INTRODUCTION

MRS allows quantifying cerebral metabolites, thus helping to characterize brain tissue under healthy as well as pathologic conditions and follow up changes during aging, disease progression and therapeutic intervention. In general it is accepted that the Signal-to-Noise-Ratio (SNR) will influence the precision of the quantification: high SNR will lead to better precision than low SNR. Low SNR spectra are expected to increase the variance of calculated concentrations but the mean value should remain stable ^1,2.Nevertheless own observations as well as significant variation in the reported absolute concentrations between studies made us suspect a possibly systematic bias of SNR on the estimations of different metabolites.

METHODS

In order to pinpoint the influence of SNR we re-evaluated grey matter spectra from a 3T system (Siemens PRISMA) as well as data from 9.4T system (Bruker Biospec). Human spectra (PRESS and STEAM) and spectra from long-tailed macaques (PRESS) were acquired at 3T, while the 9.4T system was used to obtain STEAM spectra from common marmosets and rats. Firstly, we identified high SNR example spectra from all the above species. These spectra were evaluated by LCModel. In the next step we added random noise at 20 different levels to the measured spectra by adding complex numbers to the measured raw data (=FID). This procedure was repeated 100 times for each noise level resulting in 2000 modified spectra which were then again analyzed by LCModel. The resulting concentrations for 12 metabolites were inspected for systematic aberrations.

RESULTS

Adding noise at 20 different levels resulted , as expected, in a degradation of the observed SNR (see Figure 1), which is defined in LCModel as the ratio of the maximum in the spectrum-minus baseline over twice the root mean square of the residuals. In the two human examples using PRESS at 3T we started with SNR values of 58 and 71, while 40 was the SNR-starting point in the human STEAM spectrum. In all examples adding noise decreased these values down to about 5, after 20 steps. Due to smaller voxel sizes in all animal experiments single spectra showed much lower starting values for SNR. Therefore phase coherent averaging of spectra from animals was done in order to create spectra with higher SNR. These averaged spectra served as baseline and were subsequently degraded by adding noise at 20 levels. Plotting metabolite concentrations calculated by LCModel for all 2000 modified spectra against the SNR demonstrates, as expected, the increasing variance of the calculated values with decreasing SNR (Figure 2). Besides this expected decrease in precision we recognized additional systematic changes in the average concentrations. At 3T, these changes were maximal for Glx, ranging from 8 mM (100%) in the human PRESS examples, 1.5mM (30%) in the human STEAM spectra, and 10 mM (100%) in macaque PRESS spectra. All other metabolites showed much lower influence of the SNR on estimated concentrations (see Figure 3&4). At 9.4T we again found some systematic but less pronounced differences in Glx concentrations (1.5 mM/15% in marmosets, 3 mM/25% in rats). In contrast to 3T examples significant differences were also found for creatine (1.5 mM/30% in marmosets, 2.5 mM/70% in rats). Metabolites with low concentrations (i.e. GPC <2mM) could have extremely high relative changes (>100%) but probably were not reliable (Figure 5).

DISCUSSION/CONCLUSION

We demonstrated that degradation of SNR caused substantial bias in estimated metabolite concentrations in localized proton NMR spectroscopy. These effects were mainly related to data analysis then to data acquisition since the effects could be found using different MR sequences (PRESS, STEAM), different field strengths (3T, 9.4T) and different species (human, monkeys, rats). Especially Glx concentrations were prone to SNR alterations, with lower SNR leading to higher Glx concentrations. Because other metabolite concentrations seem to be much less affected, or not at all, elevated Glx values could easily be misinterpreted as brain region- or species-dependent when differences in SNR are not considered. The common practice to adapt voxel size to target brain regions more specifically directly translates into SNR changes, when not intercepted by increasing measurement time. Therefore differences in metabolite concentrations from different brain regions should be treated with great care. Also comparisons of metabolite concentrations from different species may be affected by this imprecision because the use of different field strengths and dedicated coils can result in SNR differences as well.

Acknowledgements

No acknowledgement found.