Secondary Logo

Share this article on:

Hybrid 11C-MET PET/MRI Combined With “Machine Learning” in Glioma Diagnosis According to the Revised Glioma WHO Classification 2016

Kebir, Sied, MD*†‡§∥; Weber, Manuel, MD§∥¶; Lazaridis, Lazaros, MD*§∥; Deuschl, Cornelius, MD**; Schmidt, Teresa, MD*§∥; Mönninghoff, Christoph, MD††; Keyvani, Kathy, MD‡‡; Umutlu, Lale, MD**; Pierscianek, Daniela, MD§∥‡‡; Forsting, Michael, MD**; Sure, Ulrich, MD§∥‡‡; Stuschke, Martin, MD§§; Kleinschnitz, Christoph, MD∥∥; Scheffler, Björn, MD†‡§∥; Colletti, Patrick M., MD¶¶; Rubello, Domenico, MD***; Rischpler, Christoph, MD§∥¶; Glas, Martin, MD*†‡§∥

doi: 10.1097/RLU.0000000000002398
Original Articles

Purpose With the advent of the revised WHO classification from 2016, molecular features, including isocitrate dehydrogenase (IDH) mutation have become important in glioma subtyping. This pilot trial analyzed the potential for 11C-methionine (MET) PET/MRI in classifying glioma according to the revised WHO classification using a machine learning model.

Methods Patients with newly diagnosed WHO grade II–IV glioma underwent preoperative MET-PET/MRI imaging. Patients were retrospectively divided into four groups: IDH wild-type glioblastoma (GBM), IDH wild-type grade II/III glioma (GII/III-IDHwt), IDH mutant grade II/III glioma with codeletion of 1p19q (GII/III-IDHmut1p19qcod) or without 1p19q-codeletion (GII/III-IDHmut1p19qnc). Within each group, the maximum tumor-to-brain-ratio (TBRmax) of MET-uptake was calculated. To gain generalizable implications from our data, we made use of a machine learning algorithm based on a development and validation subcohort. A support vector machine model was fit to the development subcohort and evaluated on the validation subcohort. Receiver operating characteristic (ROC) analysis served as metric to assess model performance.

Results Of a total of 259 patients, 39 patients met the inclusion criteria. TBRmax was highest in the GBM cohort (TBRmax 3.83 ± 1.30) and significantly higher (P = 0.004) compared to GII/III-IDHmut1p19qnc group, where TBRmax was lowest (TBRmax 2.05 ± 0.94). ROC analysis showed poor AUC for glioma subtyping (AUC 0.62) and high AUC of 0.79 for predicting IDH status. In the GII/III-IDHmut1p19qcod group, TBR values were slightly higher than in the IDHmut1p19qnc group.

Conclusions MET-PET/MRI imaging in pre-operatively classifying glioma entities appears useful for the assessment of IDH status. However, a larger trial is needed prior to translation into the clinical routine.

From the *Division of Clinical Neurooncology, Department of Neurology, and

DKFZ-Division Translational Neurooncology at the West German Cancer Center (WTZ), University Hospital Essen, University Duisburg-Essen;

Translational Oncology German Cancer Consortium, Partner Site University Hospital Essen;

§West German Cancer Center (WTZ), University Hospital Essen, University Duisburg-Essen;

German Cancer Consortium, Partner Site University Hospital Essen;

Department of Nuclear Medicine,

**Institute of Diagnostic and Interventional Radiology and Neuroradiology, Departments of

††Neuropathology,

‡‡Neurosurgery,

§§Radiotherapy, and

∥∥Neurology, University Hospital Essen, University Duisburg-Essen, Essen, Germany;

¶¶Department of Radiology, University of Southern California, Los Angeles, CA; and

***Department of Nuclear Medicine, Radiology, NeuroRadiology, Clinical Pathology, S. Maria della Misericordia Hospital, Rovigo, Italy.

Received for publication September 25, 2018; revision accepted October 12, 2018.

S. K., M. W., C. R. and M. G. are shared authorship.

Conflicts of interest and sources of funding: none declared.

Ethics Approval and Consent to Participate.

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. For this retrospective study formal consent is not required.

Correspondence to: Martin Glas, MD, Division of Clinical Neurooncology, Department of Neurology, University Hospital Essen, University Duisburg-Essen, D-45147 Essen, Germany. E-mail: Martin.Glas@uk-essen.de; and Domenico Rubello, Department of Nuclear Medicine, Radiology, NeuroRadiology, Clinical Pathology, S. Maria della Misericordia Hospital, Via Tre Martiri; 35043, Rovigo, Italy. E-mail: domenico.rubello@aulss5.veneto.it.

Gliomas are heterogeneous tumors in terms of malignancy and outcome. With the ever-growing exploration of the molecular basis of tumor evolution, the genetic profile of gliomas has grown to play a decisive role.1 Consequently, in the 2016 World Health Organization (WHO) classification of tumors of the central nervous system, there has been a shift from a morphology-oriented approach to a molecularly oriented one. Of these molecular markers assessed in clinical routine, isocitrate dehydrogenase (IDH) and 1p19q-codeletion have emerged as the most important variants with regards to prognosis and expected treatment response.2

The reference standard tool for the morphological portrayal of gliomas is contrast-enhanced MRI. Non-enhancing tumors imply an intact blood–brain barrier, with less aggressive tumor histology and growth.3 However, even when the blood–brain barrier is intact, the grade of malignancy of gliomas and the corresponding clinical outcomes may vary substantially.4 An IDH-wildtype glioma, for example, classified as WHO grade II with radiological features of low-grade glioma may exhibit molecular and clinical features in tandem with high-grade glioma.5 Therefore, there is a need for non-invasive techniques capable of capturing the grade of malignancy more accurately as can be done using conventional MRI.

In this context, PET imaging could help reveal high-grade malignancies by investigating metabolic tumor activity.6 In neuro-oncology and, in particular, with malignant glioma, amino acid tracers such as O-(2-[18F]fluoroethyl)-L-tyrosine (FET) and 11C-methionine (MET) have increasingly demonstrated utility.7 Verger et al. showed recently that dynamic and static FET-PET imaging constitutes a promising tool for the non-invasive prediction of IDH status in glioma, while a reliable differentiation between 1p19q codeleted and 1p19q non-codeleted glioma was not possible.8

The PET tracer used in this retrospective pilot study is MET, whose uptake is known to be increased in gliomas. A high tumor-brain-ratio (TBR), which is characterized by a high tracer-uptake in the tumor region compared to the unaffected brain tissue is suspicious for a high-grade malignancy.9 Recent studies on MET-PET in the context of glioma have shown that MET-PET serves as a valid prognostic tool10 and may allow for the distinction between radiation necrosis and recurrence.11 On the downside, MET bears the risk of a high tracer-uptake in non-tumor lesions like inflammatory tissue and due to its short-half-life it is limited to centers with access to a cyclotron.12

The aim of this retrospective pilot study is to assess how closely TBR correlates with histological results and the molecular profile of newly diagnosed glioma patients prior surgery and to assess the value of hybrid 11C-MET PET/MRI in this context. Another unique feature of this study is the adoption of a machine learning algorithm that allows for an estimation of how well our data may be generalized. To this end, we use a separate validation cohort to test the diagnostic accuracy of our model on glioma subtyping.

Back to Top | Article Outline

METHODS

Study Design

In this retrospective study, 259 patients who underwent 11C-MET-PET/MRI from 1st January 2012 until 10th February 2017 were screened for the presence of the following inclusion criteria:

  • Patients underwent 11C-PET/MRI at the University of Essen Medical Center for the diagnostic workup of newly diagnosed glioma based on MRI findings.
  • After PET/MRI, either biopsy or resection of the brain lesion was performed.
  • Postoperative pathohistological report confirmed the diagnosis of glioma WHO grade II–IV
  • Complete assessment of relevant molecular markers in accordance with 2016 WHO guidelines.1
Back to Top | Article Outline

MET-PET/MRI Analysis

All simultaneous PET/MRI scans were performed on a Siemens mMR Biograph with 3.0 Tesla and a lutetium oxyorthosilicate (LSO)-detector as previously described.13 A mean of 843.1 MBq (range, 266–1185 MBq) MET was administered via a peripheral venous access. Following a short delay, PET/MRI was performed. PET/MRI analysis included the calculation of TBR values. The maximum TBR (TBRmax) was calculated by dividing maximal SUV (SUVmax) in the tumor by SUVmax in an unaffected part of the contralateral brain region encompassing white and grey matter. The mean TBR (TBRmean) was obtained by dividing mean SUV (SUVmean) in the tumor by SUVmean in the contralateral region.

Back to Top | Article Outline

Neuropathologic Evaluation

Following surgery, tumor tissue underwent an integrated morphologic and molecular diagnostic analysis at the Institute of Neuropathology at the University of Essen Medical Center. The final diagnosis was made on specific criteria dictated by the revised WHO guidelines.1 O6-methylguanine-DNA-methyltransferase (MGMT)-promoter methylation was assessed by pyrosequencing. Isocitrate Dehydrogenase (IDH) analysis was done either via immunohistochemistry or DNA sequencing. DNA sequencing was performed primarily with young patients below the age of 55 when immunohistochemistry (IHC) yielded a negative result. Detection of 1p and 19q loss were performed by quantitative microsatellite analysis.

Back to Top | Article Outline

Molecular Groups

Patients were divided into different subgroups according to the integrated histopathologic and molecular diagnoses. Grade II and III glioma patients were grouped together according to their IDH mutation status as follows: IDH wild-type GBM glioblastoma, WHO grade II and III IDH wildtype (GII/III-IDHwt) glioma, WHO grade II and III IDH mutant and 1p19q non-codeleted (GII/III-IDHmut1p19qnc) glioma, and WHO grade II and III IDH mutant and 1p19q codeleted (GII/III-IDHmut1p19qcod) glioma.

Back to Top | Article Outline

Machine Learning Algorithm

For the purpose of assessing diagnostic performance, we split our study cohort in half with random allocation and while accounting for differences in the distribution of molecular groups. The development cohort consisted of 20 patients while the validation cohort consisted of 19 patients. We then used a supervised learning model and trained a support vector machine classifier with a linear kernel (lSVM) on one part of the cohort.14 A threefold cross-validation approach was chosen to enhance classification performance in the development cohort. A grid search over a set of C values [0–100] was used to optimize hyperparameters. The lSVM model was then validated on the remaining subcohort. To evaluate classifier output quality, we used the Receiver Operating Characteristic (ROC) metric. The ROC metric is quantified by the area under the curve (AUC) with values close to one indicating a good model, while values closer to 0.5 reflecting a non-relevant model. Since ROC curves are usually used for binary classification to study classification performance, we calculated the AUC by binarizing each of the molecular groups and evaluated each against the remainder and averaged ensuing AUC values (micro-averaging).

Back to Top | Article Outline

Statistical Analysis

In each of the four groups, we analyzed TBR values on the preoperative MET-PET/MRI and compared their values among groups. To this end, the mean TBR values with the respective standard deviations were calculated separately for each group. Furthermore, we employed a univariate analysis of variance (ANOVA) model. We used Fisher's Least Significant Difference (LSD) method to test for significant differences among groups. P values below 0.05 were considered significant. Stata (version 14) and Python (version 3.5.2) were used for statistical calculation and visualization.

Back to Top | Article Outline

RESULTS

Patients’ Characteristics

Of 259 patients, a total of 39 patients met the inclusion criteria for this retrospective study (Fig. 1). Of those, eight (21%) were later diagnosed with GBM. Thirteen (33.3%) patients segregated into the GII/III-IDHwt group. Eleven (28%) patients fell into the GII/III-IDHmut1p19qnc glioma entity. Finally, seven (17.9%) patients were diagnosed with GII/III-IDHmut1p19qcod glioma. Patients enrolled were between 24 and 75 years old (mean, 46.2 years) at the time of tumor diagnosis with IDHmut1p19qnc patients being youngest (mean, 39.9 years). A detailed overview of the assessed clinical and histopathological results is given in Table 1. Images of corresponding MET-PET and MRI scans for selected subgroups are provided in Figure 2A–D.

FIGURE 1

FIGURE 1

TABLE 1

TABLE 1

FIGURE 2

FIGURE 2

Back to Top | Article Outline

Metabolic Activity Among Molecular Subgroups

The highest value for TBR was found in the GBM subgroup (mean TBRmax (TBRmax), 3.83 ± 1.30; mean TBRmean (TBRmean), 3.72 ± 1.30). To determine whether TBRmax and TBRmean differed significantly among molecular entities, we performed an ANOVA with LSD as post-hoc analysis. Here, we found that both TBRmax and TBRmean were significantly higher in the GBM subgroup as compared to the GII/III-IDHmut1p19qnc subgroup (TBRmax in GBM group, 3.83 ± 1.30 versus TBRmax in GII/III-IDHmut1p19qnc group, 2.05 ± 0.94; P = 0.004; mean TBRmean 3.72 ± 1.30 versus 1.98 ± 0.85; Fig. 3A–C). In addition, TBRmean was significantly higher in the GBM subgroup as compared with GII/III-IDHmut1p19qcod group, while differences in TBRmax values were only of borderline significance (TBRmax in GBM group 3.83 ± 1.30 versus 2.68 ± 0.84; P = 0.060; TBRmean in GBM group 3.72 ± 1.30 versus 2.54 ± 0.79; P = 0.049; Fig. 3A–C).

FIGURE 3

FIGURE 3

TBR values in the GII/III-IDHwt group were only slightly lower than in the GBM group with a TBRmax of 3.20 ± 1.13 and a TBRmean of 3.06 ± 1.14 (TBRmax in GBM group, 3.83 ± 1.30; TBRmean in GBM group 3.72 ± 1.30). As expected, when performing ANOVA, differences did not reach statistical significance (P = 0.308 for TBRmax; P = 0.373 for TBRmean). There was no significant difference among GII/III-IDHwt, GII/III-IDHmut1p19qnc, and, GII/III-IDHmut1p19qcod groups with respect to tracer-uptake (P = 0.76).

Back to Top | Article Outline

Diagnostic Value of TBRmax in Detecting Glioma Entities According to WHO Classification From 2007

To allow for an assessment of the diagnostic value of TBRmax in detecting glioma entities that were diagnosed according to the WHO classification from 2007, we built two separate cohorts from our original cohort and fitted a lSVM as described above. The diagnostic performance is plotted on a receiver operator characteristic chart (Fig. 4A, B). As shown in Figure 4A, the diagnostic capacity of TBRmax in separating each of the tumor entities, diagnosed pursuant to WHO 2007 criteria, is 60% and rather low. The diagnostic performance is not increased significantly (AUC 64%) when subgrouping the entire cohort into low-grade (defined as WHO grade II) and high-grade (defined as WHO grade III and IV) glioma groups as evident in Figure 4B.

FIGURE 4

FIGURE 4

Back to Top | Article Outline

Diagnostic Value of TBRmax in Detecting Glioma Entities According to the Revised WHO Classification From 2016

When evaluating the TBRmax performance according to its potential in classifying glioma entities as established by the revised WHO classification from 2016, there is still a poor performance (AUC, 62%) as shown in Figure 4C. Intriguingly, the diagnostic performance increases to an AUC of 79% when considering the study cohort by whether IDH is mutant or not, irrespective of grading (Fig. 4D).

Back to Top | Article Outline

DISCUSSION

This pilot study shows that hybrid MET-PET/MRI derived TBR values may not reliably classify glioma entities prior to surgery but could successfully help discriminate between IDH mutant and wild-type glioma. Furthermore, this study adopts a novel machine learning approach in PET research by providing validated data on hybrid MET-PET/MRI performance in glioma classification. Unlike conventional reportings, the usage of a validated machine learning model allows an estimation of how well our model may be generalized on unseen data and thus provides a higher level of evidence.

This study is novel for several reasons. First, we used preoperative simultaneous hybrid PET/MRI obtained during one single session in each patient allowing for a highly standardized approach. In addition, different from the standard approach in PET analysis, we adopted a machine learning algorithm to ensure that our findings may be generalizable. The standard approach in PET analysis consists of conducting an analysis on a single study cohort. This allows inferences on that study cohort but lacks information on whether these results may be applied on a different cohort. Using machine learning, as has been done in this study, enables information on future data. This approach allows for a much more precise depiction of diagnostic performance.15

The diagnostic performance with regard to separating each glioma entity according to the 2007 WHO classification is rather poor and slightly improved when using glioma classification according to the revised WHO guidelines from 2016 that incorporates molecular details. However, it remains unsatisfactory and may, therefore, not be suited for reliable classification. It remains to be elucidated whether advanced PET techniques, such as textural analysis or dynamic PET analysis may provide superior results.

Our study highlights, that hybrid MET-PET/MRI may be well suited for identifying IDH mutant glioma. These findings are in line with previous reports showing that MET-PET bears a particularly useful potential for differentiating IDHmut from IDHwt glioma.10 In this pilot study, we provide robust data by training a support vector classifier on one part of the cohort and then validating the results on the remaining cohort that was not included in the primary analysis. Differentiating IDH mutant from IDH wild-type glioma has also been shown to be feasible using FET-PET.8 The ability to distinguish IDH mutant glioma from wild-type is clinically relevant in the event of the presence of space-occupying lesions suspicious for glioma within eloquent brain areas that are not amenable to biopsy when a decision needs to be made on initiating treatment.

Interestingly, we found that MET-uptake was higher in the GII/III-IDHmut1p19qcod group as compared to the GII/III-IDHmut1p19qnc group. This is rather counterintuitive since MET-uptake has been associated with malignancy and GII/III-IDHmut1p19qnc tumors are known to come with worse survival as compared to the codeleted counterpart. A similar finding has also been observed in a study by Saito et al., confirming that MET-uptake was significantly higher in 1p19q codeleted WHO grade II glioma as opposed to non-codeleted WHO grade II glioma.16 One possible explanation for this observation might constitute previous findings showing that 1p19q codeleted tumors go along with increased blood perfusion as compared to the non-codeleted counterpart, as evidenced by MR perfusion imaging studies showing elevated relative cerebral blood volume in 1p19q codeleted glioma.17

Shortcomings of this study are mainly the small sample size and its retrospective nature, which is why our results need confirmation in a larger study. Yet, our study cohort is homogeneous and molecularly well characterized with no missing data. One advantage of this study is that all of the MET-PET/MRI scans were performed on the same hybrid PET/MRI system. Additionally, pathohistological analyses were all carried out in one institution, thus reducing possible misdiagnoses.

In conclusion, this pilot study provides validated data on the performance of hybrid 11C-methionine PET/MRI imaging in pre-operatively classifying glioma according to the revised WHO guidelines. The diagnostic performance appears good for the separation of IDH mutant and IDH wild-type glioma. This bears important clinical implications for evaluating glioma suspicious lesions on MRI that are not accessible to biopsy.

Back to Top | Article Outline

REFERENCES

1. Louis DN, Perry A, Reifenberger G, et al. The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary. Acta Neuropathol. 2016;131:803–820.
2. Brat DJ, Cagle PT, Dillon DA, et al. Template for reporting results of biomarker testing of specimens from patients with tumors of the central nervous system. Arch Pathol Lab Med. 2015;139:1087–1093.
3. Li D, Zhang J, Ji N, et al. Combined 68Ga-NOTA-PRGD2 and 18F-FDG PET/CT can discriminate uncommon meningioma mimicking high-grade glioma. Clin Nucl Med. 2018;43:648–654.
4. Kunz M, Thon N, Eigenbrod S, et al. Hot spots in dynamic (18)FET-PET delineate malignant tumor parts within suspected WHO grade II gliomas. Neuro Oncol. 2011;13:307–316.
5. Reuss DE, Kratz A, Sahm F, et al. Adult IDH wild type astrocytomas biologically and clinically resolve into other tumor entities. Acta Neuropathol. 2015;130:407–417.
6. Dunet V, Pomoni A, Hottinger A, et al. Performance of 18F-FET versus 18F-FDG-PET for the diagnosis and grading of brain tumors: systematic review and meta-analysis. Neuro Oncol. 2016;18:426–434.
7. García Vicente AM, Jiménez Aragón F, Villena Martín M, et al. 18F-fluorocholine PET/CT, brain MRI, and 5-aminolevulinic acid for the assessment of tumor resection in high-grade glioma. Clin Nucl Med. 2017;42:e300–e303.
8. Verger A, Stoffels G, Bauer EK, et al. Static and dynamic 18F-FET PET for the characterization of gliomas defined by IDH and 1p/19q status. Eur J Nucl Med Mol Imaging. 2018;45:443–451.
9. Pöpperl G, Kreth FW, Mehrkens JH, et al. FET PET for the evaluation of untreated gliomas: correlation of FET uptake and uptake kinetics with tumour grading. Eur J Nucl Med Mol Imaging. 2007;34:1933–1942.
10. Rani N, Singh B, Kumar N, et al. Differentiation of recurrent/residual glioma from radiation necrosis using semi quantitative 99mTc MDM (bis-methionine-DTPA) brain SPECT/CT and dynamic susceptibility contrast-enhanced MR perfusion: a comparative study. Clin Nucl Med. 2018;43:e74–e81.
11. Terakawa Y, Tsuyuguchi N, Iwai Y, et al. Diagnostic accuracy of 11C-methionine PET for differentiation of recurrent brain tumors from radiation necrosis after radiotherapy. J Nucl Med. 2008;49:694–699.
12. Arora G, Sharma P, Sharma A, et al. 99mTc-methionine hybrid SPECT/CT for detection of recurrent glioma: comparison with 18F-FDG PET/CT and contrast-enhanced MRI. Clin Nucl Med. 2018;43:e132–e138.
13. Beppu T, Terasaki K, Sasaki T, et al. MRI and 11C-methyl-L-methionine PET differentiate bevacizumab true responders after initiating therapy for recurrent glioblastoma. Clin Nucl Med. 2016;41:852–857.
14. Huang S, Cai N, Pacheco PP, et al. Applications of support vector machine (SVM) learning in cancer genomics. Cancer Genomics Proteomics. 2018;15:41–51.
15. Jordan MI, Mitchell TM. Machine learning: trends, perspectives, and prospects. Science. 2015;349:255–260.
16. Saito T, Maruyama T, Muragaki Y, et al. 11C-methionine uptake correlates with combined 1p and 19q loss of heterozygosity in oligodendroglial tumors. AJNR Am J Neuroradiol. 2013;34:85–91.
17. Jenkinson MD, Smith TS, Joyce KA, et al. Cerebral blood volume, genotype and chemosensitivity in oligodendroglial tumours. Neuroradiology. 2006;48:703–713.
Keywords:

glioma; methionine; PET/MRI; machine learning

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.