Neuroradiology (2006) 48: 622631 DOI 10.

1007/s00234-006-0102-3

DIA GNOSTIC NEURORADIOLO GY

Alfonso Di Costanzo Tommaso Scarabino Francesca Trojsi Giuseppe M. Giannatempo Teresa Popolizio Domenico Catapano Simona Bonavita Nicola Maggialetti Michela Tosetti Ugo Salvolini Vincenzo A. dAngelo Giocchino Tedeschi

Multiparametric 3T MR approach to the assessment of cerebral gliomas: tumor extent and malignancy

Received: 25 November 2005 Accepted: 12 April 2006 Published online: 3 June 2006 # Springer-Verlag 2006

M. Tosetti Department of Magnetic Resonance, Scientific Institute Stella Maris, Pisa, Italy U. Salvolini Department of Neuroradiology, Azienda Ospedaliera Universitaria Umberto I, Torrette, Ancona, Italy

A. Di Costanzo (*) Department of Health Sciences, University of Molise, Via de Sanctis 2, 86100 Campobasso, Italy e-mail: alfonso.dicostanzo@unimol.it Tel.: +39-0874-404862 Fax: +39-0874-404778 T. Scarabino . G. M. Giannatempo . T. Popolizio Department of Neuroradiology, Scientific Institute Casa Sollievo della Sofferenza, San Giovanni Rotondo, Foggia, Italy F. Trojsi . D. Catapano . S. Bonavita . G. Tedeschi Department of Neurological Sciences, Second University of Naples, Naples, Italy V. A. dAngelo Department of Neurosurgery Scientific Institute Casa Sollievo della Sofferenza, San Giovanni Rotondo, Foggia, Italy N. Maggialetti Faculty of Medicine University of Bari, Bari, Italy

Abstract Introduction: Contrast-enhanced MR imaging is the method of choice for routine assessment of brain tumors, but it has limited sensitivity and specificity. We verified if the addition of metabolic, diffusion and hemodynamic information improved the definition of glioma extent and grade. Methods: Thirty-one patients with cerebral gliomas (21 high- and 10 low-grade) underwent conventional MR imaging, proton MR spectroscopic imaging (1H-MRSI), diffusion weighted imaging (DWI) and perfusion weighted imaging (PWI) at 3 Tesla, before undergoing surgery and histological confirmation. Normalized metabolite signals, including choline (Cho), N-acetylaspartate (NAA), creatine and lactate/ lipids, were obtained by 1H-MRSI; apparent diffusion coefficient (ADC) by DWI; and relative cerebral blood volume (rCBV) by PWI. Results: Perienhancing areas with abnormal MR signal showed 3 multiparametric patterns: tumor, with abnormal Cho/NAA ratio, lower ADC and

higher rCBV; edema, with normal Cho/NAA ratio, higher ADC and lower rCBV; and tumor/edema, with abnormal Cho/NAA ratio and intermediate ADC and rCBV. Perienhancing areas with normal MR signal showed 2 multiparametric patterns: infiltrated, with high Cho and/or abnormal Cho/NAA ratio; and normal, with normal spectra. Stepwise discriminant analysis showed that the better classification accuracy of perienhancing areas was achieved when regarding all MR variables, while 1 H-MRSI variables and rCBV better differentiated high- from low-grade gliomas. Conclusion: Multiparametric MR assessment of gliomas, based on 1H-MRSI, PWI and DWI, discriminates infiltrating tumor from surrounding vasogenic edema or normal tissues, and high- from low-grade gliomas. This approach may provide useful information for guiding stereotactic biopsies, surgical resection and radiation treatment. Keywords Brain tumor . Magnetic resonance imaging . Magnetic resonance spectroscopy . Cerebral blood volume . Diffusion magnetic resonance imaging

623

Introduction
The therapeutic management and prognosis of cerebral gliomas depend on tumor type and grade, and on exact definition of boundary [13]. Contrast-enhanced MR imaging, or CT scanning if MR is not available or contraindicated, is the method of choice for the noninvasive assessment of gliomas [1, 2]. However, both these imaging techniques have limited sensitivity and specificity in the definition of tumor type, grade and extent [2, 4, 5]. Then, imaging modalities that highlight metabolic, structural and/or functional properties of the tumor can be useful to solve this problem. Proton magnetic resonance spectroscopic imaging (1HMRSI) enables the assessment of the spatial distribution of several brain metabolites. 1H-MRSI is valuable for differentiating active tumor from normal tissue, edema or necrosis, and may improve the identification of the spatial extent and characteristics of brain tumors [511]. However, attempts to correlate tumor types and grades with spectroscopic data have been inconclusive, although studies on pattern recognition analysis have produced encouraging results [11]. Diffusion weighted imaging (DWI) is an MR technique sensitive to motion of water within tissues, and provides information about tumor cellularity and structural integrity [12]. The apparent diffusion coefficient (ADC) can allow the discrimination of tumor mass from necrosis/cysts or normal surrounding tissue [13, 14], but data about the distinction among different tumor types [15, 16] and grades [12, 16, 17], and between tumor and vasogenic edema [12, 15], are controversial. Perfusion weighted imaging (PWI) is an MR technique that looks into the hemodynamics of tumors. Relative cerebral blood volume (rCBV), the most widely used PWI parameter, is directly correlated with tumor microvessel density and neovascularization [18]. It is significantly higher in high- than in low-grade gliomas [1922], but data concerning the differentiation of low-grade gliomas from anaplastic astrocytomas and anaplastic astrocytomas from glioblastomas are conflicting [1922]. In recent years, several authors have used a variable combination of 1H-MRSI, DWI and PWI in addition to conventional MR imaging to improve the ability to differentiate solid tumor from other intratumoral or peritumoral tissues [2325] and among tumor types and/ or grades [4, 2428]. To our knowledge, there is no study in which 1H-MRSI, DWI and PWI information have been combined in a pattern recognition analysis to enhance the discrimination within different tissues. Furthermore, most of these studies have been performed at a magnetic field of 1.5 T. Magnetic fields of higher strength would be expected to provide some advantages, such as higher signal-to-noise ratio and better spatial, temporal and spectral resolution [29].

The purpose of our study was the application of a 3-T MR multiparametric analysis, including 1H-MRSI, DWI and PWI, to determine whether the combination of metabolic, diffusion and hemodynamic data could add relevant information to conventional MR imaging in the assessment of cerebral gliomas.

Methods
Patients Included in this study were 34 consecutive patients with brain gliomas. Three patients had to be excluded because of gross head motion. The age of the remaining 31 patients (20 men and 11 women) ranged between 15 and 76 years (mean age SD, 5515 years). Diagnosis was based on surgical resection or stereotactic biopsy performed in all patients after MR evaluation. Histopathological assessment according to the criteria of the revised World Health Organization revealed: 1 grade I ganglioglioma, 4 grade II fibrillary astrocytomas, 4 grade II oligodendrogliomas, 1 grade II oligoastrocytoma, 1 grade III oligodendroglioma, 2 grade III anaplastic astrocytomas, and 18 grade IV glioblastomas. For statistical purposes, grade I and II gliomas were defined as low-grade (10 patients), and grade III and IV as high-grade (21 patients). The local ethics committee approved the study, and informed consent was obtained from all patients. Imaging All MR studies were performed with a 3.0-T MR scanner (Signa Horizon LX, General Electric Medical Systems, Milwaukee, Wis.) with the standard quadrature head coil. In addition to padding, chin and forehead straps were used to reduce patient motion. Preoperative conventional MR, DWI, 1H-MRSI and PWI were acquired in that order during the same examination to allow exact comparison of results. Total scan-time for all sequences was about 22 min. 1 H-MRSI, DWI and PWI acquisition data were analyzed by an automated software package (Functool Performance, General Electric Medical Systems) in a remote workstation. All image data were checked for overall image quality and motion artifact before analysis. MR imaging included: (a) sagittal T1-weighted localizing images by fast spoiled gradient-echo (FSPGR) sequence (slice thickness 5 mm, gap 1 mm, repetition time, TR 225 ms, echo time, TE 2.5 ms, field of view, FOV 24 cm, flip angle, FA 75, number of excitation, NEX 2, matrix 512256, acquisition time 1 min 57 s); (b) transverse T2-weighted fast spin echo (FSE) images (slice thickness 5 mm, gap 1 mm, TR 5,000 ms, TE 85 ms, FOV 24 cm, NEX 1, matrix 448320, acquisition time 2 min 39 s); (c) transverse fluid-attenuated inversion

624

recovery (FLAIR) images (slice thickness 5 mm, gap 1 mm, TR 11,000 ms, TE 140 ms, inversion time, IT 2,250, FOV 24 cm, NEX 2, matrix 288192; acquisition time 5 min 53 s); (d) gadolinium-enhanced T1-weighted volume-FSPGR images (slice thickness 1.4 mm, TR 225 ms, TE 3.2 ms, FOV 24 cm, FA 15, NEX 1, matrix 320288, acquisition time 23 min) which were acquired finally, after 1H-MRSI, DWI and PWI. Two-dimensional 1H-MRSI was performed by a Proton Brain Exam-Spectroscopic Imaging (PROBE/SI) protocol (General Electric Medical Systems). The location and size of the selected volume, usually of 100200 cm3, were chosen to include as much as possible of tumor, peritumor and normal contralateral brain regions, and to exclude subcutaneous fat and regions with large variations in magnetic susceptibility (i.e. the sinuses). The excitation was a point-resolved spectroscopy (PRESS) pulse sequence, with the following MR parameters: section thickness 10 mm, TR 1,500 ms, TE 144 ms, FOV 24 cm, phase 16 (acquisition time 6 min 53 s). PROBE/SI added

two dimensions of phase encoding that allow multiple spectra to be acquired in a single scan. The resulting spectra of residual water and metabolites were baseline- and frequency-corrected, and extraction of the spectral region of interest (ROI) was applied. The spectroscopic data set, consisting of 256 low-resolution postage stamp images, numbered from 1 to 256, where each of the 256 images represented a point in the spectrum in the range of 4.3 to 0.49 ppm, was transferred to the remote workstation for the acquisition of spectra and chemical shift images. The data set was spatially zerofilled to 3232 images with a final voxel resolution of 7.57.510 mm (0.56 cm3). The peak amplitude of each metabolite was obtained computing the maximum peak values in the range: lipids and/or lactate (LL), 1.581.0 ppm; N-acetylaspartate (NAA), 2.18 2.01 ppm; creatine (Cr), 3.153.0 ppm; choline (Cho), 3.363.21 ppm. Normalization of each metabolite was obtained by dividing its value within the tumor or surrounding ROI by that within an exactly contralateral

Fig. 1 Contrast-enhanced T1-weighted ( a ) and T2-weighted ( b ) images, ADC (c) and rCBV (d) maps, localizing FLAIR image (e) and proton MR spectra (19) (f) from selected ROIs in a 54-yearold man with a left temporoparietooccipital glioblastoma. ROIs: 1 tumor pattern, 2 edema pattern, 3 tumor/edema pattern, 4 infiltrated pattern, 5 normal pattern, 6 necrotic core, 7 tumor margin; 8, 9 contralateral normal. Color scale: red strongest signal

intensity, blue weakest signal intensity. Compared to the edema ROI, the tumor ROI shows a higher Cho peak, an abnormal Cho/ NAA ratio, a lower ADC and a higher rCBV, while the tumor/ edema ROI shows low metabolite peaks, an abnormal Cho/NAA ratio and intermediate ADC and rCBV. Compared to normal, the infiltrated ROI shows a higher Cho peak and an abnormal Cho/ NAA ratio

625

and symmetrical normal ROI, when possible, or within a contralateral representative normal ROI. DWI was performed using single shot, spin-echo type echo-planar imaging resulting in axial images (slice thickness 5 mm, gap 1 mm, TR 1,1000 ms, TE 66.6 ms, FOV 24 cm, matrix 164192, acquisition time 44 s) at b=0 mm2/s and b=1,000 mm2/s as the maximum b value applied along the three orthogonal directions. Images with b=0 provided a set of T2-weighted images, whereas those with b=1,000 mm2/s provided three sets of true diffusion images in the x, y, and z directions that were geometrically averaged to obtain one DWI image. Maps of ADC were generated offline on the remote workstation. ADC was calculated according to the equation: ADC=(1/b)ln(S/S0), where S0 and S are, respectively, the mean signal intensities at b=0 and b=1,000 mm2/s in the same ROI, ln is the natural logarithm and b in 1/b is 1,000. ADC values were expressed in 103 mm2/s. PWI was obtained with single shot, gradient-echo type echo-planar imaging (slice thickness 5 mm, gap 1 mm, TR 1,700 ms, TE 48 ms, FA 90, FOV 22 cm, matrix 164164, NEX 1, slice 12, acquisition time 1 min). After the first ten acquisitions, a bolus of gadopentetate dimeglumine (0.07 mmol/kg) was injected with a mechanical injection pump (3 ml/s) through an 18- or 20-gauge intravenous catheter, followed by a 20-ml saline flush. Since PWI at 3 T will be feasible with a dose of as low as 0.05 mmol/kg [30], we took advantage of higher magnetic fields to reduce the amount of contrast agent in a clinical setting. The echoplanar images were transferred to the remote workstation where the software calculated on a voxel-by-voxel basis:

the curves of signal intensity changes over time, the T2* relaxation rate (R2*), the baseline to be subtracted to the R2* curve, and finally the area under the fitted R2* curve. R2* was calculated by the equation R2*=ln(St/ S0)/TE, where St is the signal intensity at time t, S0 is the signal intensity before contrast agent administration (excluding the first one or two images acquired while reaching the steady-state MR signal) and TE is the echo time. The baseline was estimated by calculating the average background intensity prior to the onset and after completion of the transient signal intensity change. CBV maps were generated by the numerical integration of R2* for the first-pass bolus through each pixel, and rCBV was calculated as rCBV=CBV[tumor]/CBV[contralateral tissue]. Whenever possible, the ROI was placed in mirrored white matter regions; when this was not possible, the ROI was positioned in representative normal white matter in the same transaxial plane. In two low-grade gliomas involving the mesial cortex, the reference ROI were placed in the mirrored gray matter. The rCBV of normal-appearing ROIs surrounding the tumor was obtained by dividing the value of a contralateral and symmetrical normal ROI, when possible, or of a contralateral representative normal ROI. ROIs included the non-necrotic or non-cystic mass, the tumor margins, the areas of abnormal MR signal surrounding the enhancing tumors (perienhancing abnormal ROIs), and the areas of normal-appearing tissue surrounding the enhancing (perienhancing normal ROIs) or nonenhancing tumors (peritumor ROIs). The regions surrounding the tumors, without enhancement or signal alteration in all sequences, were considered as apparently

Table 1 Cho, NAA, Cr, ADC and rCBV (meansSD) in the selected ROIs of high- and low-grade gliomas ROIs High-grade gliomas Mass Margin Perienhancing abnormal Tumor Edema Tumor/edema Perienhancing normal Infiltrated Normal Low-grade gliomas Mass Margin Peritumor Infiltrated (patient no. 1) Infiltrated (patient no. 2) Infiltrated (patient no. 3) Normal
a

No. of observations

Choa

NAAa

Cra

LLa

ADC

rCBVa

14 21 12 10 12 10 13 10 10

1.611.06 1.870.60 1.940.65 0.770.13 0.860.19 1.480.21 1.000.16 1.360.37 1.190.31 1.36 1.31 1.70 1.020.12

0.310.11 0.430.18 0.580.22 0.830.17 0.480.13 0.940.23 0.960.14 0.610.38 0.760.28 0.75 0.85 1.06 1.040.17

0.630.27 0.830.40 1.100.39 0.900.26 0.730.27 1.160.22 0.990.15 1.130.34 0.980.27 1.08 0.97 0.97 1.040.13

6.075.99 3.342.85 2.472.32 1.240.43 1.370.46 1.530.47 1.160.46 1.110.16 1.060.25 1.12 1.00 0.91 0.970.15

1.240.31 1.110.28 0.920.17 1.630.17 1.240.15 0.860.08 0.770.05 1.390.32 1.110.12 0.97 0.92 1.08 0.800.07

3.701.06 3.391.00 1.680.41 0.540.27 1.010.31 1.280.16 0.980.12 1.270.63 1.740.75 1.71 1.66 1.19 1.010.13

10

Values are ratios between pathological and normal-appearing tissue in the opposite hemisphere (see Methods section for details)

626

Table 2 Classification accuracy of perienhancing ROIs with linear discriminant analysis including just 1H-MRSI variables (leave-one-out method) (73.7% of cross-validated grouped cases correctly classified; 77.2% of original grouped cases correctly classified) ROI Predicted group membership Tumor Tumor Tumor/edema Edema Infiltrated Normal 8 0 0 1 0 Tumor/edema 1 10 0 0 0 Edema 0 2 8 0 4 Infiltrated 3 0 0 8 1 Normal 0 0 2 1 8 12 12 10 10 13 Total cases

normal. The ROIs showing spectra with high Cho (>1.3) and/or abnormal Cho/NAA ratio (>1) were considered presumed tumor or tumor-infiltrated on the basis of previous spectroscopic neuropathological correlation studies [5, 9]. Two experienced neuroradiologists, blinded to the histopathological diagnosis and grade, identified in consensus the ROIs, and when they disagreed, a third neuroradiologist adjudicated. The ROIs were automatically copied from the 1H-MRSI to the DWI and PWI images and than, to reduce error linked to the presence of distortion in echo-planar images and/or the non-optimal superimposition between images of various types, a manual repositioning based on anatomical references was performed. ROIs were of 2756 mm2 for DWI and PWI, and 56 mm2 (one pixel) for 1H-MRSI. Statistical analysis Data were analyzed using the SPSS statistical software package (SPSS, Chicago, Ill.). Variables were examined for outliers and extreme values by means of box plots, normal quantile-quantile plots and Shapiro-Wilks test. When normal distribution could not be accepted, variable transformations (square, square root, logarithmic, reciprocal of square root or reciprocal transformations) were reviewed, and the reciprocal of LL helped to improve the distribution shape. Subgroup (ROI) differences were evaluated using analysis of variance and post hoc least significant difference. The number of observations in each subgroup varied from 10 to 21 in high- and from 3 to 10 in low-grade gliomas. Stepwise, cross-validated discriminant

analysis (F to enter 3.84, removal 2.71) was performed to determine MR parameters that better discriminated among perienhancing ROIs and between high- and low-grade gliomas. Cross-validation was performed by the leave-oneout method to assess the accuracy of class membership prediction when using different sets of variables. With the leave-one-out method, discriminant analysis is performed several times, once for each sample in the data set, including all samples except one. The case that is left out is used for testing the classifiers elaborated by the analysis. Classification accuracy was defined as the ratio between the number of cases correctly classified and the total number of cases in the set.

Results
In high-grade gliomas, perienhancing, apparently edematous ROIs showed three 1H-MRSI patterns: presumed tumor tissue (tumor) ROIs, with high Cho (>1.3) and abnormal Cho/NAA ratio (>1); presumed vasogenic edema (edema) ROIs, with metabolite peaks lower than or similar to normal, and a normal Cho/NAA ratio (<1); and presumed tumor-infiltrated edema (tumor/edema) ROIs, with a Cho peak similar to or lower than normal, but an abnormal Cho/NAA ratio (Fig. 1). Tumor ROIs showed lower ADC (P<0.001) and higher rCBV (P<0.001) than edema ROIs. Tumor/edema ROIs showed higher ADC (P<0.001) and lower rCBV (P=0.025) than tumor ROIs, and lower ADC (P<0.001) and higher rCBV (P=0.055) than edema ROIs (Fig. 1; Table 1). Perienhancing, apparently normal ROIs showed two 1H-MRSI patterns:

Table 3 Classification accuracy of perienhancing ROIs with linear discriminant analysis including 1H-MRSI variables plus ADC (leaveone-out method) (84.2% of cross-validated grouped cases correctly classified; 87.7% of original grouped cases correctly classified) ROI Predicted group membership Tumor Tumor Tumor/edema Edema Infiltrated Normal 7 0 0 1 0 Tumor/edema 2 11 0 0 0 Edema 0 1 10 0 0 Infiltrated 2 0 0 8 1 Normal 1 0 0 1 12 12 12 10 10 13 Total cases

627

Table 4 Classification accuracy of perienhancing ROIs with linear discriminant analysis including 1H-MRSI variable plus ADC and rCBV (leave-one-out method) (89.5% of cross-validated grouped cases correctly classified; 91.2% of original grouped cases correctly classified) ROI Predicted group membership Tumor Tumor Tumor/edema Edema Infiltrated Normal 8 0 0 1 0 Tumor/edema 1 11 0 0 0 Edema 0 1 10 0 0 Infiltrated 3 0 0 9 0 Normal 0 0 0 0 13 12 12 10 10 13 Total cases

presumed tumor-infiltrated (infiltrated) ROIs, with Cho >1.3 and/or Cho/NAA >1, and apparently normal (normal) ROIs, with metabolite levels similar to normal (Fig. 1). Stepwise discriminant analysis showed that the set of predictors including Cho, NAA, ADC and rCBV maximized the discrimination among the subgroup of perienhancing ROIs (Fig. 3). The leave-one-out method, in fact, showed that 73.7% of the cases were correctly reclassified when considering only 1H-MRSI variables, 84.2% when considering 1H-MRSI variables and ADC, and 89.5% when considering 1H-MRSI variables, ADC

and rCBV (Tables 2, 3 and 4, respectively). Better classification accuracy was achieved when considering all MR variables. In low-grade gliomas, the tumor mass usually showed all metabolite peaks except LL, but Cho and less frequently Cr were generally higher and NAA lower than normal (Fig. 2). Peritumor infiltrated ROIs (Fig. 2) were detected in three patients (Table 1). All examined MR parameters, except ADC, showed significant differences compared to highgrade gliomas (Table 1). Cho was significantly lower in the margins (P=0.006), NAA higher in the mass (P<0.001) and

Fig. 2 Contrast-enhanced T1-weighted (a) and T2-weighted (b) images, ADC (c) and rCBV (d) maps, localizing FLAIR image (e) and proton MR spectra (19) (f) from selected ROIs in a 30-yearold man with a grade II oligodendroglioma in the left frontal lobe.

ROIs: 1 tumor mass, 2 tumor margin, 3 infiltrated, 4 normal; 5, 6 contralateral normal. Compared to normal (4), the infiltrated ROI (3) shows higher Cho peak and abnormal Cho/NAA ratio

628

Table 6 Classification accuracy of tumor margin with linear discriminant analysis including 1H-MRSI variables (leave-one-out method) (80.6% of cross-validated grouped cases correctly classified; 87.1% of original grouped cases correctly classified) Predicted group membership High-grade High-grade Low-grade 17 2 Low-grade 4 8 21 10 Total cases

separately because in seven cases the tumor was reduced to an enhancing rim surrounding a necrotic core.

Discussion
Fig. 3 Scatter plot shows the distribution of the complete set of cases by using the two first discriminant functions obtained for linear discriminant analysis. With this method, new cases with an unknown diagnosis can be classified according to their distance from the mean discriminant score (centroid) of the respective group. The ordinate and abscissa are scaled in dimensionless units; x axis, value obtained with the first discriminant function (value = 0.0855 Cho + 1.8051 NAA + 7455.63 ADC 1.128 rCBV 7.9548); y axis, value obtained with the second discriminant function (value = 1.4525 Cho 2.2847 NAA + 2830.74 ADC + 1.4924 rCBV 4.7435)

margins (P<0.001), Cr higher in the mass (P<0.001), and rCBV lower in the mass (P<0.001) and margins (P<0.001). LL signal was never detected in low-grade gliomas. Stepwise discriminant analysis showed that discriminant function including LL and rCBV maximized the discrimination between high- and low-grade tumor mass and margins. The leave-one-out method, in fact, showed that 83.3% of tumor masses were correctly reclassified when performing discriminant analysis regarding just 1H-MRSI variables and 100% when regarding 1H-MRSI variables and rCBV, while 80.6% of tumor margins were correctly reclassified when including 1H-MRSI variables and 90.3% when including 1H-MRSI variables and rCBV (Tables 5, 6, 7, 8). Better classification accuracy was achieved when considering 1H-MRSI variables and rCBV; ADC values did not change the classification accuracy. Statistical evaluation of tumor mass and margins was performed
Table 5 Classification accuracy of tumor mass with linear discriminant analysis including 1H-MRSI variables (leave-one-out method) (83.3% of cross-validated grouped cases correctly classified; 83.3% of original grouped cases correctly classified) Predicted group membership High-grade High-grade Low-grade 12 2 Low-grade 2 8 14 10 Total cases

The discrimination of tumor boundaries from normal tissue or vasogenic edema, as well as the evaluation of tissue heterogeneity and tumor grading are often a challenge in conventional MR imaging. Thus, it would be of interest to have a noninvasive, objective and reproducible method for improving the assessment of brain tumor, as well as the planning of surgery and radiation therapy. The present study sought to determine whether combining 3 T 1HMRSI, DWI and PWI information in a pattern recognition analysis, such as linear discriminant analysis, in addition to conventional MR imaging, could enhance the discrimination within tissues, improving the assessment of cerebral gliomas. The results obtained confirm and extend previous MR findings at 1.5 T [2326, 28, 31]. In particular, this study further characterizes the extreme spatial heterogeneity of brain gliomas, identifying peculiar multiparametric MR patterns, such as tumor, edema and tumor/ edema patterns in perienhancing ROIs with abnormal signal on conventional MR imaging, and an infiltrated pattern in perienhancing, apparently normal ROIs. Furthermore, stepwise discriminant analysis showed that the best classification accuracy of perienhancing ROIs with different patterns was achieved when regarding all MR parameters, particularly Cho, NAA, ADC and rCBV (Fig. 3). From the histopathological point of view, regions of altered signal outside the enhancing margins of high-grade gliomas represent a variable combination of vasogenic
Table 7 Classification accuracy of tumor mass with linear discriminant analysis including 1H-MRSI variables plus rCBV (leave-one-out method) (100% of cross-validated grouped cases correctly classified; 100% of original grouped cases correctly classified) Predicted group membership High-grade High-grade Low-grade 14 0 Low-grade 0 10 14 10 Total cases

629

Table 8 Classification accuracy of tumor margin with linear discriminant analysis including 1H-MRSI variables plus rCBV (leave-one-out method) (90.3% of cross-validated grouped cases correctly classified; 90.3% of original grouped cases correctly classified) Predicted group membership High-grade High-grade Low-grade 18 0 Low-grade 3 10 21 10 Total cases

edema and infiltrating tumor cells [32]. A number of studies have suggested that the predominance of tumor cells produces spectra with the typical tumor pattern (i.e. with high Cho peaks and abnormal Cho/NAA ratios) [5, 6, 8, 9, 33], reduces ADC, because of augmented cellularity [12], and increases rCBV, because of incremented size and/ or number of vessels [34]. Furthermore, Cho signal and Cho/NAA ratio are directly proportional to the degree of tumor infiltration beyond the margins of contrast enhancement [5]. These findings are in agreement with the multiparametric pattern that we found in tumor ROIs. If edema is prominent, the increase in unbound interstitial water in normal brain, characteristic of vasogenic edema [35], dilutes the signal of metabolites and produces normal spectra with reduced metabolite peaks [24], increases the extracellular spaces and the diffusion of water [36], and compresses the blood vessels reducing rCBV [24]. These findings are in agreement with the multiparametric pattern that we found in edema ROIs. In the case of tumor-infiltrated edema, the increase in water likely dilutes the typical tumor spectra reducing uniformly the metabolite signals and leaving unchanged the abnormal Cho/NAA ratios. Furthermore, water likely dilutes the absolute cellular density and compresses the increased vascular network of tumor producing intermediate ADC and rCBV, respectively. The finding that perienhancing regions of metastasis show lower Cho and rCBV [24, 27] and higher ADC [15, 27] than those of highgrade gliomas is in line with our hypothesis. Perienhancing regions of metastasis, in fact, represent pure vasogenic edema since they do not contain infiltrating tumor cells [35], while those of high-grade gliomas contain tumor cells infiltrating along the perivascular spaces [32]. In the perienhancing normal-appearing tissues, the presence of ROIs with an infiltrating pattern is in agreement with neuropathological studies showing tumor cells beyond the outer edge of the peritumoral zone as depicted on T2- or contrast-enhanced T1-weighted MR images [5, 37], and 1 H-MRSI findings showing metabolic abnormalities up to 21 mm from the T2 hyperintense margins of glioblastomas [10]. In this study, all examined parameters, except ADC, allowed the differentiation of high- from low-grade gliomas. However, stepwise discriminant analysis showed that LL and rCBV were sufficient for the correct classi-

fication of 100% of tumor masses and 90% of margins. This result is in line with many reports indicating that lactate and/or lipid levels correlate well with the degree of malignancy [5, 28, 38]. However, some authors have concluded that lactate and/or lipids are not useful for differentiating benign from malignant tumors [39]. The increase in rCBV in high-grade gliomas is in agreement with several PWI studies [1822]. With the integration of 3-T MR imagers into clinical practice, there is growing interest in the diagnostic performance of 3-T MR imaging. One of the main advantages of switching to high-field MR is the higher SNR which can allow shorter imaging times for a given resolution, higher resolution for a given imaging time, or combination of both. A short acquisition time is preferable for the fast imaging of ill and sometimes poorly cooperative subjects, especially if long MR protocols are used. High spatial resolution allows high-quality imaging and therefore additional diagnostic information. We did not design this study for comparing 1.5-T with 3-T MR. Notwithstanding, we obtained good quality imaging and spectra in a total acquisition time shorter than previously reported in 1.5-T MR studies [10]. A shorter examination time offers obvious advantages for patients, radiologists, technicians and hospital administrators in the routine clinical setting. Furthermore, the results of this study suggest that the multiparametric MR approach, including 1 H-MRSI, PWI, DWI in addition to conventional MR imaging at 3 T, can be an effective noninvasive tool for the better definition of the spatial extent of brain gliomas. In particular, the exact delineation of margins is of pivotal importance in the overall management of brain tumors, since it allows the planning of stereotactic biopsy and surgical resection, and the delineation of target volume for radiotherapy. A shortcoming of the study is the lack of correlation with histopathological findings. This approach, relatively easy when applied to single voxel studies, becomes rather demanding as the number of studied voxels, and therefore the number of biopsies, increases. Furthermore, the voxels and the biopsy samples have different spatial resolution, and this could hamper the accuracy of the correlation. However, a trial to correlate MR data with histopathological findings and a larger study population will be useful to further support the findings of this study. In addition, given the number and the large variability of MR measurements to be examined, a full or partially automated software package will be useful to produce color maps, which facilitate the recognition of multiparametric data according to tumor component and grade. In conclusion, the combination of metabolic, diffusion and hemodynamic information by means of linear discriminant analysis increased the capacity to discriminate among different tissues. In particular, this multiple information allowed the differentiation of infiltrating tumor from vasogenic edema or normal tissues, and high-

630

from low-grade gliomas. We suggest that such a multiparametric 3-T MR approach, adding 1H-MRSI, PWI and DWI to conventional MR imaging, may be an effective noninvasive tool for the formulation of diagnosis and prognosis, the planning of therapeutic strategies and the monitoring of therapeutic response in patients with brain gliomas.

Acknowledgement The authors are grateful to Piero Ghedin for expert technical assistance. Conflict of interest statement of interest. We declare that we have no conflict

References
1. Behin A, Hoang-Xuan K, Carpentier AF, Delattre J-Y (2003) Primary brain tumours in adults. Lancet 361:323331 2. Grant R (2004) Overview: brain tumour diagnosis and management/Royal College of Physicians guidelines. J Neurol Neurosurg Psychiatry 75 [Suppl 2]: II18II23 3. Schneider JP, Trantakis C, Rubach M, et al (2005) Intraoperative MRI to guide the resection of primary supratentorial glioblastoma multiforme a quantitative radiological analysis. Neuroradiology 47:489500 4. Law M, Yang S, Wang H, et al (2003) Glioma grading: sensitivity, specificity, and predictive values of perfusion MR imaging and proton MR spectroscopic imaging compared with conventional MR imaging. AJNR Am J Neuroradiol 24:19891998 5. Croteau D, Scarpace L, Hearshen D, et al (2001) Correlation between magnetic resonance spectroscopy imaging and image-guided biopsies: semiquantitative and qualitative histopathological analyses of patients with untreated glioma. Neurosurgery 49:823829 6. Li X, Lu Y, Pirzkall A, McKnight T, Nelson SJ (2002) Analysis of the spatial characteristics of metabolic abnormalities in newly diagnosed glioma patients. J Magn Reson Imaging 16:229237 7. Preul MC, Caramanos Z, Leblanc R, Villemure JG, Arnold DL (1998) Using pattern analysis of in vivo proton MRSI data to improve the diagnosis and surgical management of patients with brain tumors. NMR Biomed 11:192200 8. Burtscher IM, Skagerberg G, Geijer B, Englund E, Sthlberg F, Holts S (2000) Proton MR spectroscopy and preoperative diagnostic accuracy: an evaluation of intracranial mass lesions characterized by stereotactic biopsy findings. AJNR Am J Neuroradiol 21:8493 9. Dowling C, Bollen AW, Noworolski SM, et al (2001) Preoperative proton MR spectroscopic imaging of brain tumors: correlation with histopathologic analysis of resection specimens. AJNR Am J Neuroradiol 22:604612 10. Pirzkall A, Mcknight TR, Graves EE, et al (2001) MR-spectroscopy guided target delineation for high-grade gliomas. Int J Radiat Oncol Biol Phys 50:915928 11. Rees J (2003) Advances in magnetic resonance imaging of brain tumours. Curr Opin Neurol 16:643650 12. Sugahara T, Korogi Y, Kochi M, et al (1999) Usefulness of diffusion-weighted MRI with echo-planar technique in the evaluation of cellularity in gliomas. J Magn Reson Imaging 9:5360 13. Tien RD, Felsberg GJ, Friedman H, Brown M, MacFall J (1994) MR imaging of high-grade cerebral gliomas: value of diffusion-weighted echoplanar pulse sequences. AJR Am J Roentgenol 162:671677 14. Brunberg JA, Chenevert TL, McKeever PE, et al (1995) In vivo MR determination of water diffusion coefficients and diffusion anisotropy: correlation with structural alteration in gliomas of the cerebral hemispheres. AJNR Am J Neuroradiol 16:361371 15. Krabbe K, Gideon P, Wagn P, Hansen U, Thomsen C, Madsen F (1997) MR diffusion imaging of human intracranial tumours. Neuroradiology 39:483489 16. Yamasaki F, Kurisu K, Satoh K, et al (2005) Apparent diffusion coefficient of human brain tumors at MR imaging. Radiology 235:985991 17. Lam WWM, Poon WS, Metreweli C (2002) Diffusion MR imaging in glioma: does it have any role in the preoperation determination of grading of glioma? Clin Radiol 57:219225 18. Cha S, Knopp EA, Johnson G, Wetzel SG, Litt AW, Zagzag D (2002) Intracranial mass lesions: dynamic contrastenhanced susceptibility-weighted echo-planar perfusion MR imaging. Radiology 223:1129 19. Lam WWM, Chan KW, Wong WL, Poon WS, Metreweli C (2001) Preoperative grading of intracranial glioma. Acta Radiol 42:548554 20. Jackson A, Kassner A, AnnesleyWilliams D, Reid H, Zhu X-P, Li K-L (2002) Abnormalities in the recirculation phase of contrast agent bolus passage in cerebral gliomas: comparison with relative blood volume and tumor grade. AJNR Am J Neuroradiol 23:714 21. Preul C, Khn B, Lang EW, Mehdorn M, Heller M, Link J (2003) Differentiation of cerebral tumors using multisection echo planar MR perfusion imaging. Eur J Radiol 48:244251 22. Law M, Yang S, Babb JS, et al (2004) Comparison of cerebral blood volume and vascular permeability from dynamic susceptibility contrast-enhanced perfusion MR imaging with glioma grade. AJNR Am J Neuroradiol 25:746755 23. Muti M, Aprile I, Principi M, et al (2002) Study on the variations of the apparent diffusion coefficient in areas of solid tumor in high grade gliomas. Magn Reson Imaging 20:635641 24. Law M, Cha S, Knopp EA, Johnson G, Arnett J, Litt AW (2002) High-grade gliomas and solitary metastases: differentiation by using perfusion and proton spectroscopic MR imaging. Radiology 222:715721 25. Tzika AA, Astrakas LG, Zarifi MK, et al (2003) Multiparametric MR assessment of pediatric brain tumors. Neuroradiology 45:110 26. Yang D, Korogi Y, Sugahara T, et al (2002) Cerebral gliomas: prospective comparison of multivoxel 2D chemicalshift imaging proton MR spectroscopy, echoplanar perfusion and diffusionweighted MRI. Neuroradiology 44:656666 27. Chiang IC, Kuo Y-T, Lu C-Y, et al (2004) Distinction between high-grade gliomas and solitary metastases using peritumoral 3-T magnetic resonance spectroscopy, diffusion, and perfusion imagings. Neuroradiology 46:619627

631

28. Bulakbasi N, Kocaoglu M, rs F, Tayfun C, z T (2003) Combination of single-voxel proton MR spectroscopy and apparent diffusion coefficient calculation in the evaluation of common brain tumors. AJNR Am J Neuroradiol 23:225233 29. Frayne R, Goodyear BG, Dickhoff P, Lauzon ML, Sevick RJ (2003) Magnetic resonance imaging at 3.0 Tesla: challenges and advantages in clinical neurological imaging. Invest Radiol 38:385402 30. Manka C, Trber F, Gieseke J, Schild HH, Kuhl CK (2005) Three-dimensional dynamic susceptibility-weighted perfusion MR imaging at 3.0 T: feasibility and contrast agent dose. Radiology 234:869877

31. Gupta RK, Sinha U, Cloughesy TF, Alger JR (1999) Inverse correlation between choline magnetic resonance spectroscopy signal intensity and the apparent diffusion coefficient in human glioma. Magn Reson Med 41:27 32. Strugar JG, Criscuolo GR, Rothbart D, Harrington WN (1995) Vascular endothelial growth/permeability factor expression in human glioma specimens: correlation with vasogenic brain edema and tumor-associated cysts. J Neurosurg 83:682689 33. Mller-Hartmann W, Herminghaus S, Krings T, et al (2002) Clinical application of proton magnetic resonance spectroscopy in the diagnosis of intracranial mass lesions. Neuroradiology 44:371381 34. Cha S, Johnson G, Wadghiri YZ, et al (2003) Dynamic, contrast-enhanced perfusion MRI in mouse gliomas: correlation with histopathology. Magn Reson Med 49:848855 35. Strugar J, Rothbart D, Harrington W, Criscuolo GR (1994) Vascular permeability factor in brain metastases: correlation with vasogenic brain edema and tumor angiogenesis. J Neurosurg 81:560566

36. Eis M, Els T, Hoehn-Berlage M (1995) High resolution quantitative relaxation and diffusion MRI of three different experimental brain tumors in rat. Magn Reson Med 34:835844 37. Watanabe M, Tanaka R, Takeda N (1992) Magnetic resonance imaging and histopathology of cerebral gliomas. Neuroradiology 34:463469 38. Li X, Vigneron DB, Cha S, et al (2005) Relationship of MR-derived lactate, mobile lipids, and relative blood volume for gliomas in vivo. AJNR Am J Neuroradiol 26:760769 39. Schlemmer H-P, Bachert P, Herfarth KK, Zuna I, Debus J, van Kaick G (2001) Proton MR spectroscopic evaluation of suspicious brain lesions after stereotactic radiotherapy. AJNR Am J Neuroradiol 22:13161324