T2–FLAIR Mismatch, an Imaging Biomarker for IDH and 1p/19q Status in Lower-grade Gliomas: A TCGA/TCIA Project

Diffuse lower-grade gliomas (LGGs) are infiltrative brain neoplasms that include histologic classes astrocytomas, oligodendrogliomas, and oligoastrocytomas and World Health Organization (WHO) grade II and III neoplasms. Histologic classification of LGGs is limited by sampling errors during tumor biopsy or resection and more importantly suffers from high intra- and interobserver variability (1–3). In recent years, genomic analysis has dramatically improved understanding of diffuse gliomas. Several studies, including that by The Cancer Genome Atlas (TCGA) LGG Analysis Working Group, have established that LGGs can be grouped into three robust molecular classes on the basis of mutations in isocitrate dehydrogenase 1 and 2 (IDH1 and IDH2, hereafter collectively referred to as IDH) and codeletion of chromosomes 1p and 19q (4–9). These are: (i) LGGs with IDH mutation and 1p/19q codeletion (IDHmut-Codel); (ii) LGGs with IDH mutation and no 1p/19q codeletion (IDHmut-Noncodel); and (iii) LGGs without IDH mutation and lacking 1p/19q codeletion (IDH-wild type; IDHwt). Among these classes, IDHwt neoplasms are associated with the most aggressive clinical behavior and worst outcome, similar to that of glioblastomas (WHO grade IV). IDHmut-Codel gliomas are associated with the most favorable clinical outcome, and possibly improved sensitivity to procarbazine, lomustine, and vincristine chemotherapy compared with non-codeleted neoplasms (10, 11). IDHmut-Noncodel gliomas are associated with an intermediate outcome, worse than those with 1p/19q codeletion, but far more favorable than IDHwt neoplasms. This molecular classification appears to capture the biologic characteristics and clinical behavior of LGGs more accurately, and with greater fidelity, than traditional histopathologic methods (4), and has been integrated into the 2016 WHO classification of brain tumors (12).

In the context of a shift toward a molecular classification of LGGs, a reappraisal of noninvasive imaging biomarkers of LGGs is warranted. Specifically, the improved molecular understanding of LGGs may advance our understanding of the various imaging phenotypes associated with LGGs. In turn, the capacity of imaging to noninvasively predict the clinical behavior of LGGs may be improved. This would have a high impact on patients with respect to treatment planning and prognostic counseling, especially in settings where detailed molecular assay of glioma specimens is not yet routinely undertaken.

The purpose of this study was to evaluate for the presence of reproducible features on conventional T2-weighted imaging (T2WI) and fluid-attenuated inversion recovery (FLAIR) MRI sequences that correlate with LGG molecular subtype and clinical outcome.

The first portion (test set) of the study was an NIH/NCI-approved retrospective review of the LGG databases of TCGA and The Cancer Imaging Archive (TCIA; ref. 13). All data in these NCI databases are anonymized and, therefore, individual institutional IRB approval is not required for this retrospective review, although it should be noted that all data were originally submitted to TCGA and TCIA by the contributing institutions under an IRB-approved protocol. TCGA LGG tumors were included in this study if they had a corresponding diagnostic MRI study in TCIA.

For the second portion of the study (validation set), IRB approval was obtained from NYU Langone Medical Center for a retrospective review of relevant patient data.

Statistical analysis was carried out utilizing R software (https://cran.r-project.org/, v 3). P values < 0.05 were considered statistically significant. False discovery rates (q) are estimated using the Benjamini–Hochberg method and reported for analysis of high-throughput molecular data.

MRI examinations were available for 199 cases of LGG in the TCIA database. Clinical information, histopathology, and molecular classifications (including IDH and 1p/19q-based molecular subtypes) were obtained from supplementary material from the 2016 pan-glioma paper (5, 14). Missing clinical data and progression free survival times were updated from XML files available through the TCGA data matrix (15).

Seventy-four cases were excluded: (i) MRI examination missing either T2WI or FLAIR sequence (n = 64); (ii) unavailable preresection MRI examination (a prior small diagnostic biopsy was not sufficient for exclusion; n = 7); (iii) infratentorial location of LGG (n = 1); (iv) LGG was excluded from TCGA (n = 1); and (v) IDH and 1p/19q status was unavailable (n = 1). A total of 125 TCGA-profiled LGGs were ultimately included for this analysis.

MRI examinations were analyzed by two independent neuroradiologists with 17 and 3 years of experience, respectively. Both readers have been certified by the American Board of Radiology in both Diagnostic Radiology and also the subspecialty of Neuroradiology. The readers reviewed the cases blinded to the histopathologic diagnosis, molecular classification, and patient outcome.

Readers evaluated the T2WI and FLAIR sequences of each MRI examination. They determined the following characteristics of the LGGs, using a binary scoring system for each: (i) presence or absence of homogenous signal intensity on T2WI; (ii) presence or absence of complete/near-complete hyperintense signal on T2WI, and relatively hypointense signal on FLAIR except for a hyperintense peripheral rim (“T2–FLAIR mismatch”; Fig. 1); (iii) margins of lesion sharp or indistinct; and (iv) presence or absence of peritumoral edema.

After independent data collection, interreader agreement was calculated using the κ coefficient (κ; “irr” package). 0 ≤ κ ≤ 0.2 indicated slight agreement, 0.2 < κ ≤ 0.4 indicated fair agreement, 0.4 < κ ≤ 0.6 indicated moderate agreement, and κ > 0.6 indicated substantial agreement. Discordant results were resolved by consensus for the imaging characteristics which demonstrated moderate or substantial interreader agreement.

On the basis of the results obtained from analyzing LGGs in the TCGA/TCIA database, a separate analysis of LGGs in patients who were managed at our institution (NYU Langone Medical Center) was subsequently undertaken for the purpose of validating the T2–FLAIR mismatch sign only.

LGG cases were derived from an internal database maintained by the Neuropathology department in our institution. One hundred and fifteen LGGs accrued from the years 2011 to 2014 were eligible for analysis. Thirty-three cases were excluded: pediatric patients (n = 22); no available preoperative imaging (n = 8); infratentorial location of LGG (n = 2); no available pathologic data (n = 1). Eighty two cases had preoperative MRI scans with both T2WI and FLAIR sequences, histopathologic diagnosis, and WHO grading, and were included for analysis.

Preoperative MRI scans of these cases were analyzed by the two neuroradiologist readers, who were blinded to the histopathologic diagnosis, molecular classification, and patient outcome. Readers evaluated T2WI and FLAIR sequences to determine the presence or absence of the T2–FLAIR mismatch sign. After independent data collection, interreader agreement was calculated, and discordant results were resolved by consensus.

For TCGA/TCIA cohort, histology and grading of tissues submitted to TCGA were confirmed by neuropathology review, as previously described (4, 5). For the NYU (validation) cohort, the histopathologic diagnosis and WHO grade were ascertained from the electronic medical record.

In addition, an independent neuropathologist performed a blinded review of histopathologic specimens of 30 IDHmut-Noncodel LGGs from the TCGA/TCIA cohort, 15 positive and 15 negative for the T2–FLAIR mismatch sign. Specimens were assessed for cellularity, edema, hemosiderin, microcysts, mucin, vacuolization, and vessel alterations. Tissue samples were of variable quality and consistency, including some cases associated with a single available slide.

Molecular classification for TCGA/TCIA cohort was derived from the 2016 pan-glioma paper (5) and includes (a) somatic mutation in the isocitrate dehydrogenase genes (IDH1 or IDH2), from whole-exome sequencing, (b) codeletion of chromosomal arms 1p and 19q, from Affymetrix SNP6.0 arrays, and (c) the seven methylation specific classes defined for IDH-mutated tumors (Codel, G-CIMP-high, G-CIMP-low) and IDH wild-type tumors (classic-like, mesenchymal-like, LGm6-GBM, and PA-like LGG). Categorical association between consensus MRI metrics and molecular subtype was determined using Fisher exact test for 2 × 2 tables.

A grid plot of genomic alterations (16, 17) was generated for the TCGA cohort, focusing on glioma-specific somatic mutations and copy-number variants (4). The total number of alterations in this gene set was compared between groups by the Wilcoxon rank-sum test. MRNA sequencing profiles for 125 cases were obtained from the Broad Institute TCGA Genome Data Analysis Center (18). Gene set enrichment analysis (GSEA) was performed using the differential expression ranking metric (19). We analyzed pathways from the NCI/Nature Pathway Interaction Database obtained from MSigDB (19, 20). Family-wise error rate (FWER) is reported to account for multiple comparisons. Total protein and phosphoprotein levels, measured by reverse phase protein array (RPPA), were obtained from the Broad Institute TCGA Genome Data Analysis Center (18). Differences between groups were assessed by two-sample t tests. The Core Analysis of Ingenuity Pathway Analysis software was conducted on lists of differential genes found by RNAseq or RPPA data.

For the NYU (validation) cohort, 1p/19q status and IDH status were retrieved from the electronic medical record. Both markers were tested in a Clinical Laboratory Improvement Amendments–certified NYU molecular pathology laboratory. 1p/19q was tested using PCR loss of heterozygosity. IDH status was first screened by IDH1 R132H mutant specific clinically validated antibody (21, 22). In negative cases, IDH mutation status was assessed by a clinically validated IonTorrent targeted next-generation sequencing panel (23, 24).

Overall survival and progression-free survival curves for the TCGA/TCIA cohort were constructed by Kaplan–Meier estimation, and log-rank tests were used to compare survival curves segregated by MRI metrics, Age-adjusted survival estimates were generated using Cox proportional hazards regression modeling. Analyses were performed in R using the “survival” package.

Table 1 displays characteristics of the 125 LGGs from the TCGA/TCIA cohort. Further refinement by DNA methylation profiles yielded 37 Codel, 62 G-CIMP-high, and 3 G-CIMP-low among the IDH-mutant tumors, and 14 mesenchymal-like, six classic-like, three PA-like gliomas among the IDHwt tumors.

Independent MRI analysis revealed interreader agreement measurements for lesion homogeneity [κ = 0.234 (0.111–0.358)], T2–FLAIR mismatch sign [κ = 0.728 (0.538–0.918)], lesion margins [κ = 0.292 (0.135–0.449)], and peritumoral edema [κ = 0.173 (0.096–0.250)]. Discordant cases with respect to the T2–FLAIR mismatch sign were resolved in consensus, and the readers determined that the T2–FLAIR mismatch sign was present in 15 cases (12%) and absent in 110 cases (88%; Figs. 1 and 2). Consensus review was not undertaken for the other imaging metrics due to low levels of interreader agreement. Reader results are displayed in Supplementary Table S1.

Among the 15 cases with positive T2–FLAIR mismatch sign, tumor histology included six astrocytomas, six oligodendrogliomas, and three oligoastrocytomas (of note, designation of “oligoastrocytoma” was assigned in our cohort based on prior WHO classification, and is discouraged based on the 2016 WHO classification). There were 11 grade II and 4 grade III lesions. All 15 cases with a positive T2–FLAIR mismatch sign were IDHmut-Noncodel neoplasms. The T2–FLAIR mismatch sign was significantly associated with IDHmut-Noncodel molecular subtype (P < 0.0001). The T2–FLAIR mismatch sign as a marker of the IDHmut-Noncodel subtype showed a positive predictive value (PPV) of 100% and a negative predictive value (NPV) of 54%. Table 2 displays the distribution of LGGs by histopathologic assignment and the T2–FLAIR mismatch sign, and Table 3 displays the distribution of LGGs by molecular subtype and the T2–FLAIR mismatch sign. Of the 15 T2–FLAIR mismatch positive gliomas, 10 harbored IDH1-R132H mutations and five harbored other IDH mutations, including three IDH1-R132S mutations, one IDH1-R132W mutation, and one IDH1-R132C mutation. With respect to DNA methylation subtype, all 15 LGGs with positive T2–FLAIR mismatch were G-CIMP-high.

Table 1 displays the clinical characteristics of the 82 LGGs from the NYU cohort.

Interreader agreement for the assessment of the T2–FLAIR mismatch was substantial [κ = 0.747 (0.536–0.958)]. After resolving discordant cases in consensus, the readers determined that the T2–FLAIR mismatch sign was present in 10 cases (12%) and absent in 72 cases (88%).

Among 10 cases with a positive T2–FLAIR mismatch sign, tumor histology included six astrocytomas, three oligodendrogliomas, and one oligoastrocytoma. There were five WHO grade II and five WHO grade III lesions. IDH and 1p/19q status was available in all of the positive T2–FLAIR mismatch cases; all 10 of these cases were IDHmut-Noncodel lesions. The T2–FLAIR mismatch sign was significantly associated with the IDHmut-Noncodel molecular subtype (P < 0.00001; PPV = 100%, NPV = 76%). Of the 10 T2–FLAIR mismatch positive gliomas, eight harbored IDH1-R132H mutations and two harbored other IDH mutations, including one IDH1-R132L mutation, and one IDH1-R132C mutation. Table 2 displays the distribution of LGGs by histopathologic assignment and the T2–FLAIR mismatch sign, and Table 3 displays the distribution of LGGs by molecular subtype and the T2–FLAIR mismatch sign.

Genomic alteration profile of LGGs differentiated by the T2–FLAIR mismatch sign showed no significant differences between the groups in rates of mutation or genomic alteration overall when the analysis included all 125 LGGs in the cohort (P = 0.18 and P = 0.27, respectively) and also when the analysis was restricted to IDHmut-Noncodel LGGs (P = 0.73 and P = 0.91, respectively).

Gene expression profile analysis of the 125 LGGs in our TCGA/TCIA cohort using GSEA identified the mTOR pathway as highly enriched with genes upregulated in mRNA profiles of the positive T2–FLAIR mismatch cases (P = 0.01, FWER = 0.695). Among the genes most significantly upregulated in positive T2–FLAIR mismatch mRNA profiles are NRAS (98.8%-tile), RAF1 (92.7 %-tile), MTOR (96.8 %-tile), RICTOR (98.5 %-tile), and RPTOR (81.7 %-tile). All of these genes were in the core set found in the mTOR pathway leading-edge analysis. These results were not reproduced when we restricted our analysis to IDHmut-Noncodel neoplasms alone (Supplementary Tables S2 and S3).

Proteomic analysis of the 125 LGGs in the TCGA/TCIA cohort using RPPA revealed increased levels of proteins in the mTOR signaling pathway among the positive T2–FLAIR mismatch cases, including mTOR (P < 0.0001, adj P = 0.004), RPTOR (P = 0.02, adj P = 0.3), and RICTOR (P = 0.03, adj P = 0.4). Decreased levels of erbB-2 (P < 0.0001, adj P = 0.007) were also noted among the positive T2–FLAIR mismatch cases. When the RPPA was restricted only to IDHmut-Noncodel neoplasms, mTOR (P = 0.002, adj P = 0.3), RICTOR (P = 0.04, adj P = 0.7), and RPTOR (P = 0.04, adj P = 0.7) remained increased among the positive T2–FLAIR mismatch cases, as were several proteins in the AMPK signaling pathway (FOXO3, MET, FASN, ACACA, and mTOR; Supplementary Tables S4 and S5).

Histopathologic analysis of 30 IDHmut-Noncodel neoplasms from the TCGA/TCIA cohort revealed no feature that was significantly different between the T2–FLAIR mismatch positive and T2–FLAIR mismatch negative cases. There was a nonsignificant trend for the presence of abundant microcysts in the T2–FLAIR mismatch positive cases compared with the T2–FLAIR mismatch negative cases (P = 0.128). In the slides that were available for review in the TCGA/TCIA cohort, 11 out of 15 T2–FLAIR mismatch positive cases contained regions composed of numerous, confluent, variably-sized microcystic spaces with only modest intervening neuropil (Supplementary Fig. S1). However, our finding did not reach statistical significance and larger T2–FLAIR mismatch positive cohorts will be required for confirmation.

The median progression-free survival and overall survival of T2–FLAIR mismatch positive gliomas were 38.9 months (30.0–indeterminate) and 65.7 months (43.9–indeterminate), respectively. There were insufficient data to calculate the upper bound of 95% confidence intervals. No significant difference in progression-free survival or overall survival was detected between positive and negative T2–FLAIR mismatch cases when analyzing all of the cases from the TCGA/TCIA cohort, however, this comparison is dependent upon the relative proportion of IDH-mutant and IDHwt gliomas within the cohort. Further survival analysis was undertaken in the IDHmut-Noncodel subset alone (which contained all positive T2–FLAIR mismatch cases), and here also there was no significant difference in survival times between positive and negative T2–FLAIR mismatch cases. The median survival times associated with T2–FLAIR mismatch positive gliomas are similar to established outcomes in IDHmut-Noncodel gliomas, reflecting the molecular subtype that the T2–FLAIR mismatch sign specifically identifies (Supplementary Fig. S2).

Our results indicate that, among LGGs, the T2–FLAIR mismatch sign represents a highly specific marker for the IDHmut-Noncodel molecular subtype of IDH-mutant gliomas, with 100% PPV in both the test and validation sets. Among the imaging metrics assessed in this study, the T2–FLAIR mismatch sign showed a uniquely high level of interreader agreement. Imaging correlates that distinguish between the two IDH-mutant glioma molecular subtypes (1p/19q codeleted and 1p/19q non-codeleted) with high specificity are lacking, and an additional molecular test to interrogate the 1p/19q status is currently required. Development of a robust noninvasive biomarker that can distinguish between these two prognostically distinct IDH-mutant glioma subtypes could assist in the diagnostic evaluation and initial management of these patients.

Multiple recent studies have also attempted to correlate MRI features of gliomas with underlying molecular status. For example, frontal lobe and non-midline tumor locations are more frequent in IDH-mutant lesions (4, 25, 26). IDH-mutant lesions also more frequently display sharp margins, homogenous internal signal intensity, and lesser degree of contrast enhancement when compared with IDH-wild type lesions (26–28). IDHmut-Codel and IDHmut-Noncodel gliomas appear to differ with respect to tumor margins and internal heterogeneity (29). Nonetheless, it is generally accepted that conventional MRI findings do not possess sufficient specificity to predict the underlying histologic or molecular subtype of an LGG in an individual patient (30). In contrast, our results introduce a simple imaging feature which appears to be highly specific for underlying molecular status (although this feature is of low sensitivity). We know of no conventional MRI feature that possesses as high a level of specificity in predicting underlying IDH and 1p/19q status as appears to be the case for the T2–FLAIR mismatch sign.

Advanced MRI techniques such as perfusion-weighted imaging (31–33), MR-spectroscopy (34–36), and sodium-imaging (37) have also revealed differences between IDH-mutant and wild-type neoplasms. Among these advanced techniques, the most promising appears to be 2-hydroxyglutarate (2HG) detection utilizing proton MR-spectroscopy. Detection of 2HG is a highly specific marker for underlying IDH mutation. Unfortunately, 2HG detection with MR-spectroscopy is a significant technical challenge due to spectral overlap of 2HG with background metabolites, and is not yet a standardized and routinely performed imaging sequence in most clinical practices. Moreover, it does not predict the 1p/19q status of gliomas.

Our finding that the T2–FLAIR mismatch sign identifies a molecular subset of gliomas (100% PPV for IDHmut-Noncodel) with a far more favorable survival profile than IDHwt glioma (in keeping with the well-known fact that IDH-mutant gliomas are far more indolent than IDHwt gliomas (4, 6, 8, 38–41)) might provide useful information to clinicians during the initial management of patients. Unambiguous radiographic evidence that a glioma is IDH-mutant and, therefore, a relatively favorable subtype would assist patient counsel. Moreover, recent studies indicate that patients with IDH-mutant gliomas significantly benefit from gross total tumor resection compared with partial resection (42), and the survival benefit achieved with gross total resection may have greater impact in the IDHmut-Noncodel subset (43). Therefore, the T2–FLAIR mismatch sign might provide useful information for the neurosurgeon prior to glioma resection. Although surgery should be as extensive as safely possible, regardless of IDH mutation status, a surgeon may be willing to extend further (e.g., supplementary motor area) if it is known that a tumor is IDH mutant.

Our study attempted to assess imaging features besides the T2–FLAIR mismatch sign. These included imaging features related to internal signal texture (homogenous or inhomogeneous), tumor margin characteristics (sharp or indistinct), and presence/absence of peritumoral edema, that have in prior studies been correlated with histologic and molecular class as well as the clinical behavior of gliomas (27, 44–46). However, the interreader agreement for each of these metrics was poor in our study. It should be noted that the readers in our study assessed all metrics in a qualitative and binary manner, one which we thought was most directly applicable to clinical neuroradiology practice. Prior studies assessing some of these imaging metrics have instead used quantitative and semi-quantitative methods that allowed for a graded assessment of tumor characteristics. It is possible that the T2–FLAIR mismatch sign showed such high interreader agreement due to its fairly striking and characteristic imaging appearance, one that conforms readily to a binary (yes/no) scoring system.

To our knowledge, a version of the T2–FLAIR mismatch sign in glioma has only been reported once previously. In a case series of protoplasmic astrocytomas, Tay and colleagues reported that all cases showed at least partial “suppression” of FLAIR signal relative to the signal on T2WI (47). Protoplasmic astrocytomas are rare WHO grade II diffuse gliomas which typically affect young adults and follow a relatively indolent course (48). Therefore, it is likely that this case series was largely composed of IDH-mutant tumors (even though the study belongs to premolecular classification era and does not mention molecular analysis). The MRI figures displayed in their article strongly resemble the appearance of the T2–FLAIR mismatch sign reported in our study.

It is unclear why only some, but not all, IDHmut-Noncodel lesions manifest the T2–FLAIR mismatch sign. One possibility is that the T2–FLAIR mismatch sign identifies some as-yet undiscovered molecular subgroup within the IDHmut-Noncodel class. In our exploratory RPPA analysis, tumors in the IDHmut-Noncodel class showed significantly increased levels of proteins in the mTOR pathway among the positive T2–FLAIR mismatch cases, including mTOR, RICTOR, and RPTOR. The GSEA also revealed increased expression of genes in the mTOR pathway among T2–FLAIR mismatch positive cases (though only when the analysis included all cases from our cohort). Several recent studies showed IDH-mutant gliomas frequently acquire activating mutations in the mTOR pathway at progression (49–51). However, our findings remain preliminary and whether mTOR is activated in a subset of IDH-mutant gliomas requires validation.

Our study has several limitations. It followed a retrospective design, and the results would benefit from prospective validation. However, our usage of the publically available TCGA/TCIA database readily permits other groups to reanalyze this data and independently verify our results. Our study utilized subjective MRI interpretation by two readers of variable experience, rather than quantitative or semi-quantitative image analysis. Although in theory, this might limit the reproducibility of our results, our analysis utilized routinely available MR imaging sequences (T2WI, FLAIR) and the assessment of relatively simple imaging features, potentially permitting wide clinical applicability and more straightforward future validation efforts compared with more technically demanding and investigational MRI sequences. Finally, although the T2–FLAIR mismatch sign appears to represent a highly specific sign of IDH-mutant lesions among LGGs, its specificity amongst all possible brain neoplasms cannot be ascertained based on our data.

Among LGGs, the T2–FLAIR mismatch sign represents a highly specific noninvasive imaging biomarker for the IDHmut-Noncodel molecular subclass of IDH-mutant glioma. This simple and robust MRI biomarker can be identified using standard clinical MRI techniques and may enable a more informed pretreatment management plan and patient counsel.

L. Cooper reports receiving speakers bureau honoraria from Varian Medical Systems. No potential conflicts of interest were disclosed by the other authors.

Conception and design: S.H Patel, L.M. Poisson, R. Jain

Development of methodology: S.H Patel, M. Snuderl, R. Jain

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S.H Patel, D.J. Brat, C. Thomas, A.M. Franceschi, J.G. Golfinos, R. Jain

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S.H Patel, L.M. Poisson, D.J. Brat, Y. Zhou, L. Cooper, M. Snuderl, C. Thomas, A.S. Chi, R. Jain

Writing, review, and/or revision of the manuscript: S.H Patel, L.M. Poisson, D.J. Brat, L. Cooper, M. Snuderl, A.M. Franceschi, B. Griffith, J.G. Golfinos, A.S. Chi, R. Jain, A. Flanders

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S.H Patel, A.M. Franceschi, R. Jain

Study supervision: S.H Patel, R. Jain

L. Cooper received financial support from National Cancer Institute grant U24CA194362. The remaining authors received no financial support for their contributions to this article.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.