Glioma Grading: Sensitivity, Specificity, and Predictive Values of Perfusion MR Imaging and Proton MR Spectroscopic Imaging Compared with Conventional MR Imaging

Methods

Results

The sensitivity, specificity, PPV, and NPV for determination of a high-grade glioma with conventional MR imaging were 72.5%, 65.0%, 86.1%, and 44.1%, respectively. For a minimum C2 error, a threshold value of 1.75 for rCBV provided sensitivity, specificity, PPV, and NPV of 95.0%, 57.5%, 87.0%, and 79.3%, respectively. For a minimum C1 error, a threshold value of 2.97 for rCBV provided sensitivity, specificity, PPV, and NPV of 72.5%, 87.5%, 94.6%, and 51.5%, respectively (Table 1). A threshold value of 2.97 provided the same sensitivity as that of conventional MR imaging but higher specificity and PPV. A threshold value of 2.18 provided the same specificity as that of conventional MR imaging but higher sensitivity, PPV, and NPV (Table 1).

A threshold value of 1.08 for Cho/Cr provided the minimum C2 error and 97.5%, 12.5%, 77.0%, and 62.5% for the sensitivity, specificity, PPV, and NPV for determination of a high-grade glioma. A threshold value of 1.56 provided the minimum C1 value and 75.8%, 47.5%, 81.2%, and 39.6% for the sensitivity, specificity, PPV, and NPV for determination of a high-grade glioma. Threshold values that provided the same sensitivity and specificity as those of conventional MR imaging were 1.61 and 1.88, respectively (Table 2).

A threshold value of 0.75 for Cho/NAA provided minimum C2 error and 96.7%, 10.0%, 76.3%, and 50.0% for the sensitivity, specificity, PPV, and NPV for determination of a high-grade glioma. A threshold value of 1.6 for Cho/NAA provided minimum C1 error and 74.2%, 62.5%, 85.6%, and 44.6% for the sensitivity, specificity, PPV, and NPV for determination of a high-grade glioma. Threshold values that provided the same sensitivity and specificity as those of conventional MR imaging were 1.66 and 1.78, respectively (Table 3).

The combination of rCBV, Cho/Cr, and Cho/NAA resulted in sensitivity, specificity, PPV, and NPV of 93.3%, 60.0%, 87.5%, and 75%, respectively, for minimum C2 error, and 70.8%, 92.5%, 96.6%, and 51.4%, respectively, for minimum C1 error. For the same sensitivity as that of conventional MR imaging (72.5%), the specificity and PPV were higher, and for the same specificity as that of conventional MR imaging (65.0%), the sensitivity, PPV, and NPV were higher than those of conventional MR imaging alone when using the combination of rCBV, Cho/Cr, and Cho/NAA (Table 4, Fig 3).

• Download figure
• Open in new tab
• Download powerpoint

Fig 3.

ROC curves for rCBV plus metabolites, rCBV alone, Cho/Cr, and Cho/NAA demonstrate superior sensitivity and specificity of rCBV plus metabolites and rCBV alone compared with conventional MR imaging (cMRI, green asterisk) for glioma grading.

The mean, standard deviation, and ranges for rCBV, Cho/Cr, Cho/NAA, and NAA/Cr are shown in Table 5. A statistically significant difference was noted in the rCBV, Cho/Cr, Cho/NAA, and NAA/Cr between low- and high-grade gliomas (P < .0001, .0121, .001, and .0038, respectively) (Table 5). Within the high-grade glioma group, there were 73 anaplastic gliomas (anaplastic astrocytomas, n = 26; anaplastic oligodendrogliomas, n = 7; anaplastic mixed oligoastrocytomas, n = 40) and 47 glioblastoma multiforme. No statistically significant difference was noted in the rCBV, Cho/Cr, Cho/NAA, and Cho/Cr ratios between the anaplastic gliomas and glioblastoma multiforme groups (P = .549, .302, .363, and .915, respectively).

Discussion

Current methods of grading gliomas have inherent limitations. The current reference standard of histopathologic grading can be inaccurate when biopsy samples are not taken from the most malignant tumor region or when the tumor is not completely resected. This is a particular problem with glioma because of the infiltrative proliferation of the tumor. Although histopathologic grading is often performed on the enhancing portion of the tumor, vascular networks in the peritumoral region serve as a path for tumoral infiltration along perivascular spaces. The region of highest vascularity and malignancy may then be within the so-called peritumoral or perienhancing region (22).

Radiologic grading of tumors with conventional MR imaging is not always accurate, with sensitivity in identifying high-grade gliomas ranging from 55.1% to 83.3% in other studies (7, 8, 11) and 72.5% in this study. Yet, accurate tumor grading has important implications for treatment planning: Patients with an erroneous diagnosis of high-grade glioma will undergo unnecessary adjuvant therapy; patients with an erroneous diagnosis of low-grade glioma will be treated conservatively, with concomitant morbidity and mortality.

Results of several previous studies suggest that rCBV measurements may improve grading. Sugahara et al (18) correlated maximal rCBV values histologically and angiographically in 30 patients, with mean values of 7.32, 5.84, and 1.26 for glioblastomas, anaplastic astrocytomas, and low-grade gliomas, respectively. Aronen et al (16) found mean maximal rCBV values of 3.64 and 1.11 in high- and low-grade gliomas, respectively (n = 19). Knopp et al (11) had similar mean maximal rCBV values of 5.07 and 1.44 in high- and low-grade gliomas, respectively (n = 29). These values are comparable to our findings, with mean maximal rCBV values of 5.18 and 2.14 for high- and low-grade gliomas, respectively (Table 5, n = 160). Lev and Rosen (23) used an rCBV threshold value of 1.5 in discriminating among 32 consecutive patients with glioma. Thirteen (100%) of 13 astrocytomas were correctly categorized as high-grade gliomas. Three of these did not enhance after administration of contrast material. Of the nine low-grade astrocytomas, seven were correctly classified. The sensitivity and specificity with use of an rCBV of 1.5 as a threshold value were 100% and 69%, respectively. This compares with the results from this study of 95.0% and 57.5% sensitivity and specificity, respectively, by using 1.75 as the threshold value. More recently, Shin et al (24) demonstrated mean rCBV ratios of 4.91 in high-grade gliomas and 2.00 in low-grade gliomas, in 17 patients; these ratios were similar to the ratios in this study of 5.18 and 2.14 in high- and low-grade gliomas, respectively. The threshold or cutoff value of 2.93 with use of ROC curve analysis is comparable to our value of 2.97 minimizing for C1 error (Table 1).

Besides vascular proliferation, cellularity, mitotic activity, nuclear pleomorphism, and necrosis are important criteria in histopathologic grading of gliomas. Ki-67 labeling is used in histologic examination as a marker for cellular proliferation. A higher rate of Ki-67–positive cells corresponds to greater malignancy in gliomas. Metabolite ratios, in particular Cho levels, have correlated with Ki-67 levels in gliomas (58). MR spectroscopic measurements of Cho/Cr and Cho/NAA ratios should therefore be helpful in the grading of gliomas.

However, to date, there have been few systematic attempts to compare the sensitivity, specificity, PPV, and NPV of perfusion MR imaging and MR spectroscopy with those of conventional MR imaging in glioma grading. To make a comparison, we determined threshold values from logistic regression and ROC analyses. However, it is not always clear what criteria should be used in determining an “optimum” threshold. For example, one could choose to minimize C1 error, which minimizes the average of the false-positive and false-negative error rates. This would be appropriate if the consequence of misclassifying low-grade gliomas is the same as that of misclassifying high-grade gliomas and the two are equally likely to be presented to you for classification. Alternatively, one could choose to minimize the C2 error, the total number of misclassified tumors observed in the data. This adjusts for a difference in the relative frequency of low- and high-grade gliomas in the patient population. In particular, if high-grade gliomas are much more prevalent, then a high misclassification rate of high-grade gliomas would result in a high total number of misclassified gliomas. In the real world, high-grade gliomas are more common than low-grade gliomas. Hence, choosing the threshold to minimize C2 error will tend to yield high sensitivity and relatively low specificity. These are the reasons for presenting two sets of results with threshold values for rCBV and metabolite ratios that minimize both C1 and C2 error values. For application in clinical practice, however, a threshold that minimizes C2 error is most appropriate (Tables 1–3).

The sensitivity of rCBV in glioma grading in our study was 95.0% (minimizing for C2 error), indicating a high true-positive rate and low false-negative rate. Hence, if the rCBV is above 1.75, there is a high probability that the tumor will be a high-grade glioma. Conversely, when rCBV is below 1.75, the tumor is unlikely to be high-grade glioma. However, the relatively low specificity means that false-positive rates are relatively high and true-negative rates are correspondingly low. In other words, some low-grade gliomas will be falsely identified as high-grade gliomas. However, this is the lesser of two evils. First, low-grade gliomas are relatively less common than high-grade gliomas, so fewer errors will be made in absolute terms. Second, a low-grade glioma misidentified as a high-grade glioma will be treated aggressively with some increase in morbidity. However, a high-grade glioma misidentified as a low-grade glioma will be treated conservatively, resulting in potential rapid death.

The high NPV (79.3%) is likewise a significant finding, as gliomas with low rCBV (<1.75) are unlikely to have high-grade components. Hence, it is an excellent tool for excluding the presence of a high-grade glioma (23), an issue that often confronts the neuroradiologist and neuropathologist. The low total error when using rCBV alone suggests that rCBV in itself is an accurate predictor of tumor grade, with a 14.4% chance of error when using 1.75 as the arbitrary threshold value. When one chooses threshold values that demonstrate the same sensitivity as that of conventional MR imaging (72.5%), the specificity, PPV, and NPV are superior to those of conventional MR imaging. Similarly, when one chooses threshold values that demonstrate the same specificity as that of conventional MR imaging (65.0%), the sensitivity, PPV, and NPV are superior to those of conventional MR imaging (Table 1).

The sensitivities (minimized for C2 error) of Cho/Cr and Cho/NAA of 97.5 and 96.7%, respectively, in this study confirm that metabolite ratios can be useful in determining tumor grade. However, the low specificities (12.5% for Cho/Cr and 10.0% for Cho/NAA) are due to the high levels for Cho that we, and others, have observed in some low-grade gliomas. Again, however, high sensitivity in identifying high-grade gliomas is more important than high specificity because of the relatively fewer cases of low-grade glioma and the more serious consequences of false-negative findings.

Despite the low specificities, a significant difference was noted in Cho/Cr, Cho/NAA, and NAA/Cr ratios (P = .012, .001, and .004, respectively) for differentiating between low- and high-grade gliomas (Table 5). A review of the literature, taking into account differences in MR spectroscopic technique such as the choice of TE and method for determination of metabolite ratios, demonstrates that the mean maximal values obtained for Cho/Cr and Cho/NAA and mean minimum values for NAA/Cr (Table 5) in our study are comparable to previously published data in differentiating between low- and high-grade gliomas (7, 31, 45, 49, 59). Other data in the literature present metabolite levels and direct metabolite concentrations in a slightly different manner than presented in the current study, but they clearly also show differences in Cho levels between low- and high-grade gliomas (35, 54). In the current study, we had a number of reasons for using Cho/Cr and Cho/NAA ratios for assessing tumor grade. First, histopathologically, the linear correlation of Cho with Ki-67 labeling index or cellular proliferative activity suggests that Cho may be a strong predictor of tumor grade (57, 58). Second, although it may be possible to provide better discrimination combining other metabolites such as alanine, lipids, lactate, myo-inositol, glutamine, and glutamate, it is often not possible to perform MR spectroscopy at multiple TEs in a clinical setting to provide metabolites such as myo-inositol, glutamine, and glutamate detectable best at short TE. Third, to analyze multiple metabolites, linear discriminant analysis (50) or automated spectral analysis software (69) is usually necessary but not commercially available as yet.

One of the challenges in spectroscopy, even with current automated techniques, is obtaining reliable and reproducible intrapatient and interpatient data. To ensure quality control in the data, measurements of normal Cho/Cr and NAA/Cr levels were also obtained in the contralateral unaffected white matter as part of the MR spectroscopy multivoxel, 2D CSI measurement. Despite possible regional variations in metabolite ratios within the brain, it was reassuring to find that normal values for Cho/Cr and Cho/NAA obtained in the low- and high-grade glioma groups were almost identical. Normal Cho/Cr and NAA/Cr levels were 0.88 and 1.72 for low-grade gliomas and 0.87 and 1.73 for high-grade gliomas (Table 5). This not only provides a means for ensuring that the MR spectroscopic data are reliable, but also allows for further comparison between the abnormal data and normative data. In this study, the comparison was made only between low- and high-grade gliomas; however, a comparison between metabolite ratios within the tumor and normal voxels in the same patient is shown in Table 5.

The role of necrosis in glioma grading is important (70). The presence of necrosis is one important distinction between anaplastic astrocytomas and glioblastoma multiforme. In our series, lipid and lactate were found in 5.0% of low-grade gliomas and in 16.6% of high-grade gliomas. There is certainly a difference in finding lipids and lactate between the two groups, and although formal quantification and analyses were not performed in this study, lipids and lactate do correlate with necrosis in high-grade glioma and may also be useful in differentiating glioma grades (27, 45, 71). The frequency of elevated lipid and lactate in the high-grade glioma group was relatively low compared with that of previous studies (45, 71). This may be explained by the population bias in the sample of 73 anaplastic gliomas versus 47 glioblastomas. Furthermore, as this study did not focus on lipids and lactate, formal quantification of these metabolites was not performed, and hence small amounts of lipids and lactate that may be obscured by baseline noise may not be detected.

In a clinical setting, where decisions such as extent of tumoral resection and addition and dose of postoperative chemotherapy, radiation therapy, and interval of follow-up must be made, threshold values can be used as important supplementary information in the noninvasive, neuroradiologic grading of gliomas. A lesion that may have nonspecific conventional MR imaging findings (lack of contrast material enhancement, no mass effect, no necrosis, and no edema), but that demonstrates an rCBV value of 7.72 and Cho/NAA ratio of 2.60, for example, has high sensitivity and specificity for being a high-grade glioma (Fig 1). A comparative lesion (Fig 2) with conventional MR imaging findings suggestive of a high-grade glioma (contrast material enhancement, necrosis, mass effect, and edema) but that demonstrates an rCBV of 1.70 and a Cho/NAA ratio of 0.90 is more in keeping with a low-grade glioma. Hence, using these threshold values in cases that may have nonspecific conventional MR imaging findings increases our confidence in glioma grading.

Of further clinical interest is the differentiation between anaplastic gliomas and glioblastoma multiforme in the high-grade glioma group. As with our findings, previously published data in a group of 26 high-grade gliomas (11) also indicate no significant difference in rCBV between anaplastic astrocytomas and glioblastoma multiforme. Discriminating between these two groups has been demonstrated with in vitro (72) and in vivo (71, 73) MR spectroscopy. A limitation in our study is the perfusion and spectroscopic imaging of a number of patients (n = 43) after surgery. Although it is possible that surgery and postoperative changes may affect tumor vascularity and cellularity, we believe that because these patients only had either partial resection or biopsy, the final statistical analysis would not have been affected greatly by this. We examined the preoperative patients separately, and the results and conclusion did not differ significantly. Another potential limitation of our study is the semiquantification of metabolite ratios by using an internal reference method within the same voxel. The effects of tumor heterogeneity, regional differences in absolute and relative metabolite concentrations, and various pathologies on metabolite T2 relaxation times is difficult to assess. Hence, using Cr as a reference metabolite may increase variability and inaccuracy. Methods to normalize Cho to contralateral normal Cho or contralateral normal Cr as well as methods to provide absolute quantification of metabolites in gliomas have been described (59, 74). Normalizing our results to contralateral normal Cho did not seem to improve statistical significance substantially; however, it would seem that absolute metabolite quantification methods should improve the sensitivity and possibly the specificity of MR spectroscopy (74). Finally, the lower sensitivity and specificity of Cho compared with rCBV suggests that although Cho levels may correlate directly with tumor cellularity, tumor cellularity may not correlate directly with tumor grade. Tumor grade seems to be more reliably correlated with necrosis, nuclear atypia, mitoses, and vascular hyperplasia (13); hence, rCBV may provide stronger correlation with tumor grade than does Cho.

This study demonstrates that rCBV and metabolite measurements can improve preoperative tumor grading. MR spectroscopy and perfusion MR imaging are useful adjuncts to conventional MR imaging in planning postoperative chemotherapy, antiangiogenic therapy, and radiation therapy (75–77). Since these techniques avoid some of the problems of sampling error associated with histopathologic examination, it is conceivable that such methods may provide a more accurate overall assessment of tumors. Demonstrating that rCBV and MR spectroscopy are sensitive techniques in this study is an important step in this regard. However, the question of whether one can improve the long-term outcome for patients with gliomas remains. Long-term outcome studies are required to determine whether rCBV, Cho levels, or pathology is the best predictor of patient outcome.

Conclusion

Preoperative grading of gliomas based on conventional MR imaging is often unreliable. Independently and in combination, rCBV measurements and Cho/Cr and Cho/NAA ratios can significantly improve the sensitivity and predictive values of preoperative glioma grading. The rCBV measurements are the best parameter for glioma grading. This is extremely important for determining the most optimal therapy regimen and the regularity and aggressiveness of postoperative follow-up and treatment. At our institution, perfusion MR imaging and MR spectroscopy not only have become a routine part of glioma imaging, but also have become an important adjunctive tool in providing information in gliomas and regions of the brain that may have eluded histopathologic assessment. Perfusion MR imaging and MR spectroscopy can potentially overcome the limitation of sampling error with histopathologic grading of tumor by the ability to sample the entire lesion noninvasively in vivo. Ongoing data collection from longitudinal studies is crucial to determine if rCBV and metabolite ratios can, in the long run, be comparable to histopathologic examination in predicting tumor behavior and patient prognosis.