Repeatability of ³¹P MR spectroscopic imaging in the human brain at 7 T with and without the Nuclear Overhauser Effect

Abstract

An often employed strategy to enhance signals in ³¹P MR spectroscopy is the generation of the nuclear Overhauser effect (NOE) by saturating the water resonance. However, NOE allegedly increases the variability of the ³¹P data, because variation is reported in NOE enhancements. This would negate the SNR gain it generates. We hypothesized that variation in NOE enhancement values is not due to variability in the NOE itself, but that it is attributable to measurement uncertainties in the values used to calculate the enhancement. If true, the expected increase in SNR with NOE would improve the repeatability of ³¹P MR spectroscopy measurements. To verify this hypothesis, a repeatability study of native and NOE-enhanced ³¹P MRSI was performed in the brain of 7 healthy volunteers at 7T.

The repeatability coefficient (RC) and the coefficient of variation in repeated measurements (CoV_repeat) were determined per method, and the 95% limits of agreement (LoA) between native and NOE-enhanced signals were calculated. The variation between the methods, defined by the LoA, is at least as great as that predicted by the RC of each method. The sources of variation in NOE enhancements were determined using variance component analysis.

In the 7 metabolites with a positive NOE enhancement (9 metabolite resonances assessed), CoV_repeat improved on average by 15%. The LoAs could be explained by the RCs of the individual methods for the majority of the metabolites, generally confirming our hypothesis. Variation in NOE enhancement was mainly attributable to the factor repeat, but between-voxel effects were also present for phoshpoethanolamine and (glycero)phosphocholine. CoV_repeat and fitting error were strongly correlated and improved with positive NOE.

Our findings generally indicate that NOE enhances the signal of the metabolites, improving the repeatability of metabolite measurements. Additional variability due to NOE was minimal. These findings encourage the use of NOE-enhanced ³¹P MRSI.

Introduction

Phosphorus (³¹P) MR spectroscopy is a non-invasive technique to study tissue metabolism under various physiological and pathophysiological conditions. Human organs and tissues frequently investigated using ³¹P MRS include the liver (1,2), skeletal muscle (3,4), the heart (5) and the brain (6,7), and studies concerning the prostate (8), the placenta (9), and breast (10) have also been reported recently. Metabolic changes can be followed over time with dynamic ³¹P MRS, and MR spectroscopic imaging (MRSI) allows the two- or three-dimensional mapping of metabolites to study their spatial distribution.

The low intrinsic sensitivity of in vivo ³¹P MRS and MRSI however impedes the regular use of these techniques. Obviously, advances in radiofrequency (RF) coil design and increases in magnetic field strength have been embraced to gain ³¹P sensitivity enabling reduced overall acquisition times or increased spatial resolution. Sensitivity of ³¹P spectra can also be improved by adding proton (¹H) irradiation to the MR sequence (11–13). Proton irradiation may activate two physical mechanisms: decoupling and the nuclear Overhauser effect (NOE).

The first process, proton decoupling, is particularly useful at clinical field strengths (1.5-3 T), where the line width of the resonances of interest after shimming can be of similar size as the weak heteronuclear ³¹P-¹H J-couplings (5-7 Hz). Without proton decoupling the coupled metabolites split into doublets or triplets, giving rise to broad resonances and spectral overlap. Many metabolites present in an in vivo ³¹P spectrum contain one or more weak ³¹P-¹H J-couplings, i.e. phosphoethanolamine (PE), phosphocholine (PC), glycerophosphoethanolamine (GPE), glycerophosphocholine (GPC), 2,3-diphosphoglycerate, α-adenosine triphosphate (αATP) and nicotinamide adenine dinucleotide (NAD⁺ and NADH). The improvement in spectral resolution obtained by decoupling reduces the spectral overlap of the ³¹P metabolites and thus allows better peak assignment and consequently better spectral quantification. Decoupling can be achieved by applying high-power broadband proton irradiation during ³¹P signal acquisition, creating high specific absorption rates (SAR) at higher field strengths.

The second process of signal improvement, the NOE, uses low-power proton irradiation to saturate the water resonance in between ³¹P signal acquisitions. The saturated water protons can increase the steady state magnetization of the ³¹P nuclei through dipolar interactions, resulting in enhanced ³¹P metabolite signal intensities. In vivo enhancements up to 80% have been reported (14), which encourages the use of NOE-enhanced ³¹P MRS or MRSI. Because NOE requires considerably less power than decoupling, it is also suited for higher magnetic field strengths.

Absolute NOE enhancement values depend on the extent to which the dipolar ³¹P-¹H interaction mediates T₁ relaxation of the ³¹P spins (15). The enhancement values therefore vary between metabolites, and are dependent on the mobility of the metabolites in the tissue and magnetic field strength (14,16–22). Sufficient saturation of the water resonance is essential to maximize the NOE enhancement (12,13), and needs to be optimized for each sequence. At ultrahigh magnetic field strengths it is even more important to determine the minimum transmit RF amplitude (B₁⁺) for constant NOE, because of the known spatial inhomogeneities in the B₁⁺ field (22).

The in vivo NOE signal enhancement indeed seemed very beneficial initially (14), but considerable variation was later reported in NOE enhancement values per metabolite, even if the measurements were calibrated well (19,20,22). This has raised concerns that using NOE might add variability to the ³¹P data, negating the gain in signal-to-noise ratio (SNR). We hypothesized that variation in NOE enhancement values is not due to variability in the NOE itself (i.e. non-constant NOE enhancements because of natural variation or inadequate NOE calibration), but that it is attributable to measurement uncertainties in the values used to calculate the NOE enhancement. If true, this would argue for the use of NOE as the SNR is expected to increase, which would potentially decrease measurement uncertainties. To test this hypothesis repeated measurements with both methods, i.e. native, non-enhanced ³¹P MRS/ MRSI and NOE-enhanced MRS/ MRSI are needed. The measurement uncertainties of the two methods, estimated by their repeatabilities, limit the amount of agreement possible between the two methods, and thus define a minimum variation in NOE enhancement values per metabolite (23).

Ultimately, one wishes to answer the question whether or not to use NOE in ³¹P MRS/MRSI studies in any tissue at any field strength. Because of the dependency of NOE enhancements on dipolar relaxation processes and saturation efficiency, spatial variations in NOE enhancement might be possible, and these may depend on the tissue type and state (healthy and diseased) and the field strength. To reduce complexity regarding all these issues, the research question in this study was narrowed down to the possible beneficial effect of NOE on ³¹P MRSI of the healthy human brain at 7 T. To answer this question we performed repeated measurements of native ³¹P MRSI and NOE-enhanced ³¹P MRSI in the brain of healthy volunteers at 7 T.

Experimental

Hardware

Measurements were performed on a 7 T whole-body MR research system (MAGNETOM, Siemens Healthcare, Erlangen, Germany). An eight-rung high?pass quadrature birdcage coil tuned to ³¹P (120.3 MHz) was used (24), which was designed to fit within a homebuilt eight-channel ¹H array head coil with meander elements (25). The safety of the coil combination was assessed previously (24). ¹H MR imaging and ¹H saturation for NOE generation was performed with an add-on system for RF shimming, which also performed real-time SAR monitoring on both the ¹H and the ³¹P channels (26,27).

Subjects

Seven healthy volunteers (6 male, 1 female, mean age 30 years, range 21-39 years) participated in this study. The study was conducted in accordance with all guidelines set forth by the approving institutional review board, and signed informed consent was obtained from all subjects before the measurements.

NOE generation

To ensure efficient water saturation for NOE, a wideband alternating-phase low-power technique for zero residual splitting (WALTZ-4, historically used for decoupling) (28) was adopted. The WALTZ-4 train was applied during the full TR, except during the ³¹P RF pulse and the 204 ms of signal acquisition. The durations of the individual pulses of one WALTZ sub-train were 2.5, 5.0 and 7.5 ms, with 2.5 ms pauses in between segments, requiring 14 WALTZ-4 trains of 90 ms each per repetition time. The minimum power needed to generate constant NOE enhancements for this setting of WALTZ-4 was determined to be 30 Hz in previous phantom experiments (partly presented in (22)).

MR measurement protocol

Anatomical images for planning of the 3D ³¹P MRSI measurements were acquired using a magnetization prepared 3D T1-weighted gradient echo (MPRAGE) sequence. Magnetic field (B₀) phasemap shimming was performed to optimize the B₀ homogeneity in the posterior region of the brain. ¹H transmit RF amplitude (B₁⁺) was maximized in the same region using RF shimming. The amplitude of WALTZ-4 pulses to generate a constant NOE was set to reach at least 30 Hz in the posterior part of the brain by using absolute B₁⁺ maps (fig. 1a). These maps were acquired using a magnetization prepared gradient echo sequence (TR = 5000 ms, TE = 5 ms, magnetization preparation flip = 90°, excitation flip angle = 10°) (29).

Figure 1.

Overview of preparation steps for native and NOE-enhanced ³¹P MRSI and data analysis in one subject. A: Transversal absolute ¹H B₁⁺ map after B₁⁺ calibration to obtain at least 90° flip angles in the posterior part of the brain, allowing minimally 30 Hz NOE power. B: Sagittal MPRAGE image. The location of the absolute ¹H B₁⁺ map is indicated by the red line and the slice selection for flip angle calibration of ³¹P is indicated by the white lines. The yellow arrows indicate signal dropout due to low B₁⁺. C: Transversal MPRAGE images with the spectroscopic grid as an overlay. The voxels indicated in blue were used for data analysis. The approximate real voxel size is indicated by the yellow dotted circle.

Flip angle calibration for ³¹P was performed using a slice selective pulse acquire sequence (6 ms excitation pulse, TR = 15 s), selecting a 40 mm thick slice through the posterior brain (fig. 1b). The amplitude of the RF pulse was varied between shots to find the maximum magnitude of the phosphocreatine (PCr) peak. This maximum was assumed to correspond to 90° excitation, and all ³¹P RF amplitudes in the subsequent MRSI sequences were scaled relative to this value.

For the 3D ³¹P MRSI measurements, a pulse acquire sequence was used with a 300 μs block pulse for excitation (flip angle = 45°) and a TR of 1500 ms. The field of view (FOV) was set to 240 × 240 × 240 mm and the 3D phase encoding matrix to 12 × 12 × 8, creating isotropic dimensions in the transverse plane. The delay between the excitation pulse and the start of data acquisition was 410μs. An elliptical k-space acquisition scheme was applied (1 average) and k-space was Hamming filtered (100%). The nominal voxel size of 20 × 20 × 30 mm is therefore enlarged to a ellipsoid-shaped voxel volume of approximately 38 cm³ (30). The MRSI acquisition of 7 min 48 s was repeated four consecutive times. The first and third MRSI acquisitions were obtained without ¹H irradiation (native), the second and the fourth with ¹H irradiation to generate the NOE.

Data analysis – voxel selection and spectral fitting

Three non-neighboring MRSI voxels were selected per volunteer (fig. 1c). Care was taken to select only voxels that received a minimum ¹H B₁⁺ of 30 Hz in the case of NOE generation, that were situated completely within the brain, and that showed good spectral quality in all four MRSI datasets. Different voxel locations were chosen per volunteer. The average B₁⁺ in each selected voxel was estimated using the absolute B₁⁺ map.

The ³¹P spectra of the selected voxels were fitted with Metabolite Report, a work in progress package from Siemens Healthcare (Erlangen, Germany). Metabolite Report performs automated, prior knowledge based, complex fitting in the time domain and has previously been applied to 7 T ³¹P spectroscopy data of the prostate (8). The spectral resonances PE, PC, GPE, GPC, inorganic phosphate (Pi), α-, γ-, β-ATP and NAD⁺/ NADH were fitted using Gaussian lineshapes, while PCr was fitted with a Lorentzian lineshape. Prior knowledge of the relative peak frequencies and the ³¹P J-coupling patterns in the ATPs was used. The line widths of PE, PC, GPE, and GPC were assumed to be the same and thus constrained to each other. The quality of the spectral fit was assessed using the Cramér-Rao lower bound (CRLB) values and qualitatively by visual inspection by spectroscopist. The CRLB values of the fits with and without NOE were compared using a paired t-test per metabolite.

Only metabolite fits with a CRLB below 30% were considered reliable. The original spectra and the fitting results were visually inspected, and if a metabolite peak was visually present and its fit was assigned to the correct resonance, giving a minimal residue, the fitting result was accepted. Only metabolites which passed both quality checks were used in further analyses.

The NOE enhancement (η) per metabolite in each selected voxel was calculated by

$$\eta =\frac{{\overline{NOEenh}}_{repeat}-{\overline{native}}_{repeat}}{{\overline{native}}_{repeat}},$$ (1)

with ${\overline{NOEenh}}_{repeat}\text{and}{\overline{native}}_{repeat}$ representing the mean metabolite peak integral of the two repeated measurements using NOE-enhanced and native MRSI respectively.

Theory of repeatability and method comparison

To be able to answer our research question, we used an approach presented by Bland and Altman to assess the agreement between two methods of clinical measurement (23). The term “agreement” might be misleading here, since a difference between the native and NOE-enhanced method is expected because of the NOE enhancements for most metabolites. However, the agreement between methods, or the lack of it, has two components: the bias, i.e. the mean difference between the methods, and the random variation. The random variation is at least as great as that predicted by the repeatability of each method, but may be larger because of heterogeneity between individual measurement units (e.g. subjects, but in this study individual voxels as well). For correct interpretation of the measure of agreement between methods it is thus essential to first determine the repeatability of each individual method.

Statistical analyses of repeatability and method comparison

Useful measures for the repeatability of a method are the repeatability coefficient (RC) and the coefficient of variation in repeated measurements (CoV_repeat). The RC defines the width of the interval within which we expect the absolute difference between two measurements with the same method to lie in 95% of the cases. It is defined as $\sqrt{2}\cdot 1.96\cdot {\sigma }_{repeat}=2.77\cdot {\sigma }_{repeat}$, where σ_repeat is the standard deviation of repeated measurements. Once the RCs are known for two methods, their repeatability intervals (2*RC) can be directly compared to the interval of the 95% limits of agreement (LoA) between the two methods (see below), allowing to assess how much of the observed variation between the methods can be explained by the repeatability of the individual methods (23).

In the present work, σ_repeat was estimated for each metabolite and each method (native and NOE-enhanced) by decomposing the total observed variance into the factors “between-subject”, “between-voxel” (or “within-subject”), and “repeat” (or “within-voxel”):

$${\sigma }_{tot}^2={\sigma }_{between-subject}^2+{\sigma }_{between-voxel}^2+{\sigma }_{repeat}^2$$ (2)

This was done by fitting a 3-level no-predictors linear mixed model to the data of each metabolite and method, in which “subject” and “voxel” were treated as level-3 and level-2 random intercepts, respectively.

The CoV_repeat is a normalized measure of σ_repeat to study relative repeatability:

$$Co{V}_{repeat}+\frac{{\sigma }_{repeat}}{\mu }\ast 100\%$$ (3)

with μ the mean of the metabolite integrals of all voxels and repeated measurements, without or with NOE (μ_native and μ_NOEenh). The native and NOE-enhanced CoV_repeat were determined per metabolite and for metabolite ratios, using standard error propagation methods. The apparent CoV_repeat of NOE-enhanced metabolite intensities corrected for their enhancement (which would be needed for absolute quantification) was calculated using error propagation as well. A relation between quality of the metabolite fits and the repeatabilities of the metabolites was explored by the Pearson correlation between CRLB and COV_repeat.

The agreement between native and NOE-enhanced ³¹P MRSI was quantified using their 95% limits of agreement (LoA). The interval of 95% LoA is determined by the overall bias between the two methods $\left(\stackrel{‒}{d}\right)$ and the variation around this mean difference, expressed by the standard deviation of the differences (σ_d):

$$95\%\text{LoA}=\stackrel{‒}{d}\pm 1.96\ast {\sigma }_d.$$ (4)

The LoA were calculated using ${\overline{native}}_{repeat}\text{and}{\overline{NOEenh}}_{repeat}$ (the mean integrals of the two repeated measurements per voxel with each method). The overall bias $\stackrel{‒}{d}$ was estimated using

$$\stackrel{‒}{d}=\langle {\overline{NOEenh}}_{repeat}-{\overline{native}}_{repeat}\rangle ,$$ (5)

where the brackets denote the average over all voxels. Since we include all data in calculating the agreement, using the two repeated measures for ${\overline{native}}_{repeat}\text{and}{\overline{NOEenh}}_{repeat}$, the observed standard deviation of the differences between pairs of means of repeated measures $s_{\stackrel{‒}{d}}$ needs an adaptation to attain the standard deviation of the differences between the methods σ_d (23):

$${\sigma }_2^d=S\frac{2}{d}+\frac{{\sigma }_{repeat\_native}^2}{2}+\frac{{\sigma }_{repeat\_NOEenh}^2}{2}$$ (6)

The 95% LoA and the repeatability intervals were compared for each metabolite. If these ranges are similar, then the (lack of) agreement is fully explained by the (lack of) repeatability. If the range of the 95% LoA is considerably wider than the repeatability intervals of the individual methods, then (some) additional heterogeneity exists between the individual voxels and subjects in NOE enhancement (i.e. NOE would introduce additional signal variation).

The presence of NOE-induced additional signal variation between voxels and subjects was also assessed by variance components analysis of measured NOE enhancements, using the same model as described above (Eq. 2).

The effect of NOE on the peak integrals of individual metabolites was tested for significance using similar 3-level mixed models as above, with “NOE” as an additional fixed factor. To test whether the minimum ¹H power criterion of 30 Hz for constant NOE was sufficient in vivo, linear mixed models of the NOE enhancements of the individual metabolites were constructed, which included local B₁⁺ as a fixed covariate. The possible effect of local B₁⁺ on NOE enhancement was assessed using these models. Statistical test outcomes with p-values < 0.05 were considered significant.

Results

All subjects completed the full scan protocol, without having to repeat any measurements. The minimum power condition of 30 Hz for NOE was reached in every subject in the posterior part of the brain (example in figure 1). All measurements could be performed within the simulation-based SAR limits (15W/6 min into the coil), governed by local SAR limits and an additional 40% safety margin. The power on the ¹H channel had the highest contribution to the total SAR. High quality spectra were obtained throughout large parts of the brain, but to meet the 30 Hz criterion on ¹H, only voxels from the posterior brain region were selected for quantification. The βATP intensity was strongly attenuated or not present because it was outside the effective bandwidth of the excitation pulse and it was therefore excluded from further analysis. The PC and NAD resonances were sometimes only visible as a small shoulder on the right side of respectively PE or αATP. An example of a spectral fit is presented in the supplementary material.

Small variations in spectral patterns between repeated measurements using the same method were observed, and NOE enhancements could also be observed visually (fig. 2). For all metabolites except αATP the quality of the spectral fit was significantly better in the NOE-enhanced ³¹P MRSI data than in the native ³¹P MRSI data (table 1). From 378 metabolite signals (7 volunteers * 2 measurements * 3 voxels * 9 signals) in the native ³¹P MRSI data, 15 metabolite fits had a CRLB above 30% (PC: 2, NAD: 13), while from 378 signals in the NOE-enhanced ³¹P MRSI data, this occurred 2 times (PC: 1, NAD: 1). None of the metabolite signals with a CRLB below 30% was discarded by visual inspection.

Figure 2.

Example of repeated spectra of the same voxel in one volunteer using native ³¹P MRSI (left, native) and using NOE-enhanced ³¹P MRSI (right, NOEenh). Variations between the individual measurements per method are visible, and the NOE-enhanced spectra show a clear increase in signal for several metabolites. PE: phosphoethanolamine, PC: phosphocholine, Pi: inorganic phosphate, GPE: glycerophosphoethanolamine, GPC: glycerophosphocholine, PCr: phosphocreatine, γ- and α-ATP: γ- and α-adenosine triphosphate, NAD: nicotinamide adenine dinucleotide.

Table 1.

NOE enhancements (±, mean ± SD) and quality of the spectral fit expressed in percentage Cramér-Rao Lower Bounds (CRLB, mean ± SD) of native and NOE-enhanced spectra.

	PE	PC	Pi	GPE	GPC	PCr	γATP	αATP	NAD
η	0.39±0.22	0.19±0.31	0.21±0.20	0.30±0.22	0.36±0.20	0.22±0.05	0.07±0.07	0.00±0.05	0.21±0.23
CRLB native	9.4±1.7	20.2±5.8	13.2±2.6	15.0±3.4	11.0±2.7	3.3±0.8	2.4±0.4	2.9±0.7	26.5±7.4
CRLB NOEenh	6.8±1.0	16.7±4.1	10.8±1.8	11.7±2.6	8.1±1.8	2.7±0.6	2.3±0.3	2.8±0.7	21.2±5.5
p-value	<0.001	0.001	<0.001	<0.001	<0.001	<0.001	0.001	0.56	<0.001

NOE enhancement factors varied between metabolites (table 1, fig. 3), with mean enhancements up to 39%. The fitted ³¹P metabolite integrals were significantly higher (p≤0.003) using NOE for all metabolites except αATP. PE, PC, Pi, GPE, GPC and NAD showed considerable variation in NOE enhancement between the voxels, but outliers did not result from the same voxels.

Figure 3.

Nuclear Overhauser Effect (NOE) enhancements (η) per voxel of ³¹P metabolite resonances in human brain at 7 T. NOE enhancement resulting from the first and second set of native and NOE-enhanced measurements are shown in black and red, respectively.

The repeatability coefficients of native and NOE-enhanced ³¹P MRSI were comparable for all metabolites except for GPE and NAD (fig. 4a). GPE was considerably less repeatable with NOE, whereas NAD was better repeatable with NOE. Because of the signal increase with NOE, the CoV_repeat (relative repeatability within voxels) was better in the NOE-enhanced data for PE, PC, Pi, GPC, PCr, and NAD (table 2, fig. 4b). Similarly, the relative repeatability of metabolite ratios containing these metabolites improved with NOE (table 2). In general, CoV_repeat was best for PCr, γATP and αATP, i.e. metabolites with the highest signal integrals and smallest fitting errors (table 2). Correcting NOE-enhanced metabolite intensities for their enhancement factors for absolute quantification purposes causes larger variation in all metabolites compared to the native or NOE-enhanced data (table 2). A strong correlation existed between mean CRLB and COV_repeat (R²=0.91, p<0.0001, fig. 5).

Figure 4.

A: 95% Limits of agreement (LoA) between the two methods and repeatability intervals (2*RC) of each method. B: Variance components in NOE enhancement.

Table 2.

Coefficients of Variation in repeated measurement (CoV_repeat) of native and NOE-enhanced data per metabolite and for metabolite ratios, as well as the apparent CoV_repeat of NOE-enhanced metabolite intensities corrected for their enhancement.

			CoVrepeat
	Native	NOEenh	Corrected		Native	NOEenh
PE	11,8%	7,3%	17.4%	PE/PC	23,1%	19,3%
PC	19,9%	17,8%	31.6%	GPE/GPC	16,1%	18,1%
Pi	16,2%	12,2%	20.6%	PE/PCr	12,8%	8,5%
GPE	11,9%	15,8%	23.3%	Pi/PCr	16,9%	13,0%
GPC	10,9%	8,8%	17.0%	(PE+PC)/PCr	23,2%	17,0%
PCr	4,8%	4,4%	6.2%	PCr/ γATP	8,1%	8,0%
γATP	6,5%	6,7%	9.5%	NAD/ αATP	28,4%	18,5%
αATP	6,6%	6,1%	7.8%
NAD	27,6%	17,5%	25.7%

Figure 5.

Relation between fitting error of the metabolites (expressed in percentage Crémer-Rao Lower Bound, CRLB, average across subjects) and the relative repeatability (expressed in the coefficient of variation in repeated measurements, COV_repeat) of the metabolites. Black symbols: native. Red symbols: NOE-enhanced.

The agreement between native and NOE-enhanced ³¹P MRSI, expressed by the 95% LoA, was in the range of the repeatability intervals of the individual methods for all metabolites except PE (fig. 4a). The variation in NOE enhancements between subjects was dominated by the “repeat” component for all metabolites (mean 85%, minimum 60% of total variance for PE, fig. 4b). Between-subjects variations were negligible (mean 3%, maximum 10% of total variance), while between-voxel differences explained a noticeable fraction of the variance only for PE (40%), GPC (34%) and PC (19%). Absolute ¹H B₁⁺ (≥30 Hz) was not a significant predictor for NOE enhancement levels (p>0.11) for any of the metabolites except PC (p=0.04). For the latter, ¹H B₁⁺ only partially explained the between-voxel variations: controlling for this factor still yielded unexplained between-voxel variations of 10% of the total variation.

Discussion

This study was performed to answer the question “should we use NOE in ³¹P MRS(I) or not?”. To a great extent, this question can be answered with yes, but we will discuss several aspects which should be taken into consideration in deciding upon using NOE or not.

Significantly higher ³¹P signals were measured using NOE for most metabolites, while none of the metabolites showed a (significant) decrease in signal. Moreover, the signal enhancement significantly improved fitting quality of the metabolite resonances with a positive NOE enhancement. These favorable aspects of NOE at 7T have been reported before (19,22) and especially the first reason has historically led to the use of NOE in in vivo studies. However, when quantifying NOE enhancements per metabolite by using native and NOE-enhanced ³¹P MRSI data, considerable variation in enhancement values was observed in this study, reproducing findings in many previous studies of different tissues at different field strengths (16–19,21,22). This variation may indicate lack of agreement between both methods, and in unreplicated studies this can be interpreted as NOE adding variability to the data. However, it might also be caused by poor repeatability of (one of) the methods. The design of this study allowed us to put this variation into the perspective of the repeatability of the individual methods. For a good repeatability study, all measurements need to be performed with the same system and method, within a short period of time. This is in contrast to reproducibility studies, in which the variation in measurements is observed under changing conditions (31). Our hypothesis was that variation in NOE enhancement values is caused by the limited repeatability of the underlying methods (native and NOE-enhanced ³¹P MRSI). The measurement uncertainties of the ³¹P metabolite values will propagate in the calculation of the NOE enhancement.

The similarity of the 95% LoA of both methods and the 95% repeatability intervals of each method as well as the dominance of the factor repeat in the NOE enhancement confirmed our hypothesis. The metabolites PC, Pi, GPE, GPC and NAD showed considerable variation in NOE enhancements, but this apparent lack of agreement between the methods could be traced back to the repeatabilities of both methods for these metabolites. It depends upon the application of the method what relative repeatability can be considered sufficient, but for all these metabolites the CoV_repeat was above 10% with native ³¹P MRSI, and improved for all but GPE with NOE. Similarly, the CoV_repeat of ratios containing these metabolites improved with NOE. GPE showed both a higher absolute and relative repeatability with NOE, although a NOE enhancement of 30.1% was observed. We suspect this finding to be incidental, resulting from the limited sample size and low SNR of this metabolite. NOE enhancements are known to be dependent on dipolar relaxation time (15), which might be affected by pH or temperature, but all circumstances were kept the same between the repeated measurements, so it is unlikely that the dipolar relaxation time has changed within a voxel between the two repeated NOE-enhanced ³¹P MRSI measurements. The relative repeatability of metabolites with endogenously higher signal intensities and smaller fitting errors was better (PCr, γATP and αATP) compared to the metabolites with lower signal intensities, even though their absolute repeatability was slightly worse. These metabolites also showed the smallest variation in their NOE enhancement values. The strong linear correlation between quality of fit, expressed by the CRLB, and the relative repeatability indicated that CRLB values present a good measure of relative repeatability of ³¹P MRSI.

PE, and to a smaller extent PC and GPC, showed an improvement in relative repeatability with NOE, but both the Bland-Altman based analysis and the decomposition of variance in NOE enhancements indicated the presence of additional heterogeneity between voxels next to repeatability variations. We tested and excluded local ¹H B₁⁺ to be the reason for additional variation with NOE. A different reason for additional variation between voxels with NOE could be spatial differences in dipolar interaction between the ¹H and ³¹P spins related to tissue type. Because of the large true voxel size in this study, selected voxels contained a mixture of gray and white matter, and they were sometimes also contaminated with skull tissue or cerebral fluid. The diffuse border of the voxels (30,32) precluded analysis of the exact tissue distribution within the spectroscopic voxels by segmentation. To study the possible effect of tissue type on NOE enhancements would require a much higher spatial resolution and thus considerably longer acquisition times.

The effect of proton decoupling on the repeatability of ³¹P MRSI, solely or in combination with NOE, was not investigated in this study. The beneficial effect of proton decoupling decreases with field strength, and it adds considerably to the SAR burden. In this 7 T study we opted to saturate the water resonance to generate the NOE with a strategy well-known from decoupling; the WALTZ-4 irradiation scheme. This irradiation scheme provides a relatively broadband saturation, which reduces the sensitivity of water saturation to local B₀ inhomogeneities. At clinical field strengths one could also use continuous wave irradiation to saturate the water resonance. Since the power needed for water saturation with WALTZ-4 contributed significantly to the SAR in our study, it would be of interest to determine if a lower ¹H B₁⁺ would be sufficient for in vivo experiments compared to phantom experiments.

In the current study, we homogenized B₀ and ¹H B₁⁺ in the posterior part of the brain only, whereas for clinical studies homogenization over the whole brain is probably more often required. At clinical field strengths the ¹H B₁⁺ and B₀ distributions are intrinsically more homogeneous than at 7 T, facilitating this prerequisite. Other coil configurations or parallel transmit techniques combined with tailored RF pulses may also increase the size of the ¹H region that can be excited homogeneously at 7 T (33). Automatic scanner adjustments of clinical MR scanners obviate the need to calibrate the ¹H pulse amplitudes for NOE, which was a relative time-consuming procedure in this study. The minimum ¹H B₁⁺ for constant NOE using a specific saturation strategy should however be determined in a pre-study. In cases where non-uniformities in the ¹H RF field are expected, e.g. when using ¹H surface coils or at ultra-high field strengths, careful examination of the field distribution and amplitude is needed before employing NOE. The necessity of ³¹P flip angle calibration in every single subject has to be assessed in a larger cohort. In the 7 subjects within this study, the ³¹P B₁⁺ was very similar. It is very common to use a pre-determined, constant B₁⁺ for multinuclear experiments at clinical field strengths, or a pre-determined fixed multiplication factor between the X-nucleus and the automated ¹H B₁⁺ calibration, because of the lack of automated procedures for multinuclear coils.

The results of this ³¹P MRSI study in the brain at 7T can guide human ³¹P MRS(I) investigations of the brain and of other tissues including those performed at different field strengths. The improvement of relative repeatability of ³¹P metabolites with positive NOE enhancements allows detection of smaller differences in metabolite intensities or ratios. This is of interest in follow-up studies or in the comparison of groups (e.g. healthy vs. patient). NOE-enhanced metabolite ratios obviously differ from native ratios, but they are more repeatable. When reporting NOE-enhanced metabolite ratios, it is of importance to clearly describe the way NOE is generated. This allows others to perform similar experiments, thus enabling between-site comparisons. When absolute quantification of ³¹P metabolite concentrations is preferred, it has to be realized that data should be corrected for the NOE enhancement, which introduces an additional source of variation to the quantification. Although it is known that NOE enhancements can vary between tissues (14,16–22), it remains uncertain if variations within an organ can exist as well. As discussed above we found an indication that this might be the case for NOE enhancements of PE, PC and GPC. It is similarly unknown if NOE is different in diseased tissue compared to normal tissue or during activation (exercise), or if there are changes in NOE with age. In diseased tissue, T₁ relaxation times of ³¹P may change (34), which may have an effect on the absolute NOE enhancements. A larger repeatability study with a more heterogeneous study population, including patients, would be needed to confirm if the above-mentioned possible differences in NOE can be distinguished from the measurement uncertainties of native and NOE-enhanced ³¹P MRSI. It would also be of interest to study reproducibility of NOE enhanced ³¹P MR data by measuring subjects on different days (20,35). This allows calculation of sample sizes needed to reveal significant differences in certain metabolite ratios for paired (e.g. response to therapy) and independent (e.g. patient vs. control) studies.

In conclusion, the signal enhancement generated by NOE improved relative repeatability of brain ³¹P MRSI in healthy volunteers. Variations in NOE enhancements per metabolite could almost completely be explained by the repeatability of native and NOE-enhanced ³¹P MRSI. For these reasons, the use of NOE-enhanced ³¹P MRSI is encouraged.

Abbreviations

ATP

adenosine triphosphate

B₀

magnetic field

B₁⁺

transmit radiofrequency amplitude

CoV

coefficient of variation

CRLB

Cramér-Rao lower bound

$\stackrel{‒}{d}$

mean bias between native and NOE-enhanced measurements

η

NOE enhancement

GPC

glycerophosphocholine

GPE

glycerophosphoethanolamine

LoA

limits of agreement

NAD

nicotinamide adenine dinucleotide

${\overline{native}}_{repeat}$

mean integral of non-enhanced repeated measurements per voxel

NOE

nuclear Overhauser effect

${\overline{NOEenh}}_{repeat}$

mean integral of NOE-enhanced repeated measurements per voxel

MRSI

magnetic resonance spectroscopic imaging

μ

mean metabolite integral over all voxels and measurements

PC

phosphocholine

PCr

phosphocreatine

PE

phosphoethanolamine

Pi

inorganic phosphate

RC

repeatability coefficient

RF

radiofrequency

SAR

specific absorption rate

σ_d

standard deviation of differences between methods

σ_repeat

standard deviation of repeated measurements

SNR

signal-to-noise ratio

WALTZ

wideband alternating-phase low-power technique for zero residual splitting