- Top of page
- MATERIALS AND METHODS
LCModel and AMARES, two widely used quantitation tools for magnetic resonance spectroscopy (MRS) data, were employed to analyze simulated spectra similar to those typically obtained at short echo times (TEs) in the human brain at 1.5 T. The study focused mainly on the influence of signal-to-noise ratios (SNRs) and different linewidths on the accuracy and precision of the quantification results, and their effectiveness in accounting for the broad signal contribution of macromolecules and lipids (often called the baseline in in vivo MRS). When applied in their standard configuration (i.e., fitting a spline as a baseline for LCModel, and weighting the first data points for AMARES), both methods performed comparably but with their own characteristics. LCModel and AMARES quantitation benefited considerably from the incorporation of baseline information into the prior knowledge. However, the more accurate quantitation of the sum of glutamate and glutamine (Glx) favored the use of LCModel. Metabolite-to-creatine ratios estimated by LCModel with extended prior knowledge are more accurate than absolute concentrations, and are nearly independent of SNR and line broadening. Magn Reson Med 51:904–912, 2004. © 2004 Wiley-Liss, Inc.
In clinical magnetic resonance spectroscopy (MRS) of the brain, short echo times (TEs) are employed to optimize the signal-to-noise ratio (SNR), reduce signal attenuation due to transverse relaxation and scalar coupling, and enable the quantification of more than the three dominant singlet resonances (i.e., N-acetylaspartate (NAA), creatine (Cr), and choline (Cho)). Generally, it is difficult to quantify short-TE spectra because of the overlapping metabolite signals and the contribution of macromolecule and lipid components. In addition to software of MR tomographs and various in-house developments at research sites (1–3), two sophisticated and well documented software packages are widely used: LCModel (4) and Magnetic Resonance User Interface (MRUI) (5). These software packages are used worldwide by many groups not only because of their availability and performance, but also because they provide results with a broader basis of comparability. The commercially available software package LCModel (6) fits spectra in the frequency domain using a basis set of spectra of in vitro metabolite solutions acquired under conditions identical to those under which in vivo data are acquired. AMARES (7), which is part of the MRUI package (5), has the advantage of being free of charge to nonprofit organizations. This advanced quantitation toolbox analyzes spectra in the time domain utilizing a priori information that can be introduced flexibly. LCModel employs a “black box” approach, and thus requires less user interaction than AMARES.
For application purposes, however, it is important to determine how quantitation depends on linewidth and SNR, and how the two methods handle the broad macromolecular and lipid signal contributions. Of course, the best way to account for macromolecular signal contribution is to acquire the macromolecular spectrum by inversion recovery (8) or saturation recovery (9) prepared measurement schemes. However, in clinical MRS, the acquisition of the macromolecular spectrum is impractical, because it is not implemented as a standard in MRS pulse sequence libraries, and, more importantly, it is time-consuming. Therefore, the ability to achieve reliable spectrum quantification in the presence of a large (and at best approximately known) macromolecular baseline is of vital importance for routine in vivo MRS. LCModel and AMARES exploit the short T2 of the macromolecular components to account for this lack of information. Whereas LCModel fits a smooth spline as background signal, AMARES allows the influence of the first time domain data points to be weighted to the fit.
Although published macromolecular spectra patterns and data (8–12) are neither identical nor well classified, they show great similarities. The same is true for lipid resonances that are usually barely present in healthy brain tissue spectra. They may arise from residual out-of-volume excitation (13), or they may be strongly elevated in brain diseases, such as tumors. The similarities motivated the integration of spectral baseline information into the prior knowledge and, indeed, in the case of LCModel, it was reported that the fit was improved by the addition of macromolecular and/or lipid signals to the basis set (11, 12, 14, 15). To the best of our knowledge, no such approach has yet been realized for AMARES. In addition to the strongly reduced residuals in the lipid region of spectra with high lipid content (12, 15), reported improvements of LCModel quantitation are based on decreased Cramér-Rao lower bounds (14), or decreased variances for in vivo spectra (11). The enhanced precision for prominent metabolites underlines the effectiveness of this approach; however, for a more comprehensive characterization of quantitation software, it is also desirable to determine accuracy. The simulation approach only addresses the quantitation part and avoids the multiple elements of uncertainty during the acquisition part, such as drift of scanner performance and patient motion.
In this work we investigated the influence of SNR, linewidth, and baseline information on the quantitation of short-TE brain spectra by LCModel and AMARES. We describe the preparation of the simulated spectra and the settings for LCModel and AMARES, and present results for absolute concentrations, metabolite-to-creatine ratios, separate quantitation of glutamate (Glu) and glutamine (Gln), and quantitation of macromolecular constituents. The accuracy and precision of the data are then discussed within the framework of the SNR and linewidth, the quantitation approach applied, and the simplification and idealization due to the simulation. Finally, practical guidelines are given.
- Top of page
- MATERIALS AND METHODS
For the sake of comparability, all concentrations and metabolite-to-creatine ratios were normalized to one. In Figs. 2 and 3, the means and SD of absolute quantitation are summarized for the four different fit strategies. The influence of decreasing SNR on the quantitation is shown in Fig. 2 for 4 Hz line broadening. This Gaussian broadening led to FWHM of singlet peaks slightly larger than 4 Hz, which is an accepted level of quality for in vivo spectra of brain tissue at 1.5 T. The effects of increasing linewidths are elucidated in Fig. 3 for an SNR of 25. The metabolite-to-creatine ratios for the concentrations displayed in Figs. 2 and 3 are presented in Figs. 4 and 5. In Figs. 2–5, only data for the sum of Glu and Gln (abbreviated as Glx) are shown. Separate Glu and Gln values are shown in Fig. 6.
Figure 2. Normalized absolute intensities of the metabolites for 4 Hz Gaussian line broadening vs. SNR determined with four different quantitation schemes. The means and SDs of 100 simulations are shown. The linewidth is slightly larger than the applied line broadening. The SNR is the ratio of maximum amplitude in the spectrum and the SD of noise. See text for a detailed explanation.
Download figure to PowerPoint
Figure 3. Normalized absolute intensities of the metabolites for an SNR of 25 vs. line broadening determined with four different quantitation schemes. The means and SDs of 100 simulations are shown.
Download figure to PowerPoint
Figure 4. This figure corresponds to Fig. 2 and shows normalized metabolite-to-creatine ratios of the metabolites for 4 Hz Gaussian line broadening vs. SNR determined with four different quantitation schemes. The means and SDs of 100 simulations are shown. See text for a detailed explanation.
Download figure to PowerPoint
Figure 5. This figure corresponds to Fig. 3 and shows the normalized metabolite-to-creatine ratios of the metabolites for an SNR of 25 vs. line broadening determined with four different quantitation schemes. The means and SDs of 100 simulations are shown.
Download figure to PowerPoint
Figure 6. Absolute intensities for Glu, Gln, and Glx. The vertical axis is compressed compared to previous figures. Shown are the means and SDs of 100 simulations for 4 Hz Gaussian line broadening vs. SNR, as determined with LCModel with macromolecules and lipids in the basis set. The significantly improved precision of Glx vs. Glu or Gln data is obvious.
Download figure to PowerPoint
At first glance, it is obvious that the incorporation of macromolecules and lipids in the prior knowledge of LCModel and AMARES significantly improves the accuracy of the results compared to the spline fit of the baseline in LCModel and the weighting of the first data points in the AMARES fit. For a more complete description, however, the dependency of the SD and of the deviation of the means from the true values (i.e., precision and accuracy, respectively) on SNR and Gaussian line broadening must be inspected in more detail.
The top row in Fig. 2 shows the means and SD of concentrations determined with the complete basis set. Even with high SNR and an approximately 4 Hz linewidth, the mean overestimation of concentrations is about 4%, which varies only slightly with SNR. The reason for this inaccuracy is not only line broadening but also noise, because the corresponding plots of Fig. 3 also show overestimation even when no broadening is applied. Only under unrealistic conditions (i.e., SNR of about 800, and 1.2 Hz linewidth of singlet signals), metabolic concentrations are estimated with an accuracy of <0.5% for NAA, Cr, and mI, and 2% for Glx. This residual inaccuracy is caused by the signal of macromolecules and lipids, since the fit is perfect (i.e., accuracy better than 0.05%) if the baseline part is removed from this simulated spectrum. If the baseline is fitted with a spline function instead of specific lipid and macromolecule prior knowledge, the overestimation of the metabolites is even more pronounced, especially for Glx, as shown in the second row of Fig. 2.
The moderate variation in accuracy with SNR and Gaussian broadening (as shown in Figs. 2 and 3, respectively) is somewhat reduced with the extended basis set. The most prominent difference between Figs. 2 and 3 is the characteristic of the SD. However, taking the acquisition times into account, this behavior suggests only minor adverse effects due to line broadening.
The surprisingly good and nearly SNR-independent accuracy of all of the metabolite-to-creatine ratios when the extended basis set was used is depicted in the top row of Fig. 4. The second row shows that the use of a spline baseline instead of lipid and macromolecule signals led to less reliable results. As demonstrated in Fig. 5, the dependence of the metabolite-to-creatine ratios on line broadening is small. Because the high level of accuracy of metabolite-to-creatine ratios is based on equal trends in accuracy of concentrations, and not on exact absolute quantitation, metabolite-to-creatine ratios should be favored if accuracy is of major concern.
Concentrations obtained with AMARES also have significantly improved accuracy if macromolecular and lipid information is part of the a priori knowledge. Absolute concentrations are accurate at a 2% level for NAA, Cr, Cho, and mI, but only at a 10% level for Glx. Except for the Glx data points at SNRs of 12.5 and 25 estimated with complete prior knowledge, almost no variation of means with SNR can be observed. The concentrations of NAA, Cr, and Cho in Fig. 3, which were calculated with complete prior knowledge, exhibit a trend from a systematic underestimation to quasi-exact values with increasing Gaussian broadening. This behavior is a consequence of the inadequacy of the Gaussian lineshape model for fitting Voigt lineshapes (19). The high accuracy of the NAA, Cho, and mI concentrations is maintained for metabolite-to-creatine ratios, but Glx accuracy remains disappointing. The strong variation of metabolite-to-creatine ratios with applied broadening in the bottom row of Fig. 5 uncovers the limited capability of the frequency-selective option for handling large broad baseline signals.
LCModel vs. AMARES
A comparison of LCModel without macromolecule and lipid signals in the basis set, with AMARES running with the frequency-selective option (i.e., the second and bottom rows in Figs. 2–5), reveals results of comparable quality. AMARES gives more exact NAA and Cr concentrations, and LCModel offers more accurate Glx/Cr ratios. As demonstrated in the top and third rows of Figs. 2–5, including a priori information about the baseline leads to more accurate absolute values with both methods, with about the same precision level for NAA, Cr, Cho, and mI. However, while LCModel results in a slight overestimation, AMARES slightly underestimates the concentrations. Glx concentration is definitely more precisely and accurately determined by LCModel. The metabolite-to-creatine ratios show the same general behavior as the concentrations. LCModel is superior to AMARES for determining Glx/Cr ratios.
Glu and Gln quantitation is one important reason for acquiring short-TE spectra. The plot of Glu and Gln concentrations in Fig. 6 for LCModel quantitation with macromolecules and lipids in the basis set (i.e., the case with best Glx quantitation) shows a severe deterioration of precision (an average 77% for Glu as compared to Glx). Accuracy is also slightly degraded. Hence, in the light of the data in Fig. 6—especially the very low precision of Gln quantitation—the value of reporting Glu and Gln data instead of Glx data remains questionable.
Although it was not the main purpose of the simulations, the incorporation of macromolecular and lipid information into the a priori knowledge provides data regarding these constituents. The normalized absolute intensities given in Fig. 7 exhibit poor accuracy and very low precision, except for mm1 and mm2 in the AMARES data, and mm2 fitted by LCModel. Although lipid content was held on a small level in the simulated spectra, LCModel extracts lipid concentrations with only 10–20% underestimation. However, the high SD detracts from the practical relevance of the estimated lipid concentrations.
Figure 7. Normalized absolute intensities of the macromolecule (abbreviated as mm1–mm4) and lipid (lip) signals for 4 Hz Gaussian line broadening vs. SNR. The means and SDs of 100 simulations are shown.
Download figure to PowerPoint
- Top of page
- MATERIALS AND METHODS
Both absolute and relative quantitations benefit from the incorporation of macromolecular and lipid information into the prior knowledge. The accurate determination of NAA and Cr concentrations by AMARES is encouraging, but is partially offset by incorrect Glx quantitation. Despite the slight variation of absolute metabolite concentrations with SNR and broadening, the correctness of the LCModel metabolite-to-creatine ratios is especially appealing. Obviously, the use of metabolite-to-creatine ratios not only circumvents the general problem of finding a valid reference point for absolute concentrations and reduces the influence of T2 to quantitation, it also benefits from the characteristic of the fit algorithm implemented in LCModel. The deteriorated precision of the metabolite-to-creatine ratios compared to concentration values (as shown by the enlargement of a greater number of error bars in Figs. 4 and 5 compared to Figs. 2 and 3), is more than offset by the higher accuracy of the results. The up to 1.6-fold higher coefficients of variation for metabolite-to-creatine ratios compared to concentrations are the same values found by Schirmer and Auer (20) in an in vivo intrasubject study.
Although the simulated spectra were restricted to the most prominent metabolites, they exhibit the main characteristics of the spectra of healthy controls. For routine quantitation, low concentration metabolites can be strongly elevated in pathologic brain tissue and should be included in the basis set. It has been reported that different sets of metabolites in the LCModel basis set can slightly change the estimated metabolite concentrations, and that these changes can be amplified by strong overlap in the spectrum (11). Hence, for real in vivo spectra, higher inaccuracies than were found in the simulation have to be expected. For example, Glx quantitation is additionally challenged by spectral overlap with signals from N-acetylaspartylglutamate (NAAG) and γ-aminobutyric acid (GABA), while mI quantitation is influenced by the overlap of glucose and scyllo-inositol. Nevertheless, the restricted basis set used in the simulation allows the analysis of the main features of both fit procedures, and should not limit the general validity of the typical findings presented here.
The idealizations of the simulation setup in terms of the choice of the starting point spectrum, as well as in terms of the selected a priori knowledge, translate into idealized simulation results. Hence, the accuracy and precision values shown here must be interpreted as lower bounds. The influence of variation in the starting parameters was not systematically investigated. The present results were obtained from fits with complete prior knowledge as well as reasonable start parameters for peak position, linewidth, and intensity. A preliminary test revealed only minor fluctuations in quantitation with regard to inaccurate start parameters of the macromolecular peaks for high-quality spectra, but increasing bias with diminishing quality of the spectra. A similar observation was reported by Soher et al. (21) for nonparametric baseline assessments, demonstrating a relationship between the accuracy of the baseline starting values and determined metabolite intensities.
The “black box” philosophy inherent to LCModel reduces user interaction and individual bias to a minimum. Only the preparation of the basis set (if it is not imported (22)) requires some effort. The absence of established macromolecular and lipid phantoms for in vivo MRS of the brain justifies the spline fit to all unknown spectral components. However, considering the improved fit quality that can be achieved with extended prior knowledge, the additional work needed to add synthetic spectra of macromolecules and lipids to the basis set is worthwhile. (During the preparation of this article, the new version of LCModel (6.0) was released with macromolecular and lipid components added to the basis set, which is even more advantageous.)
In contrast to LCModel, the preparation of the AMARES prior knowledge is laborious. AMARES offers only Lorentzian or Gaussian lineshapes for a given peak in the prior knowledge. Since for in vivo MRS spectra the more appropriate Voigt lineshape (19) has not yet been implemented, one has to start with a tradeoff. Here the use of Gaussian lineshapes is advised because lines of typical human brain spectra at 1.5 T exhibit only a small deviation from Gaussian lineshapes (17). Although this strategy offers only an approximation, the accuracy of the determined concentration for the narrow singlet peaks justifies this choice. However, one cannot rule out the possibility that the lack of a more sophisticated lineshape adaptation feature is the reason for the unsatisfying Glx results. AMARES with Lorentzian lineshapes was used by Mierisová et al. (23) to analyze short-TE rat brain spectra acquired at 4.7 T. Their linewidth FWHM of 12 Hz corresponds to 3.8 Hz at 1.5 T; however, the lines have a stronger Lorentzian character due to the shorter T2 of metabolites at 4.7 T.
In clinical studies, there is not much potential for enhancing the SNR, since a larger voxel volume would enhance volume averaging and the spectrum would lose specificity. Extensive averaging is limited by the patient's discomfort and the available scan time. Linewidth can, in principle, be narrowed by proper shimming; however, in clinical practice, automatic shimming does not always result in optimal correction, manual shimming is time-consuming, higher-order shims often are not available, and tissue heterogeneity sets limits for attaining homogeneously broadened peaks.
With respect to the variation in precision, it should be noted that the stepwise reduction of SNR by a factor of 2 at constant linewidth in Fig. 2 translates into a drastic stepwise reduction in total scan time by a factor of 4, whereas the stepwise increase in line broadening by 2 Hz in Fig. 3 results in a stepwise diminishing prolongation of scan time. Beginning with the starting point spectrum of the simulation, a line broadening of 6 Hz instead of 4 Hz requires a bit less than double the scan time to maintain SNR. Taking the acquisition times into account, the SD roughly decreases with the square root of scan time. Since this is the expected effect for well-resolved peaks, separability of the dominant resonances NAA, Cr, and Cho was not a problem up to the investigated 8 Hz broadening. However, the strongly overlapping, similarly shaped Glu and Gln signals were not well separated. The acceptable separation of Glu and Gln found by Stanley et al. (24) probably originated from the exclusion of baseline components in their simulation. Because the influence of linewidth on precision is of minor importance in the investigated interval up to 8 Hz broadening, it is advisable to define a target SNR and discard spectra from group analysis where a minimum SNR is not obtainable by shimming or by an acceptable prolongation of scan time. For LCModel, the reported slight dependence of absolute concentrations on linewidth (20, 25) was confirmed by the simulation. This indicates that precision can be further increased by the exclusion of badly resolved spectra from the statistical analysis. This should be the method of choice for clinical studies with large patient cohorts; however, for uncommon diseases, the small size of the patient groups often compels researchers to practice more moderation in accepting borderline case spectra. In such cases, the current results suggest that metabolite-to-creatine ratios should be used instead of concentrations. Additionally, metabolite-to-creatine ratios should reduce the linewidth bias if the spectra of a control group exhibit significantly higher or lower resolution.
In addition, the choice of software packages should be guided by the requirement for a reliable “black box” approach, as offered by LCModel, or by the need for a flexible software tool for spectroscopists, as in the case of MRUI with the AMARES program. Moreover, we emphasize that AMARES and LCModel are continually being revised. We hope that the current work will help provide a more rational basis for decisions regarding, for example, which spectra should be discarded from analysis, whether a particular metabolic question can be reliably answered by MRS, or which quantitation approach should be selected.