher field (4 T) (19), demonstrating the importance of utilizing carefully optimized methodology in addition to high field strength for optimal spectral quality.

While between-subject CVs determine the ability of the method to reveal group differences in cross-sectional studies, low test-retest CVs are critical for clinical utility and planning of longitudinal studies as they will allow detection of small changes with disease progression or in response to treatment in individuals. Interestingly, between-session CVs improved relative to prior studies, but the between-subject CVs (Table 2) remained comparable to other studies from hippocampus (16) and other brain regions (25). Therefore the improvements afforded with the novel MRS methodology primarily will benefit longitudinal studies, such as clinical trials, when pre- vs. post-treatment differences are evaluated. Our data also indicate that the relatively high between-subject variance is due to real inter-individual differences in metabolite concentrations, except in the case of Glu, rather than due to variance in the voxel composition and spectral quality (SNR, LW). Therefore the lower between-session CVs than between-subject CVs demonstrate an ability to measure real inter-individual differences in metabolite concentrations in the healthy brain, as well as the potential to reveal subtle changes caused by pathology in individuals. Consistently, the CRLBs of the metabolites quantified with higher quantification precision (CRLB < 25 %, Fig. 5c), were also lower than between-subject CVs, as in our prior studies (25,39).

Finally, this study demonstrated that an experienced operator can select the VOI highly reproducibly between sessions and subjects (Supplementary Fig.). Note however that potential operator dependence of the VOI selection is a limitation of the current protocol. While acquisition of data by a single operator is common in the research setting, methods to automatically prescribe VOI, especially in regions with large susceptibility changes such as the hippocampus (20), are critical for wider utility of the method in the clinical setting with multiple operators.

CONCLUSION

High spectral quality can be obtained reproducibly from a ~4mL hippocampal voxel with minimal partial voluming using standard 3T hardware, an in-house implemented short echo, single voxel pulse sequence and 5 min data averaging. These spectra allow the reliable quantification of a neurochemical profile from this challenging VOI. Namely, tNAA, NAA, tCho, tCr, myo-Ins and Glu can be quantified with high scan-rescan reproducibility (CV ≤ 6%) and quantification precision (CRLB < 9%). Six other metabolites, including glucose, glutathione and taurine, are quantified with CRLB < 30%. Finally, for the most reliably quantified metabolites, the method has sufficient sensitivity to detect inter-individual differences in the healthy brain. Thus this novel methodology can be utilized in a broad range of clinical research applications focused on hippocampal neurochemistry altered by diseases such as Alzheimer's disease, epilepsy, schizophrenia and depression, and particularly will benefit studies of the neurochemical response evoked by therapies.

Abbreviations used

Asc

ascorbate

Asc+GSH

ascorbate+guthathione

Asp

aspartate

Cr

Creatine

CRLB

Cramer–Rao lower bound

CSD

chemical shift displacement

CSF

cerebrospinal fluid

CV

coefficient of variation

FID

free induction decay

GABA

γ-aminobutyric acid

Glc

glucose

Glc+Tau

glucose+taurine

Gln

glutamine

Glu

glutamate

Glx

glutamate+glutamine

GM

gray matter

GPC

glycerophosphocholine

LCModel

linear combination model

LW

linewidth

MPRAGE

magnetization prepared rapid gradient echo

Myo-Ins

myo-inositol

NAA

N-acetylaspartate

NAAG

N-acetylaspartylglutamate

OVS

outer-volume supression

PCho

phosphocholine

PCr

phosphocreatine

PE

phosphorylethanolamine

SD

standard deviation

scyllo-Ins

scyllo-inositol

SNR

signal-to-noise ratio

Tau

taurine

tCho

total choline

tCr

total creatine

tNAA

total N-acetylaspartate

VOI

volume of interest

WM

white matter