Molecular Classification of Patients with Grade II/III Glioma Using Quantitative MRI Characteristics

Abstract

Background:

Molecular markers of WHO grade II/III glioma are known to have important prognostic and predictive implications and may be associated with unique imaging phenotypes. The purpose of this study is to determine whether three clinically relevant molecular markers identified in gliomas—IDH, 1p/19q, and MGMT status—show distinct quantitative MRI characteristics on FLAIR imaging.

Methods:

Sixty-one patients with grade II/III gliomas who had molecular data and MRI available prior to radiation were included. Quantitative MRI features were extracted that measured tissue heterogeneity (homogeneity and pixel correlation) and FLAIR border distinctiveness (edge contrast; EC). T-tests were conducted to determine whether patients with different genotypes differ across the features. Logistic regression with LASSO regularization was used to determine the optimal combination of MRI and clinical features for predicting molecular subtypes.

Results:

Patients with IDH wildtype tumors showed greater signal heterogeneity (p = 0.001) and lower EC (p = 0.008) within the FLAIR region compared to IDH mutant tumors. Among patients with IDH mutant tumors, 1p/19q co-deleted tumors had greater signal heterogeneity (p = 0.002) and lower EC (p = 0.005) compared to 1p/19q intact tumors. MGMT methylated tumors showed lower EC (p = 0.03) compared to the unmethylated group. The combination of FLAIR border distinctness, heterogeneity, and pixel correlation optimally classified tumors by IDH status.

Conclusion:

Quantitative imaging characteristics of FLAIR heterogeneity and border pattern in grade II/III gliomas may provide unique information for determining molecular status at time of initial diagnostic imaging, which may then guide subsequent surgical and medical management.

Introduction

WHO grade II and III gliomas include a heterogeneous group of infiltrative neoplasms with astrocytic and oligodendroglial morphology. These tumors have a wide range of both progression-free survival (PFS) and overall survival (OS); some respond to therapy with OS greater than 13 years, while others precipitously progress to glioblastoma (GBM) [1,2]. Although the prognosis for grade II/III gliomas was previously thought to depend mostly on histopathological grade, it is now recognized that OS is highly influenced by specific molecular markers.

Over the past several decades, tremendous progress has been made in revealing the underlying molecular alterations that influence prognosis in grade II/III gliomas [3]. Recent studies have shown that the molecular characterization of glioma [i.e., isocitrate dehydrogenase (IDH), codeletion of chromosome arms 1p and 19q (1p/19q), and methylguanine methyltransferase promoter methylation (MGMT status)] is more robust for the prediction of clinical outcomes in comparison to histological classification [1,3–7]. Codeletion of 1p/19q was initially determined to be a prognostic marker associated with oligodendroglial morphology and shown to predict response to chemoradiation [8–11]. IDH gene mutations are thought to be an early step in gliomagenesis, and are estimated to occur in 79% to 94% of grade II/III gliomas [12]. The status of IDH mutation [i.e., IDH mutant (mt) vs IDH wild-type (wt) tumors] has also been well-validated as both a prognostic and predictive marker in glioma[5,13–16]. In fact, it has been shown that most adult grade II/III glioma patients who are IDH-wt resemble GBM, molecularly and clinically [1,17] and all are presumed to have 1p19q intact [17]. In addition, MGMT promoter methylation status has known prognostic and predictive value. In patients with grade III gliomas, response to alkylating chemotherapy is better with MGMT promoter methylation; however, this may be due to the common co-occurrence of IDH mutations and MGMT promoter methylation [7,18].

Because both molecular and histological classification require invasive measures (i.e., biopsy/resection), there is increased interest in identifying non-invasive surrogates for tumor genotypes [19–25]. In recent years, distinct imaging features have been identified on MRI and other imaging modalities (i.e., MR spectroscopy) that have shown initial promise for classifying tumors by specific molecular markers [26,27]. Qi et al. showed that IDH-mt tumors were more likely to have less contrast enhancement (CE) in patients with astrocytic tumors than IDH-wt tumors [26]. Also, Tietze et al. showed in vivo MRS for 2-hydroxyglutarate identified IDH status in almost 88% of the gliomas [27]. In previous studies, it was also noted that IDH-mt tumors, regardless of the tumor grade, were more likely to have sharp tumor margins and homogeneous signal intensity relative to IDH-wt tumors [26,28] On the contrary, “indistinct” tumor margins have been described in patients with 1p/19q codeleted tumors [29–31]. Similarly, 1p/19q codeleted tumors are reported to have more heterogeneity of T2 signal compared to intact tumors [32,33]. It has been proposed that the increased heterogeneity in 1p/19q codeleted tumors is associated with calcification and paramagnetic susceptibility, which increases heterogeneity of the T2 signal [32,33]. In addition to IDH and 1p/19q codeletion status, MGMT status has been associated with specific T2 features [34]. Noushmehr et al. found that in GBM, methylated tumors are more likely to show a higher level of tumor infiltration and more indistinct borders relative to unmethylated tumors [16]. However, T2/FLAIR features associated with MGMT status have not been studied in patients with grade II/III gliomas.

While previous studies have reported that IDH, 1p/19q, and MGMT status appear phenotypically different on conventional imaging in subsets of glioma patients [35,36], few studies have evaluated all three markers simultaneously in patients with grade II or III glioma. Importantly, most of the existing studies have been based on qualitative descriptions, which may vary greatly across readers and sites, explaining some of the contradictory findings described in previous reports [26]. Thus, there is a need for reliable, quantitative imaging features that can differentiate tumor genotypes without clinician and/or researcher bias.

In this study, we investigate the ability of quantitative FLAIR and T1-post-contrast images to aid in differentiating molecular subtypes of patients with grade II/III gliomas. In particular, given recent evidence that features of the FLAIR signal may be especially informative, we introduce several novel imaging parameters obtained from FLAIR texture analysis that measure the heterogeneity (i.e., signal homogeneity/heterogeneity and pixel correlation) and border patterns (Edge Contrast; EC) within the FLAIR signal. We test the ability of our quantitative texture features to differentiate molecular subtypes compared to FLAIR and CE volumes. We accomplish this by comparing all patients who are IDH-wt to those who are IDH-mt. Next, we test the contribution of 1p19q status by comparing patients who are IDH-mt and 1p19q codeleted to those who are IDH-mt and 1p19q intact. Patients who are IDH-wt are excluded from this analysis due to evidence that all are 1p19q intact [1]. Finally, in an exploratory analysis, we examine our quantitative imaging features in patients who are MGMT methylated versus unmethylated. Based on our previous work [36] and other published studies [37], we hypothesize that: 1) measures of FLAIR heterogeneity will be greater in patients with IDH-wt compared to IDH-mt tumors. In those with IDH-mt tumor, we hypothesize that heterogeneity will be greater in 1p/19q-codeleted tumors compared to those with 1p/19q-intact tumors; 2) FLAIR borders will be less distinct in patients with IDH-wt compared to IDH-mt tumors. In those with IDH-mt, FLAIR borders will be less distinct in 1p/19q-codeleted compared to 1p/19q-intact tumors. Finally, we hypothesize that a combination of quantitative features derived from FLAIR and CE images will provide better classification of patient tumors according to molecular status compared to any single imaging feature.

Methods and Materials

Patients

This retrospective study was approved by the institutional review board. From 2010 to 2017, 115 patients with grade II/III gliomas were identified at our institution who had MR imaging, including 3D IR-SPGR pre- and post- contrast and 3D FLAIR sequences. Patients were excluded if they did not have the specific MR sequences performed at our institution prior to radiation (n = 42), there was significant artifact on imaging (n=3), or molecular information was not available (n= 9). Sixty-one patients with grade II/III gliomas met inclusion criteria to form the final study cohort (see figure 1). The final cohort included 32 males and 29 females, and the average age was 46.18 years with a range of 23 to 71 years (see table 1). All patients had undergone either biopsy (N = 20) or resection (N = 41) prior to MRI acquisition. However, all imaging was performed before start of radiation or chemotherapy.

Figure 1.

Consort diagram illustrates the exclusion and inclusion criteria and the final cohort.

Table 1.

Demographic and clinical characteristics of the patient sample.

Characteristic	Total (N=61)
Sex – no. (%)
Male	32 (52.5)
Female	29 (47.5)
Age – year
Median	45.0
Range	23.0–71.0
Histopathology – no. (%)
Oligodendroglioma	19 (31.1)
Astrocytoma	13 (21.3)
Anaplastic Astrocytoma	12 (19.7)
Anaplastic Oligodendroglioma	10 (16.4)
Oligoastrocytoma	5 (8.2)
Anaplastic Oligoastrocytoma	2 (3.3)
Tumor Grade – no. (%)
Grade II	35 (57.4)
Grade III	26 (42.6)
Location – no. (%)
Left hemisphere	36 (59.0)
Right hemisphere	25 (41.0)
1p/19q status – no. (%)
Codeleted	29 (47.5)
Intact	30 (49.2)
Not available	2 (3.3)
IDH status – no. (%)
Mutant	43 (70.4)
Wild-type	11 (18.0)
Not available	7 (11.4)
MGMT status – no. (%)
Methylated	23 (37.7)
Unmethylated	9 (14.8)
Not available	29 (47.5)
Procedure – no (%)
Resection	41 (67.2)
Biopsy	20 (32.8)

Molecular Analysis

OncoScan microarray analysis (Affymetrix, Santa Clara, CA) or fluorescence in situ hybridization (FISH) was performed on formalin fixed paraffin embedded tissue to determine the 1p/19q status. For most patients, IDH status was determined by whole exome next generation sequencing on a panel of known cancer genes that included IDH1 and IDH2. Two patients had IDH status determined by immunohistochemistry for the R132H mutation. MGMT promoter methylation status was determined by Methylation-specific PCR. Two of the patients that were found to have the 1p/19q codeletion by FISH, prior to routine testing for IDH, were included in the 1p/19q codeletion and IDH-mt group. As shown in figure 1, at least one molecular marker was missing for a subset of patients (IDH status; n = 7, 1p/19q status; n = 2, MGMT status, n = 29).

Magnetic Resonance Imaging

All MRI scans were acquired on a 3.0 T GE Signa Excite HDx scanner using an 8-channel head coil. The imaging protocol included pre- and post-gadolinium 3D T1-weighted inversion recovery-spoiled gradient recalled echo (IR-SPGR) with TE/TR = 2.8/6.5 ms, TI = 450 ms, flip angle = 8 degrees. FOV = 24 cm, voxel size = 0.93×0.93×1.2 mm, and a 3D T2-weighted FLAIR sequence with TE/TR = 126/6000 ms, TI = 863, FOV = 24 cm, voxel size = 0.93×0.93×1.2 mm. Prior to analysis, raw data were corrected for bias field and distortion [38]. Subsequently, correction for patient motion was carried out using in-house software. The pre- and post-contrast 3D IR-SPGR and FLAIR images were registered to each other using rigid body registration.

Image Pre- and Post- Processing

Contrast-enhanced volumes (CE_VOL) and FLAIR hyperintense volumes (FLAIR_VOL) were segmented semi-automatically (Amira software package, Visage Imaging) on the co-registered post-contrast 3D IR-SPGR and FLAIR images, while regions of necrosis and the resection cavity were excluded. Volumes of interest (VOIs) were drawn for the entire tumor volume by 2 trained image analysts (R.D. and S.H.) and approved by a board-certified neuroradiologist with expertise in neuro-oncology (N.F.). Quantitative texture analyses were conducted to measure the signal heterogeneity and FLAIR borders using three-dimensional co-occurrence matrix (3D-COM) and EC, respectively.

Signal heterogeneity:

3D-COM was applied to extract the heterogeneity features from the FLAIR hyperintense ROI. Histogram normalization was applied to the MR images when generating the gray-level co-occurrence matrix (GLCM) prior to the texture analysis [39]. This method provides statistics on the probability density function P_d,θ (i, j), that is the probability of finding a joint relationship between a pair of pixels composed of a central pixel of gray level i and a neighboring one of gray level j. These two pixels are separated by a distance d (pixel distance) and angle θ of one of the four values (0°, 45°, 90° and 135°) [40]. Texture features extracted by the 3D-COM method are represented by homogeneity and pixel correlation. Homogeneity is a metric that indicates the level of homogeneity/heterogeneity of the FLAIR signal on a global level (i.e., within the entire VOI), whereas pixel correlation indicates the level of homogeneity/heterogeneity on a “local” level (i.e., across adjacent pixels). Higher homogeneity indicates more uniformity of the entire signal, and higher pixel correlation indicates more homogeneity across neighboring voxels.

Border distinctness:

EC is defined as the gradient magnitude of the lesion edges, where higher EC indicates a sharper and more distinct border compared to lower EC. A 3D analysis was applied to the FLAIR VOI to enhance the local precision and decrease the partial volume effect [41–44]. The technique of EC extraction has been described in our previous study [45], where the morphological operations of erosion and dilation were applied to the FLAIR_VOL binary mask using a spherical 3D mask. Then, the contour of the FLAIR_VOL binary mask was drawn in 3D indexing the surface of the FLAIR _VOL lesion. The gradients of the FLAIR image were calculated and were overlaid on the surface of the 3D binary mask to create the hyperintense surface representing EC.

Quantitative MRI measurements including FLAIR_VOL, CE_VOL, ratio of FLAIR_VOL/CE_VOL, homogeneity, pixel correlation, and EC were calculated for each patient. Despite the complexity of the model and processing pipeline, the parameters are generated in a very time efficient manner within the seconds.

Statistical Analysis

Independent t-tests were used to test for differences in continuous variables as functions of IDH, 1p19q, and MGMT status. A chi square test was used to test for differences in sex distribution. A generalized linear model (GLM) was used to test the association between imaging paramaters and molecular status of the patient using a logit linear function. The Least Absolute Shrinkage and Selection Operator (LASSO) was used to select imaging predictors for use in a logistic regression model in order to predict each genotype status. Model selection was based on the Akaike Information Criteria (AIC). A second model was run that included age and gender to determine the additional value of adding these variables to the prediction of each marker.

Results

Patients with IDH-wt tumors were significantly older than the patients with IDH-mt tumors (p = 0.006). Within the IDH-mt group, patients with 1p/19q co-deleted tumors were also older than patients with 1p/19q intact tumors (p = 0.011). There were no significant age differences between the two MGMT groups (p = 0.42). There were no significant differences in sex distribution across IDH and 1p/19q groups (p > 0.05), whereas there was a higher proportion of females in the methylated (1 male vs 8 female) compared to the unmethylated (10 male vs 13 female) group (p = 0.02) (see table 2).

Table 2.

Imaging parameters for the three molecular subgroups; 1) IDH wildtype vs mutant, 2) IHD mutant/1p/19q codeleted vs IDH mutant/1p/19q intact, and 3) MGMT methylated vs MGMT unmethylated.

	IDH wildtype (n=11) Mean(range)	IDH mutant (n = 43) Mean(range)	p-value	Test Statistic	HDH-mt 1p/19q codeleted (n = 24) Mean(range)	IDH-mt 1p/19q intact (n = 19) Mean(range)	p-value	Test Statistic	MGMT methylated Mean( range) (n = 23)	MGMT unmethylated Mean(range) (n = 9)	p-value	Test Statistic
Age	57.09(27–71)	43.23(23–71)	0.006**	t=320	44.2(30–71)	38.7(23–67)	0.011*	t=2.81	48.6(32–65)	48.6(23–67)	0.44	t=−0.83
Sex	M, F = 3, 8	M, F = 25, 18	0.181	r = 3.14	M, F = 13, 11	M, F = 13, 6	0.34	χ2 = 2.31	M, F = 1, 8	M, F = 13, 10	0.02*	χ2 = 7.39
FLAIRVOL	29052(1445–75116)	34192(0–218140)	0.63	t=−0.47	20774(0–218140)	27783(146–146030)	0.45	t=0.88	26997(1445–83239)	33727(3461–218140)	0.31	t=−0.48
CEVOL	2203(0–7077)	3090(0–27664)	0.51	t=−0.66	3420(0–27664)	2171(0–24497)	0.35	t=−1.11	2212(0–7077)	4033(0–27664)	0.17	t=−0.95
CEVOL / FLAIRVOL	0.17(0–1.15)	0.14(0–1.82)	0.79	t=0.26	0.22(0–2.05)	0.11(0–1.15)	0.38	t=0.98	0.19(0–1.15)	0.12(0–052)	0.31	t=0.49
Edge Contrast	544.7(72–1561)	1058.5(46–2741)	0.008**	t=−2.97	632(46–2741)	1369(72–1967)	0.005**	t=−328	643(72–1368)	1055(142–2741)	0.03*	t=−1.88
Homogeneity	1.9(02–4.5)	3.1(0.3–6.3)	0.013*	t=−2.75	2.4(0.2–5.1)	3.8(0.2–6.3)	0.002**	t=−3.46	2.5(0.8–4.5)	2.6(0.2–5.2)	0.44	t=−0.14
Pixel Correlation	6.8(0.9–11.3)	11.8(2.2–33.3)	0.001**	t=−3.65	12.3(2.2–28.5)	13.2(0.9–33.3)	0.17	t=−1.26	8.1(4.8–14.8)	9.8(0.9–18.3)	0.10	t=−1.26

^**p-value < 0.01

^*p-value < 0.05

IDH status:

IDH-wt status was detected in 11/54 (21%) of tumors. FLAIR borders in the IDH-wt group showed lower EC compared to the IDH-mt group (p = 0.008). Homogeneity of the FLAIR_VOL was higher in patients with IDH-mt tumors compared to patients with IDH-wt tumors (p = 0.013). Pixel correlations in the hyperintense FLAIR_VOL were also higher in the IDH-mt tumors compared to the IDH-wt (p = 0.001) group. There were no significant differences in the FLAIR_VOL, CE_VOL, or CE_VOL/FLAIR_VOL between these two groups.

1p/19q status:

The ratio of patients with IDH-mt-1p/19q-codeleted to IDH-mt-1p/19q-intact tumors was 24:19 (55%). Homogeneity and EC were lower in patients with IDH-mt-1p/19q-codeleted tumors compared to those with IDH-mt-1p/19q-intact tumors (p = 0.002 and p = 0.005, respectively).

MGMT status:

There were data on MGMT promoter methylation for 32/61 patients (52%). Twenty-three out of 32 (71%) patients with avaialbe MGMT promoter status were in the MGMT methylated group (see table 1). MGMT methylated tumors showed lower EC compared to MGMT unmethylated tumors (p = 0.03).

A pictorial summary of the univariate results for each texture feature by molecular subgroup is provided in table 3.

Table 3.

Relationship between imaging parameters and molecular information based on the univariate t-tests.

Molecular subgroups	IDH		1p/19q		MGMT
Imaging parameters	wildtype	mutant	Codeleted	intact	methylated	unmethylated
Edge Contrast	↓↓	↑↑	↓↓	↑↑	↓↓	↑↑
Homogeneity	↓↓	↑↑	↓↓	↑↑	-	-
Pixel Correlation	↓↓	↑↑	-	-	-	-

↓↓ significantly lower (p-value < 0.05), ↑↑ significantly higher (p-value < 0.05), - no significant changes (p-value > 0.1)

Logistic Regression Model

EC, homogeneity, and pixel correlation survived as predictors of IDH status among all the quantitative imaging features extracted using logistic regression with LASSO regularization (EC_coefficient = −6.36e-4, homogeneity_coefficient = −8.98e-05, pixel correlation_coefficient = −6.76e-02) (see figure 3). The stepwise AIC of the model for IDH status prediction was 51.7 using these three imaging parameters. When clinical/demographic data were also included in the model (i.e., age, sex), five features showed non-zero coefficients for predicting IDH status (age_coefficient = 4.62e-02, sex_coefficient = −1.02e-1, EC_coefficient = −3.74e-4, homogeneity_coefficient = −7.69e-05, and pixel correlation_coefficient = −4.14e-02). The stepwise AIC of this model for IDH status prediction was 44.0. The lower AIC of the second model indicates that the quality of the multivariate model for prediction of IDH status was slightly better when including age and sex with the imaging parameters compared to using only the imaging parameters. There were no non-zero coefficients in the logistic regression model using the imaging and clinical features to predict the status of 1p/19q and MGMT tumors.

Figure 3.

Depicts the features selection and regularization using LASSO on multiple logistic regression for predicting the IDH-status. The tuning parameter (lambda) has the most parsimonious model where its error is within one standard error of the minimum bionomial deviance. The optimum lambda occurs using three selected features (homogeneity, EC, and pixel correlation).

Discussion

In this study, we investigated whether three clinically validated molecular markers in patients with grade II/III gliomas—IDH, 1p19q, and MGMT status—have unique imaging phenotypes. Using novel quantitative imaging parameters of FLAIR homogeneity/heterogeneity and border patterns, we demonstrated that whereas these three groups do not differ in FLAIR or CE volumes, they differ across a range of FLAIR texture features that could help to determine genotype and facilitate treatment planning prior to surgery.

Higher signal homogeneity indicates a more uniform signal throughout the FLAIR region, whereas higher pixel correlation indicates more uniform texture at a local level (i.e., within neighboring regions of the tumor). Higher EC indicates more distinct and sharp borders surrounding the FLAIR region. Heterogeneous FLAIR signal may result from hemorrhage, necrosis, or calcification, whereas poorly defined FLAIR borders may reflect tumor infiltration from the tumor bed to adjacent tissue [31]. In this study, we apply these texture features to patients with grade II/III gliomas and demonstrate their utility in delineating the molecular subtypes described below.

IDH status:

Our results demonstrate that patients with IDH-wt tumors show greater global and local heterogeneity within the FLAIR hyperintense region in comparison to IDH-mt tumors. These findings are concordant with a previous study from Qi et al., which reported that IDH-mt tumors were more likely to show more homogeneous signal intensity compared to IDH-wt tumors [26]. In addition, we found that patients with IDH-wt tumors showed less distinct FLAIR borders compared with patients with IDH-wt tumors. These findings are supported by several other studies that have shown less distinct borders in IDH-wt tumors based on visual analysis [26,28]. Previous studies have shown the association between an “invasive” FLAIR border and tumor growth [46] in IDH-wt tumors. Interestingly, we found that a combination of imaging features including global and local FLAIR heterogeneity combined with FLAIR border patterns provides the best classification of tumors based on IDH status. As in previous studies, patients with IDH-wt tumors were older than IDH-mt group. Therefore, age also contributed to the model. These findings indicate that multiple feature of the FLAIR signal should be considered simultaneously when determining tumor genotypes—which may be easier to achieve with quantitative measures.

1p/19q codeletion:

We demonstrate that among patients with IDH-mt tumors, those with 1p/19q codeleted tumors show a more heterogeneous FLAIR signal and less distinct borders compared to the intact group. Consistent with our findings, previous studies have shown that a more heterogeneous intratumoural signal is present in patients with 1p/19q codeleted anaplastic tumors [31][32,33]. Similar findings were also reported when determining heterogeneity using a qualitative, visual analysis of tumor characteristics [32,33]. Our data show that patients with IDH-mt-1p/19q-intact tumors showed less heterogeneity in the FLAIR hyperintense region compared with patients who had IDH-mt-1p/19-codeleted tumors (see figure 2). Although the reason for the lower heterogeneity in the IDH-mt-1p/19q-intact tumors compared to the co-deleted tumors is unclear, calcification and paramagnetic susceptibility present in co-deleted tumors may increase heterogeneity in this subset of otherwise IDH-mt patients [33]. Although the goal of this paper was to characterize the imaging phenotypes associated with each individual marker, these data also speak to the value of stratifying patients by multiple molecular markers in larger patient cohorts.

Figure 2.

A) IDH-wildtype tumor demonstrating very indistinct borders and heterogeneous signal on FLAIR imaging, B) IDH-mutant-1p/19q codeleted tumor demonstrating heterogeneous signal on FLAIR imaging and C) IDH-mutant-1p/19q-intact tumor demonstrating homogeneous signal on FLAIR imaging.

MGMT prediction:

In this study, patients with MGMT methylated tumors showed less distinct borders compared to patients with MGMT unmethylated tumors. MGMT promoter methylation status predicts treatment response to alkylating agent chemotherapy, and may be used in conjunction with IDH status for guidance in clinical decision-making [18,34,47–49]. In terms of imaging characteristics, Drabycz et al. found associations between texture on FLAIR images and MGMT methylation status in patients with GBM [50]. We extend the literature by demonstrating differences in texture features in patients with grade II/III gliomas. Although it is unclear what advantage MGMT methylation status holds in patients with grade II/III gliomas, our findings of differences in EC between the methylated and unmethylated patients is of interest and could help to guide further hypotheses about the relevance of this marker in this group of patients.

With the rapid expansion of radiogenomics, there is increased interest in identifying non-invasive means of determining a patient’s tumor genotype. However, previous studies have primarily classified brain tumors based on visual analysis of imaging features, which may not provide a reliable, systematic characterization of texture features. The current work introduces quantitative imaging metrics to capture FLAIR signal homogeneity/heterogeneity and border distinctness. These measures may serve as adjuncts to standard clinical analysis and interpretation and provide information regarding the patient’s molecular status prior to biopsy/resection. These quantitative features may be particularly useful in large-scale clinical trials where measures must be quantifiable, reproducible across sites, and free of reader bias.

This study has some limitations that should be noted. First, our sample size is modest compared to many large-scale studies. Nevertheless, the fact that we demonstrate group differences as a function of genotype in a relatively small sample may speak to the robustness of our measures. Second, we did not have MGMT promoter methylation status for many of the patients, limiting the conculsions that we can draw about these data. However, we hope that are preliminary results with MGMT status will motivate future studies with robust sample sizes. As expected in patients with grade II/III gliomas, our groups were also unbalanced with respect to IDH status, with a lower number of IDH-wt tumors. However, our numbers are commensurate with those reported in the literature (5–15%) [51,52]. Third, our patients were not treatment-naïve, as many had undergone biopsy or subtotal resection. Therefore, our findings will need to be further validated in the pre-operative setting. Fourth, although several studies, including ours, have proposed novel imaging markers to classify molecular subtypes, an important extension of this work will be to derive reliable cut-offs for each imaging parameter that would be invariant to sample. However, a very large multi-site study would be needed to systematically address this. Finally, we did not include survival data in our study as most patients with grade II/III gliomas do not reach meaningful progression-free or overall survival endpoints over a short time frame (i.e., 7 years in our study). However, we intend to follow this patient cohort and report on the prognostic value of our imaging features overall and within subtypes to address specific questions related to risk statification (e.g., do patients with IDH-mt tumors with less distinct borders have poorer outcomes than patients with IDH-mt tumors with sharper borders?) in a longitudinal study. Although our current sample size was not large enough to subdivide patients within each molecular subtype by tumor grade, we plan to investigate the differences in imaging characteristics of patients with grade II versus III gliomas separately in a larger prospective study.

In conclusion, our study demonstrates the utility of novel, quantitative FLAIR texture features, specifically signal heterogeneity and border sharpness, that may serve as powerful imaging biomarkers for determining tumor molecular status in patients with grade II/III gliomas.