Discrimination between neurochemical and macromolecular signals in human frontal lobes using short echo time proton magnetic resonance spectroscopy

Magnetic resonance spectra from large (35 cm3) frontal lobe voxels in vivo were analyzed using LCModel, with and without subtraction of a “metabolite nulled” spectrum with an inversion time of 650 ms to characterize the macromolecule baseline.

Baseline subtraction decreased the signal to noise ratio (SNR), but improved the reliability of LCModel quantification of most metabolites, as reflected in the Cramer–Rao lower bounds, in particular for glutamate and glutamine.

The reported concentrations increased for glutamine, creatine, and lactate, and decreased for glutamate, myo-inositol and NAAG, but the sum of all metabolites remained constant, as did the standard deviation of the concentrations in the control group.

Macromolecule subtraction is worthwhile when SNR is high, as in the characterization of normal-appearing tissue in the brain.


Proton magnetic resonance spectroscopy can estimate the in vivo concentration of several important chemicals in the brain.

Most studies to date have been performed at long echo times, at which signal generally remains from only three main metabolites: N-acetyl aspartate (NAA), which is thought to be a marker of neuronal numbers and function; the sum of creatine and phospho-creatine (Cr), which are central to energy homeostasis; and choline-containing compounds (Cho), which probably are largely related to membrane function.

N-acetyl aspartyl glutamate (NAAG) – a putative neurotransmitter – may be detected, in white matter in particular, although it is difficult to resolve from NAA.

If present at sufficient concentration in the brain (usually only in pathological states), then lactate (Lac) is also visible at long echo times.

At short echo times, signal is detectable from several additional metabolites of interest: myo-inositol (Ins), a possible marker of glial cell number and function; glutamate (Glu), a key amino acid neurotransmitter; and glutamine (Gln), a closely related amino acid also linked to neurotransmitter metabolism.

Elevations in the combined signal of glutamate and glutamine (Glx) have been shown in some patients with epilepsy,1–3 and may be an indication of increased cortical excitability.

This is potentially of interest not only in understanding the pathophysiology of epilepsy but also in suggesting suitable drug therapies and in monitoring patient progress.

However, the signals of glutamate and glutamine overlap to a large extent both with each other and with macromolecules such as proteins, making their reliable estimation very difficult.

At 1.5 T, complete spectral resolution of glutamate and glutamine signals is impossible, so spectral modelling is needed.

LCModel is a commercially available fitting package, which estimates the concentration of chemicals in the brain with reference to their individual spectra acquired using highly concentrated solutions.4

This provides estimates of the separate concentrations in vivo of Glu and Gln, and of their combined signal Glx, along with estimates of the uncertainty in these concentrations, based on Cramer–Rao lower bounds.

The reliability of measurement of Glx is typically better than that of its individual components, and glutamine in particular would be considered generally to be unreliably determined in vivo.

The macromolecular baseline signal, which has broad peaks appearing at roughly 0.9, 1.3, 2.4, and 3.2 ppm,5 is a major limitation of the in vivo application of short echo time proton spectroscopy.

These resonances can comprise 30% of the total signal, and spectral modelling is generally unable to find a unique solution to discriminating between these broad signals and the superimposed sharper peaks from small metabolites.6

One method to discriminate signals from macromolecules and neurochemicals is through their different spin-lattice (T1) relaxation times.

Macromolecules are relatively immobile, and therefore their protons have much shorter T1 (approximately 200 ms at 1.5 T 7) than those of small metabolites (between 1 and 1.5 s at 1.5 T 8).

Therefore, spectroscopic acquisition can be preceded by an inversion pulse with a delay chosen to null the signal from one while maintaining signal from the other.

It is possible to suppress the lipids and macromolecules directly, using an inversion pulse with a short delay.9,10

However, this also reduces signal from small metabolites, and may reduce them differently depending on their differing T1.

It may be possible to optimize lipid suppression somewhat, for example by using a train of inversion pulses,11 but there is always an accompanying reduction in signal-to-noise ratio (SNR) of the small metabolites.

Another problem is that it can be difficult to integrate lipid/macromolecule suppression with water suppression: to attain shorter inversion times, it may be necessary to reduce the number or pulse width of water suppression pulses, leading to less thorough suppression or to poorer frequency specificity, respectively.

A better approach is to null the signal from small metabolites while maintaining signal from macromolecules.

The fraction of signal remaining (S/S0) at each inversion time (TI) is calculated from the estimated T1 of each species and the repetition time (TR) as: S/S0 = 1 – 2 exp(–TI/T1) + exp(–TR/T1)The resulting macromolecule-weighted spectrum can be subtracted from the original trace to leave a spectrum with only signal from small neurochemicals remaining.

Although this method of metabolite nulling has been demonstrated to be capable of separating out the macromolecule signal,5,12,13 it is not widely used in vivo, due to the doubled scan time, and the reduction in signal-to-noise ratio inherent in performing a subtraction.

However, when estimating cerebral concentrations of especially glutamate and glutamine, signal-to-noise is not as limiting as macromolecule signal overlap, so the benefits of this approach may outweigh the drawbacks.

We investigated whether metabolite-nulling improved the reliability of metabolite concentration estimation using the LCModel program, as measured by both the reported uncertainty of estimation and by the spread of concentration estimates in a control group.


The study was performed on a 1.5 T General Electric Signa Horizon Echospeed scanner (Milwaukee WI, USA) using the standard birdcage head coil.

Axial T1-weighted inversion–recovery prepared fast spoiled gradient echo images were obtained, both to guide voxel prescription and subsequently for segmentation using SPM99 (Statistical Parametric Mapping; Wellcome Dept.

Of Imaging Neuroscience, Institute of Neurology, London) to determine the fractional content of grey matter, white matter and cerebrospinal fluid (CSF) in the spectroscopic voxels.

The concentrations of all metabolites were divided by the estimated fractional brain tissue content of the voxels, to correct for the presence of CSF, as previously described.14

Voxels approximately 4.0 × 3.5 × 2.5 cm in size (Fig. 1) were studied in both frontal lobes of 10 normal volunteers.

This study was approved by the Joint Research Ethics Committee of the National Hospital for Neurology and Neurosurgery and the Institute of Neurology and all subjects gave informed consent.

Point-resolved spectroscopy (PRESS) localization was used to select the spectroscopy volume, with an echo time of 30 ms and a repetition time of 3 s.

Water signal was suppressed using 3 chemical-shift selective (CHESS) pulses in both cases, with the flip angle of the final pulse tailored to ensure slight under-suppression.

32 transients each were collected, with and without the inversion, plus 16 transients without CHESS, for a total scan time of approximately 5 min per lobe.

The optimum inversion pulse delay was calculated from eqn. (1) to be about 750 ms, assuming a range of in vivoT1 between 1.0 and 1.5 s and a TR of 3 s.

This was checked in a pilot in vivo study.

A 40 cm3 volume of interest was placed along the midline in one subject, at a level similar to Fig. 1, including equal amounts of tissue from the right and left frontal lobes in case there was a difference in the relaxation times between them.

Spectra with inversion times of 550, 650, 750, 850, and 950 ms were acquired and compared to a spectrum with no inversion.

The inversion time giving the minimum residual metabolite signal was used for all subsequent acquisitions.

Spectra were analyzed using LCModel, with and without subtraction of the metabolite-nulled spectrum.

The same reference set of spectra from metabolite solutions was used for both analyses, as well as the same calibration factor.

Comparisons were made between left and right frontal lobes, and between raw and subtracted data, using paired t-tests in SPSS 9.0.

The average of the standard deviations (SD) reported by LCModel (a reflection of the Cramer–Rao bounds) on the left and right were also compared between raw and subtracted data.


An inversion pulse delay of 650 ms was chosen to null the signal from small metabolites, based on the pilot series of acquisitions (Fig. 2).

Of the inversion times shown, 750 ms was too long, since positive peaks were clearly visible from NAA, Cr and Cho, whereas 550 ms was too short, since negative peaks appeared for NAA and Cr.

A good compromise appeared to be 650 ms, which showed a fairly smooth baseline of broad components.

Some residual positive peak of NAA appeared to remain, but this was not seen in all subjects.

This optimum inversion time was somewhat shorter than predicted by theory; eqn. (1) would suggest 750 ms to be better for the range of in vivoT1 expected.

This discrepancy may perhaps be due to inversion pulse imperfections; or the metabolites of interest may have somewhat shorter T1 relaxation times than previously reported.

In particular, the protons of small metabolites not forming the large singlet peaks may have T1 shorter than 1 s.7

The metabolite-nulled spectra consisted of broad signals similar to those previously observed and assigned to macromolecules in vivo,5,7 and the average metabolite-nulled spectra from 10 control subjects was found to differ significantly on the left and right (Fig. 3).

Narrow peaks from the small metabolites were effectively nulled on both sides: the main difference between left and right appeared to be an increase in the broad signal around 1.5 ppm on the subject's left compared to the right.

In all cases a smoother baseline was fitted when macromolecules were subtracted (Fig. 4).

In Fig. 4A, a typical constrained C-spline fit to the baseline is shown.

This is similar to the observed macromolecule signal, but clearly had uncertainty associated with it.

In the subtracted spectrum, the baseline only included residual water, which can be modelled relatively easily, giving a more consistent result.4

The SNR was reduced by the subtraction by the expected factor of √2, from a mean of 36 to 26, but remained sufficient for quantification.

The reported standard deviations given by LCModel, a reflection of the Cramer–Rao lower bounds, were reduced significantly by baseline subtraction for almost all metabolites, suggesting that quantification had been improved (Table 1).

The improvement was most marked for glutamate and glutamine: for Gln, the SD reported by LCModel changed from 31%, which would be considered unreliable, to a more acceptable 19%.

For Glx the SD was nearly halved, going from 12% to just over 6%.

The spread of measurements in the control group, however, was not much affected by the subtraction of metabolite-nulled data.

Some concentrations were increased (Gln, Cr, Lac) and some were reduced (NAAG, Glu, Ins), but the standard deviations of the measurements for each metabolite across the subject group remained fairly stable, as well as the summed concentration of all metabolites measured.

This suggests that although LCModel's C-spline estimate of baseline is not perfect, it tends to behave consistently in the presence of macromolecule.

The only systematic difference between the left and right was in Glx, and this disappeared when the metabolite-nulled baseline was subtracted.

Glx showed a reduction in the SD of the control group as well as in the LCModel SD estimate, confirming that the quantification reliability was improved.

Lactate results in controls were still unreliable as estimated by LCModel; however, use of metabolite nulling did increase conspicuousness of Lac (Fig.3) and may therefore allow better determination where it is pathologically elevated.


“Metabolite-nulled” spectra were acquired and analyzed, to investigate the macromolecular signal in magnetic resonance spectra at short echo times and its effect upon quantification of the small molecular weight metabolites of interest.

Metabolite spectra were analyzed using LCModel with and without the subtraction of the macromolecule baseline, and the concentrations and reliability estimates were compared.

Similar approaches attempting to improve LCModel fitting of metabolites have been adopted recently by other groups.15–17

These merit particular discussion, so the methods they employed will first be described.

All have attempted to incorporate macromolecule signals into the LCModel basis set, which has the major advantage that the lengthy acquisition of baseline data may not be necessary in every subject studied.

The first approach was the incorporation into the basis set merely of simulated lipid peaks at 0.9 and 1.3 ppm.15

This was shown to aid the discrimination of lipid from baseline, and to improve estimates of lactate and alanine in vivo.

This may be a particularly advantageous approach in tumours, because lipid, lactate and alanine are often elevated, and these elevations can be significant for differential diagnosis.

In addition, tumour lipid peaks are more mobile than those in normal brain and scalp, giving them a longer T1, so they are not as easily distinguished from small metabolites using T1-based techniques.

However, this technique does not much aid the measurement of glutamate, glutamine, and the other small metabolites overlapped by the other macromolecule regions, either in controls or in conditions such as multiple sclerosis where there may be pathological increases in macromolecule signal throughout the spectrum.18,19

No metabolite results in normal controls were shown using this method,15 so no direct comparison with the current study is possible.

The second approach was to include the average experimentally determined macromolecule baseline as a single element in the LCModel basis set.16

This has the advantage that the broad macromolecule signals which overlap with small metabolites are somewhat constrained by those that do not, but the disadvantage that the small (but not completely negligible) inter-subject variation in macromolecule signal cannot be accommodated.

LCModel requires the relative peaks of a metabolite (or in this case pseudo-metabolite) to be fixed absolutely, allowing no method to compensate for this variation.

However, Hofmann et al. achieved quite good reproducibility of macromolecule signal among their control subjects by using the more time-consuming but more thorough saturation recovery method to characterize them,16 which should have minimized this drawback.

Another methodological difference with the current study is that the echo time was 20 ms instead of 30 ms.

This gave higher, and presumably more accurate, metabolite concentration estimates in vivo, since the spin–spin (T2) relaxation times in vivo are shorter than in the phantom solutions of the basis set, giving more T2 weighting in the current study.

However, different radiofrequency pulses and gradients are needed for such short echo times, which gives less sharp voxel profiles and greater chemical shift displacement error.

This may have increased contamination with lipid and macromolecule signal from outside the voxel, or indeed outside the brain.

The third approach was to include a series of independently varying parameterized peaks in LCModel to account for the macromolecules and lipids.17

This is better able to accommodate inhomogeneous pathological variation in these signals.19,20

However, when they are allowed to vary independently, there is no way to completely constrain the macromolecule resonances which underlie the small metabolite signals at about 2 and 3 ppm.

The acquisition methods for that study also differed from the current study: the pulse sequence used for localization was stimulated echo (STEAM) instead of PRESS, the echo time was 15 ms, and the TR was 1.5 s.

This gave lower concentration estimates than in the current study, probably because of T1 weighting of the spectra.

The shorter echo time also may have led to increased macromolecule signal due to decreased T2 weighting.

Short echo times are more easily combined with sharp voxel definition in STEAM than PRESS, so less of the macromolecule signal should have arisen from outside the voxel; and in addition, stronger outer-volume suppression pulses were used.21

However, a trade-off remains between strong suppression of the outer volume and oversuppression of the inner volume, which depends on the size and position of the voxel and of the suppression bands.

Mader et al22. calculated a suppression factor within the voxel ranging from 3% for a white matter voxel to 19% for the pons when the extra-strong outer-volume suppression was used.

Metabolite concentrations

Macromolecule subtraction has been shown to improve somewhat the reliability of quantification of small metabolites in large frontal lobe voxels in vivo, in particular glutamate and glutamine.

The reduction in the estimated uncertainty for most metabolites, despite reduced SNR, confirms that the baseline is a major source of error in the fitting of in vivo signal.

However, the absence of large changes in the estimated concentrations of most metabolites, even glutamate + glutamine, confirms that the C-spline baseline used by LCModel performs surprisingly well.

With the method employed here, the summed metabolite signal stayed constant: the sum of metabolites on the left and right in Table 1 is 47.9 mM for the raw spectra, and 47.7 mM following baseline subtraction.

Hofmann et al16. found a global reduction in the estimated concentrations of nearly all the major small metabolites by about 10% on average when macromolecule signal was included in the basis set.

Of the metabolites in Table 1, only Cho in the white matter was not reduced significantly in that study by the incorporation of macromolecule into the basis set.

Seeger et al17. had a less significant reduction in metabolites when macromolecule was incorporated, about 5% on average, with only NAA + NAAG and Cho reaching significance (and with Cr, as in the current study, being significantly elevated).

It seems that when LCModel is provided with very broad signals in the basis set, there may be a tendency to overfit them to the data.

This is a major potential drawback for methods incorporating macromolecule signal into LCModel.

Hofmann et al.,16 Seeger et al.,17 and the current study all showed a significant reduction in NAA + NAAG when methods were incorporated to account for macromolecule.

This may seem surprising, since quantification of NAA and NAAG is based mainly on the most prominent peak in the spectrum, which should be relatively unaffected by the baseline.

Looking at the NAA and NAAG separately, Hofmann et al16. concur that the dominant effect is on NAAG, which was reduced by 21% in white matter and 45% in grey matter.

This suggests that in the raw in vivo spectra LCModel fitted some macromolecule signal as NAAG.

The other metabolites with prominent peaks were also affected by macromolecule correction methods, although to a lesser extent.

Myo-inositol showed a trend of reduction by about 5% in all 3 studies.

Cho was reduced by about 5% in the current study (non-significantly) and by a similar amount in Seeger et al.In Hofmann et al., it was unchanged in white matter and increased by about 15% percent in grey matter.

Cr was increased by about 5% by macromolecule correction in the current study and in Seeger et al., and decreased by about 5% in white matter in Hofmann et al. Probably these small differences indicate that the C-spline baseline fitted by LCModel does not precisely match the macromolecule signal, although the approximation is close.

Therefore, macromolecule corrections may improve the accuracy of measuring these “conventional” metabolites at short echo times also.

The increased conspicuousness of Lac following subtraction of the macromolecule baseline may allow better determination where it is pathologically elevated.

One caveat is that if Lac had an especially long T1 it would be overestimated due to subtraction of negative signal at an inversion time of 650 ms.

Hofmann et al16. and Seeger et al17. found that incorporating macromolecule signal into LCModel reduced the estimated Lac instead of increasing it as in the current study: it is unclear which method may be more accurate.

However, even if it were less accurate, metabolite-nulling may be preferred as being more precise: it was found to be invaluable in distinguishing between pathologically elevated lipids and abnormally high lactate in multiple sclerosis lesions.19

Although we found no significant effect of macromolecule subtraction on the estimated Glx (glutamate plus glutamine), there was redistribution between these overlapping peaks, with a reduction in the estimated glutamate and an increase in glutamine.

The elimination of broad resonances underlying the peaks would be expected to have aided in this discrimination; but the lack of a definitive gold standard makes it impossible to judge which is more accurate.

Hofmann et al16. showed large reductions in both glutamate and glutamine, again suggesting possible overfitting of macromolecule.

Seeger et al17. didn't report glutamate and glutamine separately, but found no significant change in Glx.

It is unclear why Glx was not affected by the incorporation of a broad unconstrained macromolecule peak at the same frequency in the basis set.

From the same group, Mader et al22. found that Glx interfered with macromolecule estimation at 2.1 ppm, so it seems likely that macromolecule would interfere with Glx estimation.

This inability to discriminate between macromolecule and Glx argues that the acquisition of metabolite-nulled data may be necessary to reliably determine Glx.

In the past we have demonstrated elevations in Glx in temporal lobe epilepsy1 and idiopathic generalized epilepsy.3

In future, we will apply these methods of metabolite-nulling both to ensure that this elevation is not due to artifactual macromolecule signal, and also to attempt to estimate the individual concentrations of glutamate and glutamine.

Previous studies have shown elevations in glutamine alone in epilepsy patients taking vigabatrin23 or valproate13.

Macromolecule signals

We showed a marked left–right difference in the macromolecule signal recorded, with greater signal around 1.5 ppm on the subject's left.

A very similar difference has been shown between spectra from predominantly grey and predominantly white matter in vivo, with the greater signal around 1.5 ppm in grey matter voxels being ascribed to the tissue content difference.16

In the current study, the estimated percentage content of grey matter was similar on the two sides (30 ± 3% on the left; 29 ± 5% on the right), so tissue content can be excluded as the source of this.

The difference was more likely to be due to differences in the amount and phase of macromolecule signal contamination from outside the gradient-selected region of interest.21,24

This can cause a consistent directional difference because the gradient used to select the volume of interest always varied from negative on the left to positive on the right.

A similar effect may possibly have been acting in the Hofmann study:16 more signal at 1.5 ppm was seen in the grey matter voxel, which was positioned in the occipital lobe at the posterior of the brain, than in the white matter voxel, which was located anteriorly in the centrum semiovale.

On GE 5X Signa scanners, posterior is conventionally the negative gradient direction, although whether this was so in Hofmann et al. is not known.

In both the current study and in Hofmann's work, standard outer-volume suppression pulses available on a GE 5X Signa scanner were used.

With the use of stronger outer volume saturation,22 more macromolecule signal at 1.5 ppm was also found in grey matter than in white.

However, caution is needed in interpreting subtle changes in macromolecule as true physiological differences, because in the neocortex, grey matter voxels are by definition almost always further towards the periphery of the brain: closer to the scalp and further from the magnet isocentre.

A better method to investigate grey/white matter differences in macromolecule content may be magnetic resonance spectroscopic imaging.25

Contamination with extra-cerebral signal remains a problem, but the spatial encoding helps identify the origin and extent of the contaminating signal; and the acquisition of data from multiple voxels containing variable tissue content may enable the more effective modelling of macromolecule content in grey and white matter.

Hwang et al25. collected macromolecule-weighted MRSI, and although they did not investigate the tissue content dependence of macromolecule signal per se, they commented upon finding a uniform distribution of macromolecule signal intensity within the excited region of interest, implying that no such tissue content dependence existed.

In one of our own MRSI studies at an echo time of 30 ms,14 we showed spectra obtained from voxels containing mostly grey matter and mostly white matter.

Neither showed appreciable macromolecule signal at 1.5 ppm, although signal at this frequency was excited in the scalp well outside the PRESS-selected region of interest.

When lesions are investigated, they are likely to be distributed heterogeneously in different patients, so the macromolecule signal is likely to suffer even more from variable inclusion of extra-voxel contamination in patients than in controls.

In an elegant demonstration that lipids detected in a multiple sclerosis lesion were not caused by scalp contamination, Davie et al18. studied a control voxel located in normal-appearing tissue between the lesion voxel and the scalp, and verified that the lipid/macromolecule signal was not elevated at that location compared to controls.

This sort of validation is vital to enable sound interpretation of macromolecular signal in vivo; we have chosen instead to regard macromolecular signal as largely artifactual, and to attempt to minimize its effect on metabolite analysis.

In summary, there are several methods available to investigate the distinct contributions of macromolecules and small metabolites to in vivo proton spectra.

Which of these is the best solution to the problem depends on which metabolites are of most interest, and on the specific pathology.

Methods have been proposed for modelling macromolecule signal without the separate acquisition of metabolite-nulled data; but none has yet demonstrated convincing improvement in the quantification of glutamate and glutamine, which are metabolites of central interest in seizure disorders.

Acquisition of a separate metabolite-nulled dataset is time-consuming, but not prohibitively so if fairly large volumes of interest are used, as in characterizing normal tissues.

It is likely to be less useful in studying small focal lesions than in studying larger lesions or more widespread abnormalities.