NON-INVASIVE GRADING OF ASTROCYTIC TUMOURS FROM THE RELATIVE CONTENTS OF MYO-INOSITOL AND GLYCINE MEASURED BY IN VIVO MRS

MRI and MRS are well established methodologies for evaluating intracranial lesions (1,2). MRS is nowadays receiving increasing attention from radiologists and neurologists, and its use to help MRI in even here accuracy is not perfect (7). Neuroimaging-based classification (defined as the assignment of type and grade) is frequently reported to be unreliable, especially for certain lesions such as gliomas (8, 9). An additional non-invasive method for accurately diagnosing and grading brain tumours would be a major advance, particularly in those cases. As MRS spectra of brain tumours and other focal lesions are often quite distinctive, H-MRS is a very promising non-invasive method for brain tumour diagnosis, and it is becoming widely acknowledged as a useful complement to MRI. Several in vivo diagnostic approaches for MRS have been tested, such as spectroscopic imaging (10) or single voxel MRS (11, 12). Previous studies with MRS have suggested correlations between metabolic features in vivo and the histopathological grade of astrocytic tumours (4, 13, 14), but they were generally concerned with the NAA, Cr and Cho resonances. In addition to those resonances, MRS produces much more information of potential interest for astrocytic tumour grading, e.g. the myo-inositol resonance (ca 3.55 ppm) which can be measured at short echo times (15-18). It had earlier been realized that the J-modulation dependent effect on mI resonances would strongly decrease the apparent mI peak intensity at ca. 3.55 ppm (19). In contrast, since the gly resonance is a singlet and shows no J-dependent modulation, the ca. 3.55 ppm peak intensity will not be affected by echo time the clinical decision making process is increasing. MRS of brain tumours is currently used for diagnosis, grading, assessing therapeutic response and providing prognostic information on survival (3-4). According to Stewart et al (5), approximately 60% of all intracranial tumours are of neuroepithelial origin (gliomas). Among the glioma subtypes, Chamberlain and Kormanik (6) have reported percentages of 4050% for glioblastomas (GBM), 3035% for anaplastic astrocytomas (A3), and 15-20% for low grade astrocytoma (A2). Proper grading of astrocytic tumours is important for the clinician to decide upon treatment and patient care (4, 6). Current radiological methods do not distinguish adequately between the large number of recognised types of brain tumours, and histopathological diagnosis from a biopsy remains the gold standard for diagnosis and grading, although JBR–BTR, 2011, 94: 319-329.

MRI and MRS are well established methodologies for evaluating intracranial lesions (1,2).MRS is nowadays receiving increasing attention from radiologists and neurologists, and its use to help MRI in even here accuracy is not perfect (7).Neuroimaging-based classification (defined as the assignment of type and grade) is frequently reported to be unreliable, especially for certain lesions such as gliomas (8,9).An additional non-invasive method for accurately diagnosing and grading brain tumours would be a major advance, particularly in those cases.As MRS spectra of brain tumours and other focal lesions are often quite distinctive, 1 H-MRS is a very promising non-invasive method for brain tumour diagnosis, and it is becoming widely acknowledged as a useful complement to MRI.Several in vivo diagnostic approaches for MRS have been tested, such as spectroscopic imaging (10) or single voxel MRS (11,12).Previous studies with MRS have suggested correlations between metabolic features in vivo and the histopathological grade of astrocytic tumours (4,13,14), but they were generally concerned with the NAA, Cr and Cho resonances.In addition to those resonances, MRS produces much more information of potential interest for astrocytic tumour grading, e.g. the myo-inositol resonance (ca 3.55 ppm) which can be measured at short echo times (15)(16)(17)(18).
It had earlier been realized that the J-modulation dependent effect on mI resonances would strongly decrease the apparent mI peak intensity at ca. 3.55 ppm (19).In contrast, since the gly resonance is a singlet and shows no J-dependent modulation, the ca.3.55 ppm peak intensity will not be affected by echo time the clinical decision making process is increasing.MRS of brain tumours is currently used for diagnosis, grading, assessing therapeutic response and providing prognostic information on survival (3)(4).
Current radiological methods do not distinguish adequately between the large number of recognised types of brain tumours, and histopathological diagnosis from a biopsy remains the gold standard for diagnosis and grading, although when gly is its major contributor.Barba et al (20) utilised this echo time J-coupling modulation-dependent effect to estimate the relative mI and Gly content in human brain tumours in vivo.This approach was extended using a two-phantom protocol to produce an expression that should allow estimation of the MRS visible mI/gly content ratio (21).
We now report the use of a simple expression for mI + gly content relative to creatine content for astrocytic tumour grading from in vivo MRS data, and an investigation of a biochemical rationale for this method, based on the analysis of biopsy metabolite extracts by high field NMR spectroscopy.

Patients
The Samples were centrifuged to eliminate precipitated potassium perchlorate (40,000 x g for 15 min), reextracted once and supernatants combined and freeze-dried.All operations were performed at 4ºC.The PCA-insoluble pellet was kept for further protein analysis.Prior to NMR acquisition, samples were resuspended in 400 ml of D2O.The pH* (pH meter reading uncorrected for the deuterium isotope effect) was adjusted to 7.06 ± 0.22 and TSP (3trimethylsilyl-2,2,3,3-sodium tetradeuteropropionate) was added as chemical shift and concentration standard (final concentration of TSP proportional to the initial frozen sample weight).Some samples had been analysed by the group previous to the introduction of the internal standard into the extraction protocol; they were corrected for the mean losses (25%) observed in our present PCA protocol which includes an internal standard (24).

In vivo MRS spectra
Patient spectra (n = 95; A2 n = 18, A3 n = 7, GBM n = 70) were obtained from the INTERPRET validated database (http://gabrmn.uab.es/interpretvalidateddb<accessed on 2011-01-21>).The consensus protocols developed by the INTERPRET project for acquisition of MRI and MRS data (25) are shown in Table I.Gadolinium contrast agents, when used, were given prior to MRS measurements.The MRS voxel was Biopsy samples (n = 74; A2 n = 7, A3 n = 8, GBM n = 59) were collected after surgical resection of tumours at Hospital Universitari de Bellvitge.The Institutional Review Board approved the study and patients gave signed informed consent prior to surgery.Samples were frozen in liquid nitrogen within 5 minutes of collection and maintained in those conditions until processing at the Universitat Autònoma de Barcelona (UAB).The Anatomical Pathology department of the hospital provided the diagnosis of the patient.
For the in vivo MRS cases, the histopathological diagnosis was achieved by examination of a biopsy sample from the tissue excised at operation.Diagnostic agreement was required between a minimum of two of the consultant neuropathologists involved in the INTERPRET project (see (11) for further details).For the unpaired cases (no MRS, but biopsy collected), the local pathologist's diagnosis was used.

Perchloric acid extraction of biopsies
Perchloric acid (PCA) extracts were carried out essentially as in (22).Frozen samples were pulverized with a pre-cooled mortar and pestle.The fine powder obtained was transferred into a tube containing 0.5 M PCA (6 ml/g wet weight) and homogenized.External fumarate (ca. 5 µmol/g wet weight) was added as an internal standard to account for extraction process losses (23).

Phantom studies at clinical field
Phantom studies were planned from calculations based on the concentration ranges and T2 of brain metabolites (mI, gly, Cr) reported by others (26)(27)(28), as well as previous experimental data from our group (20), and were devised in order to estimate by previous external calibration the mI/gly ratio in vivo.The calibration curve solution had a constant creatine concentration (10 mM) and variable mI and gly concentrations (Table II).The phantom used in these acquisitions was based on the EUROSPIN phantom, and is described in more detail in (29).
Acquisition of phantom spectra (inner phantom cube containing the model solutions described in Table II) was carried out in a Philips scanner (ACS NT) operating at 1.5 T. MRS; acquisition parameters were: PRESS sequence, 512 data points, SW 1000 Hz, TR 2 s, TE 30 and 136 ms.The number of acquisitions was 96 for short TE (STE) and 128 for long TE (LTE), except in the case of the phantom with a lower concentration of mI (12 mM) in which the number of acquisitions was increased to 768.The single voxel (SV) volume was placed inside the inner cube of the phantom and the voxel size was fixed to 1.5 cm 3 .Unsuppressed water reference spectra were also acquired with the same parameters except that the number of acquisitions was 16.

In vitro high resolution NMR spectroscopy
Pulse-and-acquire spectra were acquired in an ARX -400 spectrometer operating at 9.4 T (Bruker SADIS, Wissembourg, France) at the UAB ing in a line broadening of 1Hz and normalization of the spectra to Euclidean unit length (eq.1), with X being the height of any point in the spectral vector.Next, the signal-tonoise ratio (SNR) for the tallest peak in the 0-3.4 ppm range was also calculated (see (11) and (12) for further details).The noise level was defined as the average standard deviation of noise between 9 to 11 ppm.The resulting processed file had 512 points and ranged from -2.7 to 7.1 ppm.Astrocytoma cases in the database were only used if they had available both one short echo and one long echo time spectrum, which had both passed spectroscopic quality control criteria (water linewidth at half height < 8 Hz and SNR > 10, and lack of visible spectral artifacts (11)).
An additional inclusion criterion for this work (but see also (29)) was that only spectra with SNR higher than 5 for the 3.55 ppm (mI + gly) and 3.03 ppm (creatine) signals at short echo time (STE) spectra were used.Following these restrictions, 9 spectra (1 A2 and 8 GBM) were discarded.
In order to make the (mI + gly)/Cr function simple to measure, we used peak heights rather than peak areas.In principle a non-specialist user of the method could measure them with a pencil and ruler.In the present study, a modified processing module was developed to calculate SNR and peak heights for in vivo and phantom spectra.The 3.03 and 3.55 peak heights and their SNR (range -2 to -1 ppm as basis) were automatically calculated for STE and long echo time (LTE) for the resulting processed file derived from the INTERPRET data manipulation software (DMS).
Before peak height measurement in phantom spectra, a 4 Hz line broadening was applied to mimic the linewidth effects in in vivo brain tumour spectra.After that, the (mI + gly)/Cr ratio (eq.2) was calculated.facility 'Servei de Ressonància Magnètica Nuclear' (SeRMN) at 298 K. Acquisition parameters were: time domain 16 k (8 k complex points), sweep width 4854 Hz, recycling time 10 s, and water presaturation with 0.05 mW power for 1 second.Initially, a short spectrum (16 scans) with 30s of recycling time was acquired to allow for fumarate quantification and loss-related corrections.The number of acquisitions was between 64 and 4096 scans, depending on the initial biopsy sample weight.Spectra were processed with DC correction, zero filling to 16 K and 0.3 Hz line broadening before Fourier transformation with WINNMR version 6.1.0.0 (Bruker Daltonik, GmbH) in a PC.Peaks of interest were fitted to Lorentzianshaped curves, quantified with respect to the TSP reference with WINNMR, and finally corrected for the percent losses measured or estimated from fumarate content in the final extract sample.

Processing of in vivo MR spectra
In vivo and phantom spectra were automatically processed with a software module derived from the INTERPRET data manipulation software (11,12).This module carried out the required functions such as Fourier transform, residual water filtering by HLSVD between 4.3 and 5.1 ppm, offset correction, zero order (Klose algorithm) phase correction, setting the 4.2 to 5.0 ppm range to zero, exponential apodisation result-

Simulation of expected (mI + gly)/Cr from in vitro PCA extract calculated values
The (mI + gly)/Cr that would be expected in vivo was simulated from in vitro PCA extract sample data.This simulation was based in data obtained with the phantom calibration curve and an equation (equation 3, see results) was adjusted to give curve points that correlate the (mI + gly)/Cr in vivo to mI and gly concentrations measured in vitro.

Statistical analysis
Statistically significant differences were evaluated as follows: the normal distribution of values was assessed with the Kolmogorov-Smirnov test and the variance homogeneity with the Levene's test.Values with a normal distribution and homogeneous variance were compared with Student's t test or (if The mI signal at LTE was attenuated when compared to the creatine signal, due to mI J-coupling induced phase modulation, whereas glycine remained practically isointense with respect to creatine at 3.03 ppm (Fig. 1).Values of calculated mI and gly concentrations used in the phantom and the resulting (mI + gly)/Cr ratio were used to devise the calibration curve shown in figure 2. From these values, an asymptotic equation was obtained from fitting the calculated (mI + gly)/Cr for each (mI)/[gly) ratio (eq.3), with the SIGMAPLOT (Systat Software) program.

3.11
From equation [3], (mI + gly)/Cr (y) values in vivo can be predicted more than two comparisons were carried out) the ANOVA test.For values with normal distribution but inhomogeneous variance, Student's t-test was carried out.
Values with non-normal distribution were evaluated with the nonparametric Mann-Whitney's U test.Significance level was set to 0.05, and all analyses were carried out with SPSS version 11.5.1 and 14.0 (SPSS Inc, Chicago, IL).

Phantom studies and calculation of (mI + gly)/Cr
Phantom spectra were acquired under conditions of good SNR, which allowed clear identification of peaks of interest (see figure 1).The SNR range obtained for the signals of interest was 30-122 (STE) and 51-78 (LTE).

A B
from assumed mI and gly content ratios (x), and vice versa mI/gly ratios calculated from in vivo measured (mI + gly)/Cr.

In vivo measurement of (mI + gly)/Cr for calculation of (mI)/[gly)
The initial set of in vivo spectra (n = 95) available in the database was reduced to n = 86 by application of the criteria mentioned previously: unsuppressed water linewidth and SNR for the whole spectral pattern, and for the individual 3.55 and 3.03 peaks.Variations due to short echo time STEAM or PRESS sequences (STE) were assessed by separating cases according to acquisition conditions and testing (t-test) for differences between groups (data not shown), and no significant differences were found.This observation agrees with that of Ernst and Hennig (19), who reported differences in modulations between metabolites measured with STEAM and PRESS sequences, but only at higher echo times (> 50 ms).Slight differences in echo time (20-35ms for STE or 135-136 ms for LTE) were assumed to be negligible for the purposes of this work with respect to MRS pattern effects in the mI/gly or Cr regions.Table III summarizes (mI + gly)/Cr and (mI)/[gly) calculation from equation 3 for the studied groups (A2, A3 and GBM).

Gly and mI content in biopsies from in vitro measurements of PCA extracts and calculation of the predicted (mI + gly)/Cr in vivo
Typical PCA extract spectra from biopsies can be seen in figure 5, with examples of spectra from A2 and GBM.Values for (mI)/[gly) were calculated for each in vitro spectrum, and equation 3 was used for predicting the expected (mI + gly)/Cr value in vivo.Average values for gly and mI concentration ratios obtained from biopsies are summarized in Table V.The total number of paired cases (MRS and biopsy extraction) is summarized in Table IV.
The boxplot for the calculated (mI + gly)/Cr is shown in figure 4B and in Table III a statistically significant decrease in (mI + gly)/Cr between A3 and GBM can be seen, while the reverse tendency is observed between A2 and A3, the former presenting lower values.
While the (mI + gly)/Cr values for GBM and A3 are not significantly different when comparing values measured in vivo with values calculated from extract data, (mI + gly)/Cr values for A2 are significantly different, being 3-fold lower for in vivo data with respect to ratios calculated from in vitro data (Table III).

Quantification of the mI/gly ratio in vivo by combining data acquisition at two echo times
Several approaches have been used to evaluate the mI and/or gly content of the human brain and of human brain tumours using data obtained in vivo in clinical scanners (1.5-7T) (20,(30)(31)(32)(33)(34)(35).This measurement is difficult, especially at low field, because of the co-resonance at ca. 3.55 ppm of the singlet resonance of glycine with the J-coupled resonances from the protons of myo-inositol (protons 1, 3, 4 and 6).Most authors have usually resorted to independent gly and mI quantification at a single echo time (between 30 and 144 ms (31,33,34)) or have used a method based on averaging echo time results in the 30-284 ms range (31,36), and using frequency domain fitting and quantification with LCModel.The method described in (31,36) was also based on the well known fact that the Jmodulation dependent effect on mI resonances would strongly decrease the apparent mI peak intensity at ca. 3.55 ppm at short TE (15,19,20).Additionally, since the gly resonance The distribution of glial cases in the validated INTERPRET database that fulfilled the inclusion criteria for this analysis partially agrees with the distribution reported in (6), in which of 86 astrocytoma cases, 72.1% were glioblastomas, 8.1% were anaplastic astrocytomas and 19.8% were diffuse astrocytomas.The agreement is even more complete if high grade (GBM + A3) tumours are compared with low grade (A2) tumours; the numbers are then 80.2% vs. 19.8% in the INTERPRET data and 70-85% vs. 15-20% in (6).In typical in vivo spectra (Fig. 3A and 3B, mean spectra), the attenuation of the 3.55 ppm signal at LTE is higher in low-grade tumours, whereas in glioblastomas, there is usually little attenuation.Furthermore, the boxplot in figure 4A shows that (mI + gly)/Cr tends to be inversely correlated with grade, achieving statistically significant differences between A2 and GBM.
The (mI + gly)/Cr values obtained for paired in vivo/in vitro cases (most of which, 15/16, were GBMs, Table IV) were variable; some cases presented similar values and some cases did not.These values were statistically compared with Student's "t" test and significant differences (p = 0.02) were found.The (mI + gly)/Cr value tended to be higher in vivo, suggesting a lower apparent glycine content in vivo.

In vivo
In  ).Note that for the A2 cases, the 3.55 ppm signal (arrows) is high at STE and much lower at LTE, whereas for the GBM the relative signal height with respect to creatine at 3.03 ppm remains high.Spectral intensity was set to zero after 4.2 ppm to avoid contribution of artefactual signal due to suboptimal water suppression (11).
A B is a singlet and shows no J-dependent modulation, the ca.3.55 ppm peak intensity is not affected by echo time when gly is its major contributor.This effect had already been noticed in studies of glial tumours in vivo (15,20).Barba et al (20) used the echo time J-coupling modulation-dependent effect to estimate the relative mI and gly content of human brain tumours (hemangioperycitomas and meningiomas) in vivo.In that study, mI/gly content was estimated from a two phantom protocol.This approach has been extended in our present work (see also (21) for a preliminary description) to produce an expression (equation 3) that allows MRS-visible mI/gly content ratio estimation under our experimental conditions.No correction was applied in our work for possible differences in T1 or T2 values between the phantom solution and brain tumour tissue.Accordingly, the mI/gly ratio calculated for the tumour tissue should be considered an approximation to the real tissue mI/gly.If T1 or T2 values differentially change for mI and gly, this could produce an apparent mI/gly ratio derived from the measured ratio, weighted by the relative NMR visibility of the two compounds, and compared with the height of a singlet resonance which is not affected by the echo time modulation (creatine).A similar approach to the present one was described by Hattingen et al. (35) for MRSI dataset grids.They also used the change in pattern produced by recording spectra from ratio obtained in vivo (Fig. 4A) by substituting the experimental values into equation 3 has allowed us to demonstrate statistically significant differences in calculated mI/gly ratios between low grade astrocytic tumours (n = 17) and glioblastoma multiforme (n = 62) (Table III).This would agree with previous work (17) in which a tendency for a higher mI/Cr to be associated with lower grade had already been described at a single echo time (20 ms) in a smaller cohort of patients (n = 34), albeit without statistical analysis of the differences described.However, even though a trend towards a decrease in mI/gly ratio with grade exists (Fig. 4A) and statistically significance can be demonstrated for the difference between A2 and GBM, there is a great deal of overlap among grades.Therefore, if the (mI + gly)/Cr value is to be used for tumour grading, it should be combined with other relevant spectral features that have already been described (11,37,38).Indeed, the mI + gly peak height or peak area have already been used for classifier development at short (ca.30 ms) or long (ca.136 ms) echo times (11,16,37), and a joint classifier that uses information from the two echo times has recently been developed (39).We have already pointed out (37) that the manual combination of data obtained at short and long echo time significantly improves human brain tumour classification, and also the grading of astrocytic tumours.The (mI + gly)/Cr factor described here would human brain tumours at two echo times (30 and 144 ms), time domain data fitting using AMARES and QUEST, and phantom spectra, to calculate mI/gly ratios for low grade astrocytomas (A2), high grade astrocytomas (A3 + GBM) and control subjects (35).Their results, which would be qualitatively equivalent to 1/(mI + gly)/Cr using the terminology in our study, also showed a trend to increasing apparent gly content in high grade tumours, although a full quantitative comparison between both studies is made difficult by the different phantom approach used and their combination of A3 (n = 10) and GBM (n = 5) in a single high grade group.The approach described herein may have some advantages with respect to other previously developed approaches in its relative simplicity, which should allow its use with SV MRS data acquired at any clinical centre.Monitoring of the relative mI/gly content in vivo may also be of interest for other clinical conditions in which changes in the relative contents of mI or gly are expected (20,34,36).

Relevance of the apparent mI and gly content detected by MRS in vivo for non-invasive astrocytic tumour grading
The relative content of mI and gly in tumours was calculated ((mI + gly)/Cr) from the quantitative pattern change caused by the MRS visibility of mI varying at increasing echo times (19,20).The apparent mI/gly Fig. 4. -A) Boxplot for (mI + gly)/Cr automatically calculated from actual in vivo spectra ($) labels significant differences using an ANOVA test between the A2 and GBM groups.Number of cases for each group is given on the abscissa axis.B) Boxplot obtained for predicted (mI + gly)/Cr in vivo, calculated from values obtained from PCA extracts of biopsies.($) labels significant differences between the A3 and the other two groups.In these boxplots, upper and lower box limits represent 3 rd and 1 st quartiles, respectively.The central thick line is the median.Whiskers label maximum values comprised between the quartile and the product Interquartile Range (IQR) x 1.5.Outliers are represented as circles when value is within the 1.5 and 3.0 x IQR.Extreme outliers (higher than 3.0 x IQR) are represented as *.

A B
effectively combine the information contained in spectra obtained at the two echo times in the 3.55 ppm mI + gly region.Using a single echo time for astrocytic tumour grading may be misleading when using the 3.55 ppm region since, for instance, a signal that is high at short echo time would modulate and almost disappear at long echo time if it were due to mI, whereas it would be constant if it were due to gly.This confusion disappears if two echo times are used, as has previously been proposed (15,20,35).Noisy spectra can also cause overlap in calculated (mI + gly)/Cr -factor values , since the value of (mI + gly)/Cr will then approach one, independently of the grade investigated.In order to reduce the contribution of "bad quality" cases we therefore restricted our analysis to short TE spectra with SNR values for the 3.55 and 3.03 ppm resonances that were larger than 5.A further reason for overlap in (mI + gly)/Cr data is that GBM is not a homogeneous tumour type: both tumour evolution and genetic data suggest that GBMs can be divided into primary and secondary groups (40), while transcriptomic studies have demonstrated three molecularly differentiated primary GBM subtypes (41).Although a detailed in vivo analysis of the MRS pattern of these GBM subtypes is still lacking, ex vivo data obtained by HR-MAS analyses of tumour biopsies suggest that secondary GBM could display a much higher mI content than primary ones (42).This would point to the possibility that the GBM outliers with a high (mI + gly)/Cr value in figure 4A could correspond to secondary GBMs.Further studies comprising survival data may be helpful in this respect.

Comparison of mI/gly calculated from in vivo data and mI/gly measured from biopsies in vitro
We were interested in validating the in vivo estimated mI/gly with in vitro data obtained from high resolution NMR spectra of PCA extracts from brain tumour biopsies.The glycine and mI contents in Table V show a similar trend compared to other reported studies (43)(44)(45)(46)(47)(48): mI content decreased with astrocytic grade, albeit non-significantly, in agreement with (43)(44)(45)(46)(47)(48).Furthermore, gly content increases with grade, in agreement with previous studies (43,44,46,48)    in (43)), and the same trend in gly increase with tumour grade was observed in (47 and 50) (HRMAS studies).In this respect, we are confident that our results from astrocytic tumour PCA extracts, obtained from the largest biopsy dataset yet described (n = 74), represent the expected metabolite pattern of the tumour grades investigated here.With respect to the discrepancy between (mI + gly)/Cr measured in vivo and calculated from in vitro data, there are several possible explanations.As biopsies and in vivo spectra were not matched (Table IV) it could be argued that sampling was somewhat biased.Furthermore, the biopsy sample may not necessarily represent the average spectral pattern sampled from a much larger voxel by the SV in vivo MRS approach.The rather large number of cases investigated (86 in vivo and 74 in vitro) makes this unlikely, as bias should not be coherent.Another possibility to take into account would be post-ischaemic changes in mI and or gly content between the in vivo tumour pattern and the pattern obtained from the PCA extract after open surgery resection of the biopsy.We are not aware that such changes have been described after the short ischaemia experienced by biopsies during the pre-freezing time (ca. 5 min) after surgical excision.One further possibility to take into account would be a reduced in vivo NMR visibility of gly in A2.If this were to be the case, the in vivo (mI + gly)/Cr value would be providing an apparent mI/gly ratio contributed only by the NMR-visible mI and Gly pools.One of the accepted causes of NMR invisibility of a small molecular weight metabolite is binding to macromolecular structures, with concomitant reduction of its T2 value (e.g.ADP bound to F-actin in muscle microfilaments).In this respect, binding of gly to multimeric channel proteins in the plasma membrane of PC-12 cells (50) or to a glycine-sensitive anion death channel in sinusoidal endothelial cells (51) have been proposed.Increased gly content protects various types of cells from necrotic or apoptotic death induced by ATP depletion (PC-12 cells, (51)), hypoxia (hepatic sinusoidal cells, ( 52)) or ischaemia-reperfusion derived reoxygenation injury (cardiomyocytes, (53)).As these are physiological situations usually encountered by tumours during progression, it would make sense if similar gly-derived effects were taking place.Then, since the gly pool in A2 forme.The (mI + gly)/Cr ratio simulated from mI and gly concentration measured from PCA extracts of tumour biopsies suggest that part of the gly pool is not NMR visible in vivo in low grade astrocytomas.The (mI + gly)/Cr ratio may have application for astrocytic tumour grading and also for other cellular situations or pathologies in which the mI/gly ratio is expected to change.
tumours is smaller than in GBM, most gly would remain NMR invisible, bound to those plasma membrane proteins, perhaps due to a lower requirement for its protective effect.Once gly content increased above a certain threshold, or it was extracted by PCA, the NMR visibility would be recovered and the (mI + gly)/Cr would be correspondingly affected.In relation to this, it is worth mentioning here the recently described increase in apparent detection of Gly by HRMAS with increased recording time in GBM biopsies (48), which was also interpreted by the authors as suggesting higher gly visibility in GBM ex vivo with respect to the in the in vivo tumour MRS pattern.The concentration of gly that displays a protective effect in the various cell systems mentioned above is in the 3-5 mM range (51)(52)(53).This value is very close to the gly content found by us in GBM (3.28 µmol/gtw, Table V).Both explanations, the post-ischaemia changes and the reduced NMR visibility, could also account for some discrepancies observed between the expected (mI + gly)/Cr (deduced from in vivo data) and the calculated (mI + gly)/Cr (from in vitro data) for some of the paired cases in Table IV (results not shown).
It may be relevant to mention here that the mI/gly ratio has been analyzed in PCA extracts of rat glioma C6 cells (54) and has been found to be 3.2 times higher in postconfluent cells (mI/gly = 7.2, proliferation slowed down) than in log phase cells (mI/gly = 2.23, active proliferation).In this respect, actively proliferating C6 cells in culture would parallel the GBM behaviour, whereas proliferation-arrested C6 cells would partially mimic the situation in low grade astrocytic tumours.Therefore, changes in the mI/gly ratio in an astrocytic tumour in vivo with location (tumour heterogeneity) or progression towards malignancy could well be accessible noninvasively by monitoring changes in the (mI + gly)/Cr ratio.

Conclusions
An experimental protocol and an empirical formula have been developed to allow calculation of the (mI + gly)/Cr ratio from in vivo spectra of astrocytic human brain tumours.The in vivo (mI + gly)/Cr decreases with astrocytic tumour grade, and there is a statistically significant difference between the ratios in low grade astrocytoma and glioblastoma multi-

Fig. 1 .
Fig. 1. -Spectra at STE (top) and LTE (bottom) at 1.5 T of phantom solutions of two calibration curve points.A) phantom #1: creatine 10 mM, gly 80 mM and B) phantom #2: creatine 10 mM, mI 12 mM.The mI/gly peak at ca 3.55 ppm is labelled with an arrow.Phantom spectra were scaled to approximately constant creatine peak height for each pair of STE/LTE spectra.

Fig. 2 .
Fig. 2.-Calibration curve (semilogaritmic plot) obtained with phantom measurements.For proper fitting, an arbitrary value of 500,000 was used in place of "ϱ" for (mI)/[gly) in phantom solution #10, for which two different acquisitions were performed.The value for phantom #1 is not shown in the plot.

Fig. 3 .
Fig. 3. -In vivo mean spectra + /-SD (shading) at STE (left) and LTE (right) for A) A2 cases (n = 17) and B) GBM cases (n = 62).Note that for the A2 cases, the 3.55 ppm signal (arrows) is high at STE and much lower at LTE, whereas for the GBM the relative signal height with respect to creatine at 3.03 ppm remains high.Spectral intensity was set to zero after 4.2 ppm to avoid contribution of artefactual signal due to suboptimal water suppression(11).

Table II .
-Glycine, myo-Inositol and creatine concentrations in the phantom used for the calibration curve.
[(3. 03 ppm height LTE) ] (gly was detected only in one out of five low grade astrocytoma cases analyzed

Table IV .
-Number of paired cases (in vivo MRS + in vitro biopsy extract) in this study.