In vivo evaluation of EGFRvIII mutation in primary glioblastoma patients via complex multiparametric MRI signature

Abstract

Background

Epidermal growth factor receptor variant III (EGFRvIII) is a driver mutation and potential therapeutic target in glioblastoma. Non-invasive in vivo EGFRvIII determination, using clinically acquired multiparametric MRI sequences, could assist in assessing spatial heterogeneity related to EGFRvIII, currently not captured via single-specimen analyses. We hypothesize that integration of subtle, yet distinctive, quantitative imaging/radiomic patterns using machine learning may lead to non-invasively determining molecular characteristics, and particularly the EGFRvIII mutation.

Methods

We integrated diverse imaging features, including the tumor’s spatial distribution pattern, via support vector machines, to construct an imaging signature of EGFRvIII. This signature was evaluated in independent discovery (n = 75) and replication (n = 54) cohorts of de novo glioblastoma, and compared with the EGFRvIII status obtained through an assay based on next-generation sequencing.

Results

The cross-validated accuracy of the EGFRvIII signature in classifying the mutation status in individual patients of the independent discovery and replication cohorts was 85.3% (specificity = 86.3%, sensitivity = 83.3%, area under the curve [AUC] = 0.85) and 87% (specificity = 90%, sensitivity = 78.6%, AUC = 0.86), respectively. The signature was consistent with EGFRvIII+ tumors having increased neovascularization and cell density, as well as a distinctive spatial pattern involving relatively more frontal and parietal regions compared with EGFRvIII− tumors.

Conclusions

An imaging signature of EGFRvIII was found, revealing a complex, yet distinct macroscopic glioblastoma phenotype. By non-invasively capturing the tumor in its entirety, the proposed methodology can assist in evaluating the tumor’s spatial heterogeneity, hence overcoming common spatial sampling limitations of tissue-based analyses. This signature can preoperatively stratify patients for EGFRvIII-targeted therapies, and potentially monitor dynamic mutational changes during treatment.

Importance of the Study

Quantitative multivariate analysis of clinically acquired multiparametric MRI reveals a non-invasive in vivo EGFRvIII imaging signature in glioblastoma that can facilitate non-invasive patient selection for targeted therapy, stratification into clinical trials, and potential repeatable monitoring of dynamic molecular changes during treatment. The proposed approach differs from prior literature on the extensiveness of multiparametric MRI used, leading to a comprehensive radiomic signature, providing macroscopic biological insights into the structural and histological characteristics of the mutation-containing tumor, via properties related to cell density, water concentration, neovascularization, and potentially underlying apoptotic processes that lead to tissue necrosis. Moreover, this study discovered that the presence of EGFRvIII is strongly correlated to a specific spatial pattern/distribution of glioblastoma, suggesting that glioblastoma cerebral localization may be linked to the carried mutational status. Through the use of clinically acquired imaging and evaluating our model on independent cohorts, our findings can be considered of potential translational value, subject to multi-institutional validation

The understanding of the biology and genetics of glioblastoma, the most common malignant adult brain tumor, has improved significantly over the last 2 decades; consequently, World Health Organization (WHO) classification of glioblastoma based purely on histology has given way to an integrated histologically/molecularly based classification.¹ Different constellations of genetic alterations are associated with specific types of glioblastoma, and with differential sensitivity to certain therapies.¹

The most prevalent genetic alterations in glioblastoma, including point mutations and gene amplification, occur in the epidermal growth factor receptor (EGFR) gene.² The most common of these mutations results from an in-frame deletion that leads to a splice variant, EGFRvIII,³ which appears to enhance tumor proliferation, invasion, and angiogenesis, through multiple mechanisms, including elevated mitotic rate and reduction in apoptosis.³ Until now, the determination of EGFRvIII presence has required analysis of surgically resected or biopsy-obtained tissue specimens. Importantly, this type of assessment is limited by tissue sampling error, inability to capture the tumor’s spatial heterogeneity, and hesitation, due to surgical invasiveness, for longitudinal assessment during treatment and cancer progression.

The goal of this study is to non-invasively determine the EGFRvIII presence in patients with de novo (primary) glioblastoma, by multivariate analysis of preoperative multiparametric (mp)MRI data. We conducted this multivariate analysis through linear configuration of support vector machines (SVMs),⁴ to create a prediction index to determine patients harboring the EGFRvIII mutation. Analysis of mpMRI data via advanced computational analytics has been increasingly shown to provide rich and highly informative characterizations of glioblastoma and surrounding brain tissue,^5–13 as the various modalities reflect diverse and complementary tissue properties. We hypothesize that quantification of subtle, yet important and spatially complex imaging features extracted from mpMRI (ie, radiomic features) is informative when assessed multivariately and leads to non-invasively determining molecular tumor characteristics, herein the EGFRvIII oncogene, with sufficient sensitivity and specificity on an individual patient basis.

Materials and Methods

Study Setting and Data Source

We identified independent discovery (n = 75) and replication (n = 54) cohorts of preoperative baseline mpMRI scans, of 129 de novo (primary) glioblastoma patients (74 male, 55 female) with average age of 59.3 years (range: 18.7–86.9), acquired at the Hospital of the University of Pennsylvania (HUP) from 2008 to 2015. Data inclusion criteria comprised age over 18 years, histopathological diagnosis of glioblastoma (WHO grade IV), and preoperative baseline mpMRI at time of diagnosis that included pre- (T1) and post-contrast (T1-gadolinium [Gd]) T1-weighted, T2-weighted (T2), T2 fluid attenuated inversion recovery (T2-FLAIR), diffusion tensor imaging (DTI), and dynamic susceptibility contrast (DSC) MRI scans (see acquisition settings in Supplementary material). Consistent with existing literature, the EGFRvIII status (ie, EGFRvIII−/EGFRvIII+) for the discovery and the replication cohort was distributed as 51/24 and 40/14, respectively. For evaluating EGFRvIII presence, sufficient tumor tissue collected at time of surgery was required. Patients included in our study were treated according to standard of care, which included maximal safe resection, radiotherapy, and concomitant and adjuvant chemotherapy with temozolomide. All studies were approved by the HUP institutional review board.

Image Preprocessing

All mpMRI volumes were converted to Neuroimaging Informatics Technology Initiative (NIfTI) format, reoriented to the left-posterior-superior (LPS) coordinate system (as required by the computer-aided segmentation method¹⁴) and affinely co-registered (12 degrees of freedom) to the same anatomical template using the functional MRI of the Brain Software Library (FSL).¹⁵ Subsequently all volumes were smoothed¹⁶ to remove any high frequency intensity variations (ie, noise), corrected for magnetic field inhomogeneities¹⁷ and skull-stripped.¹⁸ We extracted commonly used measurements¹⁹ from the acquired DTI volumes, ie, the tensor’s apparent diffusion coefficient (ADC), axial diffusivity (AX), radial diffusivity (RAD), and fractional anisotropy (FA). Furthermore, the DSC-MRI volumes were used to extract parametric brain maps of the relative cerebral blood volume (rCBV), peak height (PH), and percentage signal recovery (PSR) after considering for leakage correction.^20,21 These DTI measurements and DSC maps were used as individual imaging volumes.

Determination of EGFRvIII Mutation Status

The most representative block per resected tissue specimen was chosen based on morphology and used for genetic analysis. An assay based on next-generation sequencing to detect EGFRvIII transcripts^22,23 was used, which was validated by Taqman reverse transcription PCR. Total nucleic acid was extracted from paraffin-embedded tissue, and complementary DNA was then synthesized from 200 ng RNA. PCR primers were designed to capture wild-type (wt-)EGFR, EGFRvIII, 3 housekeeping genes, and 3 primer sets (Supplementary Table S1), with increasing target sizes to assess the level of RNA degradation in the sample. The sequencing library preparation method was a 2-step PCR, with multiplex PCR followed by a second PCR to add Illumina sequencing index and adaptors. Subsequently, the sequencing library was quantified, sequenced on Illumina MiSeq, and analyzed using a bioinformatics pipeline developed in our lab. EGFRvIII presence was first evaluated by applying the following formula: EGFRvIII reads/(EGFRvIII reads + wt-EGFR reads), and then based on our results using normal brains and glioblastoma, a cutoff for EGFRvIII+ tumors was set at 5% EGFRvIII to wt-EGFR allele ratio.

Segmentation Labels of Tumor Sub-regions

A semi-automatic generative segmentation method (GLISTR¹⁴) was utilized to obtain the segmentation labels of the enhancing tumor (ET) and non-enhancing tumor (non-ET) core, peritumoral edema/invasion (ED), and ventricles. GLISTR simultaneously registers a probabilistic atlas of a healthy population to the patient’s mpMRI (ie, T1, T1-Gd, T2, T2-FLAIR) and partitions the patient’s brain into ET, non-ET, ED, ventricles, white matter (WM), gray matter, vessels, and cerebellum. It is based on expectation-maximization (EM) and accounts for the tumor’s diffusivity through a biophysical glioma growth model. This growth model modifies the healthy atlas into an image with a tumor and ED adapted to best match to the given set of patient mpMRI. The modified atlas is then registered into the patient images, and posterior probability estimates of tissue labels are iteratively refined based on EM, together with the deformation field, and the growth model parameters. All segmentation labels were evaluated by H.A. and M.B. and manually corrected when needed.

Radiographically, the ET and non-ET parts are defined by hyper-intense and hypo-intense areas, respectively, in T1-Gd compared with T1, but also compared with normal-appearing WM. The non-ET regions describe non-enhancing tumor, as well as transitional/pre-necrotic and necrotic regions that belong to the non-enhancing part of the bulk tumor, and are typically surgically resected together with the ET. Finally, the ED region is defined by the abnormal hyperintense signal on the T2-FLAIR volumes.

Quantitative Features Extracted from mpMRI

After segmenting the tumor sub-regions, relevant imaging features were extracted for each patient from all modalities, to capture phenotypic characteristics and create the proposed radiomic signature. These features describe (i) volumetric, shape, and size measurements, (ii) diffusion metrics related to tissue cellularity²⁴ and microstructure, (iii) physiological measures that relate to temporal perfusion dynamics of the tumor influenced by angiogenesis and breakdown of the blood‒brain barrier,^25,26 as well as (iv) tumor spatial distribution (Supplementary Table S2). All features were integrated via SVM to determine combinations most predictive of EGFRvIII and estimate the performance of our multivariate predictive model, via a 10-fold cross-validation scheme in the discovery cohort. This scheme assesses the discovery cohort as it was divided in independent training and testing cohorts, but in a more statistically robust manner by firstly randomly and proportionally dividing the cohort in training (90% of the data) and testing (10% of the data) subsets and then randomly permuting across them. Our model was also validated in the independent replication cohort.

An analytic estimation of statistical significance based on multivariate analysis (p_m) was employed to select and calculate the relative contribution of each feature in the final SVM classifier.²⁷ The classifier was trained using a linear kernel function. The parameter for the soft margin cost function (C) was optimized on the training data, based on a 5-fold cross-validated grid search; C = 2^α, where αϵ[−5,5]. This parameter controls the influence of each individual support vector that involves trading error penalty for stability.

Considering that the translational applicability of the multivariately selected features lies in enabling a clinician to directly distinguish the EGFRvIII status of patients without the need for computational analysis, we also conducted a univariate analysis of all the features in the combined population.

Evaluating Histograms of Features

To obtain a deeper biological understanding of the information used by the SVM to provide predictions, histograms of the most predictive features were generated by dividing the patient cohort into 2 groups according to the EGFRvIII status. Histograms of these features based on data from all patients were created with the value of the feature on the x-axis and its frequency on the y-axis (Fig. 1B). For certain features, values along the x-axis were partitioned to facilitate further analysis. These histograms were normalized according to the total number of voxels of each sub-region. This approach focused on capturing the heterogeneity of the radiomic features within the obtained mask for each tumor sub-region.

Fig. 1.

(A) Two examples of the imaging modalities and the segmentations in EGFRvIII+ and EGFRvIII− patients. (B) Illustrated histograms of the most distinctive features, according to EGFRvIII status. The histograms in the first column are divided (by the dashed lines) into 3 partitions, to allow for deeper analysis. In the histogram of rCBV in ET, the first partition shows higher population for EGFRvIII+, likely representing pre-necrotic areas (ie, with very low rCBV). For this partition, ADC suggests higher cellularity in EGFRvIII+. The remainder of the rCBV in ET histogram suggests that EGFRvIII+ is hypervascular, with the second partition suggesting lower water concentration for the EGFRvIII+ tumors through the T2-FLAIR signal. In the histogram of T2-FLAIR for non-ET, the first partition might represent more CSF-like fluid in EGFRvIII− and the third partition represents higher water concentration for EGFRvIII−. ADC and FA measures, for the third partition, suggest higher cellularity and different microstructure, respectively, in EGFRvIII+ tumors. Note that the histograms were created using information from all patients, whereas a single patient example is shown at the bottom of the figure for visualization purposes.

Spatial Distribution and Pattern of the Tumor

To capture the tumor spatial configuration,²⁸ we constructed 2 spatial distribution atlases (SDAs), one for each EGFRvIII status, ie, SDA+ and SDA− for EGFRvIII+ and EGFRvIII−, respectively. These SDAs describe the spatial frequency of tumor occurrence at each voxel, calculated by superimposing the tumor core (TC) segmentation labels (ie, the union of ET and non-ET) for all patients based on their EGFRvIII status (Fig. 2). These atlases are then used to calculate the similarity of a new unseen TC with each of the 2 atlases. To assess a new brain scan b, we first need to segment its TC (ie, TC_b), and then retrieve the voxel-wise frequency of each SDA for the area defined by TC_b, ie, SDA+(TC_b) and SDA−(TC_b). Four discrete relative values are then used in our analysis, based on the maximum and average frequency of each SDA for the TC_b area: (i) max(SDA+(TC_b))/max(SDA−(TC_b)), (ii) mean(SDA+(TC_b))/mean(SDA−(TC_b)), (iii) max(SDA+(TC_b))–max(SDA−(TC_b)), (iv) mean(SDA+(TC_b))–mean(SDA−(TC_b)). Moreover, the proportions of TC in each lobe, and the distance from its sub-regions to ventricles have been also utilized.

Fig. 2.

Spatial configuration of primary glioblastomas according to their EGFRvIII status. (A) EGFRvIII− on top panel, (B) EGFRvIII+ middle lower panel, and (C) frequency distribution on lower panel. The color look-up tables show the frequencies in percent. All images were displayed in the radiological convention orientation. The basal ganglia label consists of putamen, caudate nucleus, globus pallidus, subthalamic nucleus, nucleus accumbens, internal capsule, and thalamus.

To analyze the relationship between the status of the EGFRvIII mutants and the TC location on the voxel level, we have also used the existing robust approach of voxel-based lesion–symptom mapping (VLSM)²⁹ over 1000 permutations. VLSM maps, in our study, depict t-test results evaluating EGFRvIII per voxel, where patients with an EGFRvIII+ tumor in a given voxel were compared with those with EGFRvIII− status.

Peritumoral Heterogeneity

We also note a recent hypothesis-driven discovery¹³ of a peritumoral heterogeneity index (PHI) being highly informative of EGFRvIII status, based on analysis of the temporal perfusion dynamics (DSC-MRI) alone. Thus, we incorporated this PHI into our multivariate predictive model, resulting in a meta-classifier that we trained in the discovery cohort and evaluated on replication cohort similarly to the original model.

Code Availability

All software tools used for this study are publicly available. Specifically, the tools used for skull-stripping and co-registration are available in fsl.fmrib.ox.ac.uk. The Cancer Imaging Phenomics Toolkit (CaPTk—www.med.upenn.edu/cbica/captk)³⁰ was used to smooth the brain volumes, to manually seed points for initializing GLISTR, and to extract the radiomic features. GLISTR is available through www.med.upenn.edu/sbia/glistr.html. Finally, we conducted the SVM analysis using the open-source library for SVM⁴ (LIBSVM—www.csie.ntu.edu.tw/~cjlin/libsvm).

Results

Classification Performance of Predictive Model

Our predictive model’s accuracy in correctly classifying the EGFRvIII status on the discovery and replication cohorts was 85.3% (specificity = 86.3%, sensitivity = 83.3%) and 87% (specificity = 90%, sensitivity = 78.6%), respectively. Assessment of the mutation status and feature selection in the discovery cohort was conducted by 10-fold cross-validation, whereas the performance in the replication cohort was estimated after independently applying the model trained in the discovery cohort. Furthermore, a receiver operating characteristic (ROC) analysis was performed and the area under the curve (AUC) for each ROC was 0.85 and 0.86, for the discovery and the replication cohort, respectively (Fig. 3). The optimal accuracy, specificity, and sensitivity of the meta-classifier including PHI¹³ in the replication cohort were 88.9%, 90.0%, and 85.7%, respectively (AUC = 0.92).

Fig. 3.

ROC curves in discovery (left panel) and replication (right panel) cohorts. Red points display the optimal cutoff points, revealing accuracies of 85.3% (AUC = 0.85, specificity = 86.3%, sensitivity = 83.3%) and 87% (AUC = 0.86, specificity = 90%, sensitivity = 78.6%) for discovery and replication cohorts, respectively.

Selected Quantitative Features

The selected features, based on our multivariate model, are spatial tumor pattern and intensity distribution in the non-ET sub-region. Specifically, the selected features and their statistical significance²⁷ were:

• i) max(SDA+(TC_b))–max(SDA⁻(TC_b)), (p_m = 0.00026),

• ii) max(SDA+(TC_b))/max(SDA⁻(TC_b)), (p_m = 0.00028),

• iii) the 4th bin of the non-ET intensity distribution in the ADC maps, (p_m = 0.0084),

• iv) the 3rd bin of the non-ET intensity distribution in the ADC maps, (p_m = 0.029), and

• v) the standard deviation of the non-ET intensity in the T2 modality, (p_m = 0.06).

Individual thresholds for each of the features accompanied by their statistical significance based on a univariate analysis in the combined population are reported in Supplementary Table S3. These thresholds can be used to associate with EGFRvIII status in a non-computational assessment.

Biological Interpretation of the Radiomic Signature

The main findings of the obtained radiomic signature indicate that the EGFRvIII+ tumors (compared with the EGFRvIII− tumors) have signals of higher rCBV, lower ADC, higher FA, lower T2-FLAIR, and a distinctive spatial pattern (Fig. 4). A voxel-based analysis, via the Wilcoxon signed-rank test, of all sub-regions’ intensities in all employed modalities revealed significant difference between EGFRvIII− and EGFRvIII+ tumors, with the highest P-value equal to 10⁻¹¹.

Fig. 4.

Summary of descriptive characteristics of EGFRvIII+ glioblastoma.

A quantitative analysis of the tumors’ spatial distribution shows a differential localization depending on the mutation’s presence, with EGFRvIII+ tumors being more frequent in frontal and parietal lobes, and EGFRvIII− tumors more frequently appearing in the temporal lobe (Fig. 2). ROC analysis of only the tumor’s spatial configuration yielded the AUC in the discovery and the combined cohort equal to 0.72 (P = 5 × 10⁻⁴) and 0.71 (P = 5 × 10⁻⁵), respectively.

VLSM also confirmed the contrast between left perisylvian regions and frontal lobe. Fig. 5 shows representative slices from VLSM maps computed for EGFRvIII status of 129 GBM patients. These maps are t-test results evaluating EGFRvIII status on the voxel level. High t-scores (red) indicate that tumors to these voxels have a highly significant correlation to the EGFRvIII status. Only voxels that were significant at P = 0.05 are shown. The power map is calculated for alpha = 0.05.

Fig. 5.

Representative slices from VLSM maps computed for EGFRvIII mutation status of 129 glioblastoma patients. The top panel shows the voxel-based map of power across the brain, calculated for alpha = 0.05. The bottom maps are illustrations of t-test results. Patients with tumor in a given voxel were compared with those without tumor in that voxel on measures of EGFRvIII status. High t-scores (red) indicate that tumors to these voxels have a highly significant effect on EGFRvIII mutation status. Dark blue voxels indicate regions where the presence of a tumor had relatively lower impact on the EGFRvIII mutation status. T map includes only significant voxels (P = 0.05), considering thresholding based on cluster size and 1000 permutations.

Discussion

Summary of Findings and Results

This study focused on investigating the connection between the radiolographic characteristics found in EGFRvIII-mutant glioblastoma, as well as on providing a macroscopic biological insight for these changes. We identified an in vivo sensitive and specific radiomic signature of the EGFRvIII mutation in glioblastoma, and investigated biological correlates of the most distinctive imaging features. Importantly, this signature was derived using a comprehensive panel of morphological and physiological features extracted from routine clinically acquired mpMRI, without the need to develop and deliver specialized molecular ligands. The main findings of the obtained radiomic signature indicate that the EGFRvIII+ tumors (compared with the EGFRvIII− tumors) have signals of higher rCBV, lower ADC, higher FA, lower T2-FLAIR, and more variable spatial pattern. These quantitative imaging features have been associated with neovascularization, cell density, and tissue microstructure, inflammation, necrosis, and other processes related to this aggressive cancer. Importantly, individual assessment of these features is not sufficient for identifying the mutation on an individual patient basis; however, appropriate integration yielded sufficient sensitivity and specificity for identifying the mutation on individual patients, underlining the value of multivariate pattern analysis approaches (Fig. 1). An important finding is that the spatial tumor distribution was the most distinctive feature of this mutation, therefore emphasizing the value of assessing spatial characteristics of tumors via imaging and mapping to reference atlases. Another study^9,10 has also reported varying distributions between EGFRvIII+ and EGFRvIII− tumors that differ from the spatial location reported here. However, the 3 times larger sample size included here potentially enables this study to obtain more robust statistics of spatial patterns. Furthermore, all patient brain scans in the other studies were aligned to a common reference space using simple affine registration, whereas our study utilizes deformable registration that explicitly accounts for the large deformations that these tumors induce to the surrounding brain (known as mass effect), through a biophysical tumor growth model.¹⁴

Clinical Relevance and Impact

Evaluation of cancer mutations is currently typically done via tissue specimen analyses obtained from a single location within the tumor. This can be limiting, as a single specimen/sample might not reflect the presence of the mutation in such a heterogeneous tumor, which is known as sampling error. Moreover, repeated evaluation during treatment is typically not considered, due to invasiveness, thereby limiting our ability to measure the temporal heterogeneity.

The radiomic signature we identified offers promise in addressing both of these limitations, since mpMRI allows repeatable assessment/monitoring of the tumor’s entirety. Importantly, our results indicated that heterogeneity in both imaging signals and spatial tumor distribution correlates with the mutation’s presence. Although molecular imaging ligands under development for PET^31,32 and other molecular imaging modalities aim to more directly evaluate the presence of mutations, these methods are limited by the need to (i) develop specialized ligands for each target mutation and (ii) have multiple scans for evaluating multiple mutations. Our radiomic signature is derived from routine clinically acquired mpMRI, and therefore is easily translatable to the clinic. Although this study is focused on EGFRvIII, the same approach could also be used for other mutations, by deriving respective signatures from our extensive quantitative imaging feature panel³⁰ integrated appropriately for each mutation of interest.

The radiomic signature for EGFRvIII, identified in baseline preoperative scans of glioblastoma patients, has potentially important clinical implications. First, it can help identify cases more likely suffering from tissue sampling error, since these cases might have a positive imaging signature and negative tissue results. Second, the radiomic signature can identify patients with high degrees of heterogeneity in the key imaging features that could potentially indicate patients with a heterogeneous mix of EGFRvIII+ and EGFRvIII− cells. This has been observed in our next-generation sequencing data, where multiple blocks of the same specimen were sequenced,³³ although it would need to be confirmed against imaging features to establish the relationship between these 2 types of heterogeneity. Finally, the non-invasive nature of this signature allows its likely application in cases of recurrent glioblastoma (subject to future evaluation), with the goal of assessing EGFRvIII status before and during treatment, hence non-invasively monitoring dynamic mutational changes as response to targeted therapy and therefore allowing for adjustments in the provided therapy. Importantly, it has been shown that EGFRvIII expression may be lost at the time of progression in about half of patients treated by standard chemoradiation,³⁴ so non-invasively assessing the presence of the mutation before and during treatment could be critical. There is also a high probability of losing EGFRvIII expression after vaccination with rindopepimut^35,36 or infusion of genetically modified chimeric antigen receptor T cells.³³

The radiomic signature we identified may also be used to gain insights into the biological basis of target antigen loss by longitudinal evaluation of EGFRvIII expression. Importantly, it is also possible that assessment of clinical outcomes for EGFRvIII targeted trials would be enhanced by a non-invasive correlate of EGFRvIII expression, to depict spatial heterogeneity and to potentially define maintenance of EGFRvIII oncoprotein in tumors after resection.

Radiographic Changes Consistent with Molecular Pathways

Detailed evaluation of the features contributing to the radiomic signature of EGFRvIII provided a macroscopic biological interpretation that is consistent with current literature on the molecular pathways and the pathological changes associated with the EGFRvIII oncogene (Fig. 6). Specifically, consistent with our finding of increased cellularity/proliferation, EGFRvIII is known to (i) be highly associated with the Shc-Grb2-Ras pathway, which enhances DNA synthesis³; (ii) activate the phosphatidylinositol-3 kinase pathway,^3,37 which increases levels of phosphorylated Akt; and (iii) reduce levels of p27KIP1, which regulates mitosis.^3,37,38 Enhancement of glial cell proliferation and invasiveness is also regulated by inducible nitric oxide synthase signaling downstream of EGFRvIII/signal transducer and activator of transcription 3, which also elevates the population of brain tumor stem cells.^39,40 Consistent with our finding of increased cellularity, it is also reported that EGFRvIII may have a role in promoting cancer neural stem cell renewal, through abnormal spindle-like microcephaly associated protein.^39,41 The proliferative effects of EGFRvIII are strengthened by its anti-apoptotic effects through upregulation of B-cell lymphoma–extra large.⁴² This behavior is consistent with the relatively high T2-weighted signals, observed in parts of EGFRvIII− tumors, indicative of higher water concentration often related to higher rate of cell death in pre-necrotic tissue.⁴³ The enhanced aggressiveness of EGFRvIII expressing glioma cells is also associated with enhanced tumor angiogenesis,³ which we also found to be a key characteristic of EGFRvIII+ tumors, in terms of both rCBV and PH suggesting hypervascularization.

Fig. 6.

EGFRvIII associated molecular pathways, biological changes, and expected radiological changes. Note that multiple biological changes within the tumor have one or more imaging correlates on different MRI sequences.

Biological Interpretation of Features of the EGFRvIII Signature

To gain understanding about the biological processes that could give rise to the EGFRvIII radiomic signature, we performed a detailed histogram analysis of all imaging features included in all tumor sub-regions (ie, ET, non-ET, ED). The main findings from comparing the features of EGFRvIII+ with those of EGFRvIII− tumors, in both ET and non-ET (Fig. 1), identify EGFRvIII+ tumors as having:

1. Regions of higher blood volume, which points toward hypervascularity and increased angiogenesis, based on rCBV and PH;

2. Regions of imaging characteristics consistent with higher cellularity, suggestive of increased proliferation, as well as different tissue microarchitecture, based on ADC and FA;

3. Regions of lower water concentration, based on T2-FLAIR and ADC, consistent with dense and non-necrotic tissue;

4. A distinctive part of the tumor that displayed pre-necrotic imaging characteristics, based on very low rCBV; we named it pre-necrotic, as its phenotype would be consistent with tissue transitional to necrotic;

5. A spatial pattern of the tumor’s extent that was markedly, and statistically significantly, more variable than, and different from, that of EGFRvIII− tumors, potentially consistent with either more variable origin or with higher migratory behavior.

The 2 most descriptive parameters of our analysis are shown on the left side of Fig. 1B. These histograms were divided in 3 partitions based on where the 2 curves are crossed. This partitioning leads to deeper interpretation and reveals the importance and benefit gained through multivariate analysis of mpMRI data, which is not perceivable by visual inspection alone. The top-left histogram in Fig. 1B describes the distribution of rCBV in ET, where the first partition shows a higher population for EGFRvIII+ tumors and more likely represents the pre-necrotic areas that have very low perfusion (rCBV). To further evaluate these areas we assessed the ADC signals, which suggest higher cellularity for EGFRvIII+ tumors. The second and third partition of rCBV in ET suggests that EGFRvIII+ tumors have higher degree of vascularization. Specifically, for the second partition we also assessed the T2-FLAIR signal, which suggested lower water concentration for the EGFRvIII+ tumors. We also assessed the non-ET and noticed that the most descriptive modality was T2-FLAIR, where the first partition might represent more CSF-like fluid in EGFRvIII− and the third partition represents higher water concentration for EGFRvIII− tumors. Further evaluation of the latter partition, through ADC and FA measures, shows evidence of areas with higher cellularity and different microstructure, respectively, in EGFRvIII+ tumors.

Spatial Distribution and Pattern of the Tumor

An important finding of this study is that the TC spatial location is the most distinctive feature in estimating this mutational status. In particular, tumors harboring the mutation had a spatial distribution overlapping mainly frontal and parietal lobes, whereas EGFRvIII− tumors were found predominantly in the temporal lobe (Fig. 2).²⁸ Differences in location of origin, or histological differences in the stem cell–rich subventricular zone among these lobes can potentially explain these spatial pattern differences.⁴⁴ Other potential explanations might include the higher motility and infiltration ability that EGFRvIII+ tumors are thought to possess, which might cause their migration away from the common perisylvian pattern. Finally, unknown factors related to the tumor microenvironment might explain these marked differences in spatial location.

Limitations

A limitation of this study is that the data were acquired from a single institution, whereas multi-institutional data would be beneficial to further and externally validate our predictive model and imaging features. However, the use of independent discovery and replication cohorts, along with the use of clinical mpMRI protocols, provides confidence that this signature will generalize to other institutions and patient populations.

Furthermore, this study evaluated the expression of EGFRvIII, as a global binary present/absent value, which is what is typically available in standard tissue analysis. However, the identified features can be evaluated on a regional basis to capture the tumor’s spatial heterogeneity, subject to availability of multiple spatially distinct biopsies or single-cell data. Distinct histopathologic and molecular profiles have been noticed in EGFRvIII tumors,³³ and we expect these regional molecular differences to be potentially validated by our radiomic signature.

Importance of Multivariate Analysis of mpMRI

T1 and T1-Gd images contain information indicative of regional angiogenesis and of the blood‒brain barrier integrity in the tumor and surrounding regions. T2 and T2-FLAIR images are helpful for assessing both the non-ET and the extent of the ED.⁴⁵ DTI maps the diffusion process of water in the brain, affected in part by tumor cellularity⁴⁶ and by integrity of WM structures, as well as the underlying microstructure of tissue, eg, via FA. DSC techniques reflect various aspects of perfusion in the brain,⁴⁷ which provide quantitative measures of regional microvasculature and hemodynamics.^6,48,49 The histogram analysis of these measures reflects heterogeneity in the tumor and its surrounding ED. Individual features were too subtle, hence insufficient, to produce a specific enough radiomic signature. Conversely, an optimally synthesized combination of many complementary imaging features (Supplementary Table S2) achieved high accuracy, both in cross-validation and in independent cohort configurations. Furthermore, comparing the univariate assessment of the extracted features with the multivariate analytic estimation of statistical significance of each feature²⁷ reveals that subtle individual signal changes can be synthesized in an index of higher distinctive performance. Kickingereder et al also evaluated the association of mpMRI features with molecular characteristics, but without assessing EGFRvIII.⁵

Previous attempts to identify EGFRvIII from imaging were focused on analyzing data from a single MRI modality^13,49,50 or on blindly correlating to the mutant status only in the bulk of the tumor,^49,50 instead of integratively assessing mpMRI signals from the tumor entirety (including the peritumoral infiltrated tissue)¹³ and providing an approach toward understanding how these imaging signatures correlate to tumor biological processes.

Supplementary Material

Supplementary material is available at Neuro-Oncology online.

Funding

This work was supported by NINDS/NIH:R01NS042645, NCI/NIH: U24CA189523, NCATS/NIH: UL1TR001878, and the ITMAT of the University of Pennsylvania.

Conflict of interest statement

Nothing to disclose.