Diffusion MRI Phenotypes Predict Overall Survival Benefit from Anti-VEGF Monotherapy in Recurrent Glioblastoma: Converging Evidence from Phase II Trials

Glioblastoma (GBM) is the most common and lethal type of primary brain tumor in adults. Median survival for patients diagnosed with a GBM is around 14 months (1), and fewer than 10% of patients survive beyond 5 years after diagnosis (2). The current standard of care for newly diagnosed GBM consists of maximum safe surgical resection, followed by radiotherapy plus concomitant and adjuvant temozolomide. At recurrence, however, few effective therapeutic options exist, and median survival is only around 8 to 10 months (3–5).

High tumor vascularity resulting from elevated production of proangiogenic growth factors, including vascular epithelial growth factor (VEGF; refs. 6, 7), has led to the development of therapies targeting proangiogenic signaling pathways (8). Antiangiogenic therapies that specifically target VEGF and its receptors fall into two general categories: antibodies or tyrosine kinase inhibitors (TKIs; ref. 9). Bevacizumab, a humanized monoclonal antibody for VEGF-A (Fig. 1), was approved for use in recurrent GBM in 2009 after it was shown to improve progression-free survival (5, 10). Bevacizumab is often used for the treatment of recurrent GBM and almost all patients will receive bevacizumab at some point during their treatment. Aflibercept, or “VEGF trap,” is a recombinant fusion protein that has been tested in recurrent GBM (11) and similarly acts in the extracellular domain by trapping VEGF through use of VEGF-binding portions from the extracellular domains of VEGFR-1 and VEGFR-2. Cediranib and cabozantinib, which are both multitargeted pan-VEGF receptor TKIs (12–17), have also been studied as promising antiangiogenic agents that inhibit different aspects of VEGF signaling, specifically the intracellular kinase domain of the VEGF receptors (Fig. 1).

Despite promising initial data and widespread exploration of anti-VEGF therapies in recurrent GBM, randomized phase II trials have not demonstrated an overall survival (OS) benefit for unselected patients with recurrent GBM. However, recent evidence suggests certain subsets of patients with specific imaging characteristics may have a significant OS benefit from anti-VEGF therapy. For example, single and multicenter studies have shown that high pretreatment apparent diffusion coefficient (ADC) measured with diffusion MRI within the contrast enhancing portion of the tumor is a significant and independent predictor of prolonged OS in patients treated with bevacizumab (18–21). Because bevacizumab, cediranib, cabozantinib, and aflibercept are all VEGF pathway inhibitors, and all commonly have activity against VEGFR-2, we hypothesized baseline diffusion MRI characteristics in recurrent GBM may be predictive of survival benefit for all four anti-VEGF therapies. Thus, the purpose of this study was to first evaluate whether pretreatment diffusion MR measurements of ADC within the contrast enhancing tumor is an independent, significant predictor of OS benefit in recurrent GBM patients treated with anti-VEGF monotherapy by examining data in a series of multicenter, phase II trials in cediranib monotherapy (NCT00035656), bevacizumab monotherapy (BRAIN Trial, AVF3708g; NCT00345163), cabozantinib monotherapy (XL184-201; NCT00704288), and aflibercept monotherapy (VEGF trap, NABTC0601; NCT00369590). Lastly, we validated our findings in an independent cohort of recurrent GBM patients treated with bevacizumab monotherapy versus lomustine monotherapy (BELOB), confirming that diffusion MR phenotypes are exclusively predictive for anti-VEGF therapy.

A total of 258 anti-VEGF treatment–naïve recurrent GBM patients with measurable enhancing tumor (>1 cm³) and adequate diffusion MRI estimates of ADC available from 5 separate phase II clinical trials were included in this study. Data from four trials were used as training cohorts: (i) 30 patients treated with cediranib monotherapy (NCT00035656), including two pretreatment (double baseline) MR examinations; (ii) 57 patients treated with bevacizumab monotherapy (BRAIN Trial, AVF3708g; NCT00345163); (iii) 65 patients treated with cabozantinib monotherapy (XL184-201; NCT00704288); (iv) 25 patients treated with aflibercept monotherapy (VEGF trap; NABTC0601; NCT00369590); and one trial was used for validation: (v) 42 patients treated with bevacizumab monotherapy and 39 patients treated with lomustine monotherapy as part of a phase II validation cohort (BELOB).

Disease progression that led to enrollment in each study was identified on MRI ≤ 14 days before the baseline treatment. All patients on all trials failed initial standard of care including concurrent radiotherapy and temozolomide and were required to be at least 8 weeks from completion of radiation therapy before considered eligible for each trial. Karnofsky performance status (KPS) was required to be ≥60 for all trials. Patients receiving corticosteroids were required to be on stable or decreasing dose for at least 5 days prior to the baseline scan. All patients enrolled in all trials signed institutional review board-approved written consent at the respective study sites. Table 1 further outlines patient demographic information for each trial included in the current study.

A total of 30 patients of the 31 enrolled in an NCI-sponsored phase II trial of cediranib, a pan-VEGF receptor TKI with additional activity against cKIT and PDGFβ, at the Massachusetts General Hospital and Dana-Farber Cancer Institute (NCI; NCT00035656) were included in the current study based on availability of high-quality diffusion MR and anatomic data. One patient was excluded based on limited drug exposure, as described previously (22). Patients were treated with cediranib (AZD2171; AstraZeneca Pharmaceuticals) at 45 mg once daily on a 28-day cycle. For more information on the trial, therapeutic dosing information, and details on inclusion and exclusion criteria, see the following website for the trial: clinicaltrials.gov/ct2/show/NCT00035656 and final published clinical trial information from Batchelor and colleagues (22).

A total of 57 recurrent, histologically confirmed GBM patients within the bevacizumab monotherapy arm in the BRAIN trial (La Roche/Genentech, AVF3708g; NCT00345163), an open-label, multicenter (11 sites), randomized, noncomparative phase II trial performed to assess the effectiveness of bevacizumab or bevacizumab and irinotecan hydrochloride with or without concomitant enzyme-inducing antiepileptic drugs were included in the current study based on availability of both high-quality diffusion and anatomic MRI at baseline. Patients within the bevacizumab monotherapy arm were treated with a dose of 10 mg/kg every 2 weeks. Patients within the bevacizumab and irinotecan combination arm were not included in the current study. For more information on the trial, therapeutic dosing information and details on inclusion and exclusion criteria, see the following website for the trial: clinicaltrials.gov/ct2/show/NCT00345163 and final published clinical trial information from Friedman and colleagues (5).(It is important to note that a previous study similar to the current study was already published using this dataset, ref. 20; but all data have been independently reanalyzed using slightly different methodology in the current study.)

A total of 65 anti-VEGF treatment–naïve, recurrent GBM patients of the 152 originally enrolled as part of XL184-201, a multicenter (8 sites), phase II, open-label, uncontrolled study of cabozantinib (XL184; Exelixis; NCT00704288), a tyrosine kinase inhibitor with principal targets of MET, VEGF receptors, AXL, and RET, at a dose of 140 or 100 mg (free base equivalent weight, oral, daily) at first or second relapse were included in the current study based on the availability of diffusion and anatomic imaging data. Specific inclusion and exclusion criteria for this trial can be found at clinicaltrials.gov/ct2/show/NCT00704288.

A total of 25 patients with histologically confirmed GBM with evidence of unequivocal progression after chemoradiation enrolled in an NABTC- and NCI-sponsored multicenter (7 sites), phase II study of aflibercept (VEGF trap; NABTC0601; NCT00369590) were included in the current study based on the availability of diffusion and anatomic imaging data. Patients were treated with a dose of 4 mg/kg intravenously every 2 weeks. For more information on the trial, therapeutic dosing information, and details on inclusion and exclusion criteria, see the following website for the trial: clinicaltrials.gov/ct2/show/NCT00369590 and final published clinical trial information from de Groot and colleagues (11).

To confirm that diffusion MR phenotypes are predictive of response to anti-VEGF therapies but not cytotoxic chemotherapies, we examined a total of 81 recurrent GBM patients treated with either bevacizumab monotherapy (N = 42; 10 mg/kg every 2 weeks) or lomustine (N = 39; 110 mg/m² every 6 weeks) as part of a multicenter (14 sites), Dutch investigator initiated phase II trial exploring bevacizumab, lomustine, or bevacizumab plus lomustine in patients with GBM at first recurrence (BELOB; NTR1929). Patients treated with combination bevacizumab and lomustine were not included in the current study. For more information on the trial, therapeutic dosing information and details on inclusion and exclusion criteria, see the final published clinical trial information from Taal and colleagues (23).

Standard and diffusion MR data were acquired using either a 1.5T or a 3T MR scanner from an MR scanner manufactured by Siemens Healthcare, Philips Medical Systems, GE Medical Systems, or Hitachi Medical Corporation. The MRI parameters for cediranib (24–26), bevacizumab (20, 27), and cabozantinib (28) trials have been published previously. T1-weighted images were acquired using either a fast spin-echo sequence or a magnetization-prepared rapid acquisition gradient-echo sequence with repetition time (TR)/echo time (TE)/inversion time (TI) = 400–3209 ms/3.6–21.9 ms/0–1.2 seconds; slice thickness = 3–6.5 mm; slice gap = 0–2.5 mm; number of excitations/averages (NEX) = 1–2; matrix size = 176–512 × 256–512; and field of view (FOV) = 240–256 mm. Diffusion-weighted MR images (DWI) were acquired before injection of contrast with TE/TR = 80–110 ms/4–10 seconds, NEX = 1, slice thickness = 5 with 0–1 mm interslice gap, matrix size = 128 × 128, and FOV = 220–256 mm using a monopolar spin-echo echo-planar preparation. ADC maps were calculated offline from the acquired DWIs using b = 0 s/mm² and b = 1,000 s/mm² images. Post-contrast T1-weighted MR images were acquired with identical acquisition parameters shortly after contrast injection with gadopentetate dimeglumine (Magnevist; Berlex), at a concentration of 0.1 mmol/kg, or a contrast agent with equivalent relaxivity. Additional anatomical T2-weighted fast spin-echo and T2-weighted fluid-attenuated inversion-recovery (FLAIR) sequences were also acquired prior to contrast, but not used in the current study. All examinations were acquired within 14 days of starting therapy for all three trials, and 29 of the 30 patients used in the cediranib trial received two examinations, approximately 1 to 7 days apart, prior to treatment initiation for the purposes of reproducibility testing.

Contrast-enhanced T1-weighted digital subtraction maps were created by registration, normalization, and subtraction of pre-contrast from post-contrast T1-weighted images as outlined previously (29). First, affine registration was performed between pre- and post-contrast T1-weighted images by using a 12 degree-of-freedom transformation and a correlation coefficient cost function in FSL (FLIRT; FMRIB Software Library, Oxford, England; http://www.fmrib.ox.ac.uk/fsl/). Second, the image intensities for both pre- and post-contrast T1-weighted images were intensity normalized using custom c-code and bash scripts courtesy of the National Institutes of Mental Health Magnetoencephalography 3Core Facility (3dNormalize; NIMH MEG Core, Bethesda, MD; kurage.nimh.nih.gov/meglab/Med/3dNormalize). This normalization was performed by essentially dividing each voxel by the standard deviation of the image intensity from the whole brain [S_Nor(x, y, z) = S (x, y, z)/σ_WB], where S is raw image signal intensity, Nor is normalized, x, y, z are voxel coordinates, and WB stand for whole brain. Third, voxel-wise subtraction was performed between intensity normalized pre-contrast and post-contrast T1-weighted images (Fig. 2A). Lastly, voxels with T1 subtraction values greater than zero were isolated, then manual corrections were made to exclude vessels or erroneous voxels, resulting in final T1 subtraction maps used to extract tumor volumes and act as continuous volumes of interest (VOI) for ADC histogram analysis (Fig. 2C), described below. Any satellite enhancing lesions were pooled together into a single VOI for subsequent analysis.

T1 subtraction-defined enhancing tumor volumes were used to extract ADC values for ADC histogram analysis. Nonlinear regression of a double Gaussian mixed model was then performed for the extracted ADC histograms using GraphPad Prism, Version 4.0c (GraphPad Software). The model used for the double Gaussian was defined by the following equation:

where p(ADC) is the probability of obtaining a particular value of ADC in the histogram, f is the relative proportion of voxels represented by the lower histogram, N(μ, σ) represents a normal (Gaussian) distribution with mean, μ, and standard deviation, σ, ADC_L represents the lower and ADC_H represents the higher of the two mixed Gaussian distributions (Fig. 2D). Resulting model fits were visually inspected and rerun with different initial conditions until adequate convergence was obtained. Goodness of fit was determined to be adequate if the adjusted R² > 0.7. This approach is similar to those used previously (18–20, 30).

Reproducibility in ADC_L measurements within the enhancing tumor was evaluated by calculating the coefficient of variance (COV) of pretreatment, double baseline MR examinations acquired as part of the cediranib trial. Also, a paired t test was used to determine whether a significant difference in ADC_L was observed between these two baseline time points. To determine whether this variability results in meaningful changes in the classification of ADC_L phenotypes for individual patients, the proportion of patients with the same ADC_L categorization (e.g., high vs. low ADC_L) were calculated as a function of different ADC_L thresholds.

Next, optimal ADC_L thresholds were determined by calculating the Mantel–Haenszel hazard ratio and corresponding P values for patients categorized as “high” vs. “low” ADC_L by changing different ADC_L thresholds. This process was performed for both pooled patients from all trials as well as patients from each trial separately. Using continuous values of volume and ADC_L as well as phenotypes based on these thresholds, log-rank and Kaplan–Meier data and Cox proportional hazard models were used to understand the relationship between OS and factors, including treatment, patient age, pretreatment contrast enhancing tumor volume (continuous and dichotomized into large and small sizes using the median volume of the cohort), and ADC_L. Lastly, “risk groups” were defined based on a combination of baseline tumor volume and diffusion MR characteristics. Specifically, tumor volume was dichotomized into large and small sizes, based on the median tumor volume in the sample, and ADC_L phenotypes were defined using the optimized thresholds defined above. We hypothesized a combination of small tumors with high ADC_L measurements would have a significant survival advantage compared with large tumors with low ADC_L measurements when treated with anti-VEGF therapy.

Lastly, using the optimal thresholds defined above, we validated the ability for tumor volume and ADC_L phenotypes to predict response to anti-VEGF therapy by examining an independent cohort of patients treated with either bevacizumab or lomustine monotherapy as part of the BELOB trial. Using continuous values of volume and ADC_L as well as phenotypes based on these thresholds, log-rank and Kaplan–Meier data and Cox proportional hazard models were used to test whether volume and ADC_L phenotypes were significant predictors for OS in each of these two therapies.

For all analyses, P < 0.05 was considered statistically significant. No corrections for multiple comparisons were performed. Statistical analyses were performed with Stata 12 (2011) or GraphPad Prism v6.0h (GraphPad Software, Inc.). All errors are presented in standard error of the mean (SEM).

Repeated (double baseline) ADC measurements prior to initiation of therapy were available for 29 of the 30 patients enrolled in the cediranib trial. Measurements of ADC_L demonstrated a COV of 2.50% ± 0.52% SEM and did not significantly differ between the two evaluations (Fig. 3A; paired t test, P = 0.4537). The consistency of ADC_L phenotype classification, defined as the percentage of tumors with the same ADC_L classification at a particular threshold (e.g., high vs. low ADC_L), ranged from 79.3% (23 of 29 with same classification) to 100% (Fig. 3B).

An ADC_L threshold of 1.24 μm²/ms produced the largest OS differences between patient cohorts when calculating the Mantel–Haenszel hazard ratio and corresponding P values for pooled patients as a function of ADC_L thresholds (Fig. 3C and D). This threshold was optimal for the pooled patient cohort, while other thresholds appeared to emerge as specific to the particular anti-VEGF therapy employed (Fig. 3E and F). For example, patients treated with cediranib appeared to have the largest difference in OS when using an ADC_L threshold of 1.16 μm²/ms and patients treated with cabozantinib had the largest difference using an ADC_L threshold of 1.46 μm²/ms, while bevacizumab had a relatively constant HR over all ADC_L thresholds evaluated. The consistency of ADC_L phenotype classification using a threshold of 1.24 μm²/ms, optimal for generalization across all anti-VEGF therapies tested, was approximately 89.7% (26 of 29 with same classification; Fig. 3B), and 55% of these patients had an ADC_L > 1.24 μm²/ms.

Log-rank analysis after stratification based on an ADC_L threshold of 1.24 μm²/ms confirmed a statistically significant difference in OS (Fig. 4A). Notably, patients with high ADC_L had approximately 49% greater OS than patients exhibiting a low ADC_L phenotype (P < 0.0001, HR = 0.5303, median OS = 7.7 vs. 11.6 months). Subsequent univariate analyses of patients within individual treatments showed a significant difference in OS when using a threshold of ADC_L = 1.24 μm²/ms in cediranib monotherapy (Fig. 4B; P = 0.0489, HR = 0.4980), bevacizumab monotherapy (Fig. 4C; P = 0.0050, HR = 0.4545), cabozantinib monotherapy (Fig. 4D; P = 0.0107, HR = 0.4623), and aflibercept monotherapy (Fig. 4E; P = 0.0017, HR = 0.3313).

Multivariable Cox proportional hazards analysis considering treatment, patient age, enhancing tumor volume, and continuous estimates of ADC_L using all anti-VEGF treatment–naïve patients indicated that age (P = 0.0140, HR = 1.0165), volume (P < 0.0001, HR = 1.0158), and ADC_L (P = 0.0002, HR = 0.2555) were all significant predictors of OS (Table 2). Multivariable Cox regression in patients treated with cediranib using continuous tumor volume and ADC_L phenotype (higher or lower than 1.24 μm²/ms) suggested age (P = 0.0156, HR = 1.0488) and ADC_L phenotype (P = 0.0256, HR = 0.3901), but not tumor volume (P = 0.0820, HR = 1.0131), were predictive of OS (Table 3). Within patients treated with bevacizumab monotherapy as part of the BRAIN trial, data suggested that ADC_L phenotype (P = 0.0090, HR = 0.4495) was a significant predictor of OS, while volume trended toward significance (P = 0.0765, HR = 1.0157) and age was not a significant predictor (P = 0.2492, HR = 1.0143). Enhancing tumor volume (P = 0.0034, HR = 1.0114) and ADC_L phenotype (P = 0.0076, HR = 0.4195), but not age (P = 0.2170, HR = 1.0143), were predictive of OS for patients treated with cabozantinib. In patients treated with aflibercept, tumor volume (P = 0.0098, HR = 1.0353) was a significant predictor of OS, while ADC_L phenotype trended toward significance (P = 0.0677, HR = 0.4266).

Because the combined data from all four trials used as a training cohort suggested enhancing tumor volume and ADC_L phenotype were significant, independent predictive factors for OS in recurrent GBM patients treated with anti-VEGF therapy, three “risk groups” were defined based on a combination of pretreatment tumor volume and diffusion characteristics. Patients included in the risk I category had the lowest risk, corresponding to small tumors (<25cc) with high diffusivity (ADC_L > 1.24 μm²/ms); patients included in the risk II category were defined as having either small tumors (<25cc) with low diffusivity (ADC_L < 1.24 μm²/ms) or large tumors (>25cc) with high diffusivity (ADC_L > 1.24 μm²/ms); and patients included in the risk III category were had the highest risk corresponding to large tumors (>25cc) and low diffusivity (ADC_L < 1.24 μm²/ms). [Note that patients in the risk II group consisted of both large tumors with high ADC_L and small tumors with low ADC_L, for which there was no difference in OS (log-rank, P = 0.6779)]. Results demonstrate the combination of enhancing tumor volume and ADC_L further enriched the population, with small, high ADC_L tumors showing more than a three-fold longer OS than patients with large, low ADC_L tumors (Fig. 4H; P < 0.0001, HR = 0.2653, median OS = 4.7 vs. 14.9 months).

To verify ADC_L is a significant, predictive imaging biomarker for anti-VEGF therapy, we examined an independent sample of patients from the BELOB trial treated with either bevacizumab or lomustine monotherapy. Multivariable Cox regression analysis including age, continuous measures of tumor volume, and continuous measures of ADC_L (Table 4) showed that volume (P = 0.0008, HR = 1.0450) and ADC_L (P = 0.0255, HR = 0.0764) were predictive of OS in bevacizumab-treated patients, whereas only enhancing tumor volume (P = 0.0034, HR = 1.0367), not ADC_L (P = 0.2483, HR = 0.3553), was predictive of OS in patients treated with single-agent lomustine. Similar trends were observed when examining age, continuous measures of tumor volume, and ADC_L phenotypes using a threshold of 1.24 μm²/ms (Table 5). Specifically, results showed that tumor volume (P = 0.0004, HR = 1.0472) and ADC_L phenotype (P = 0.0309, HR = 0.4332) were predictive of OS in patients treated with bevacizumab, but only tumor volume (P = 0.0019, HR = 1.0394) was predictive of OS in patients treated with lomustine. Consistent with trends in the training data, risk categorization based on both tumor volume and ADC_L phenotypes showed a significant difference in OS between risk I and III patients treated with bevacizumab (Fig. 4G; P < 0.0001, HR = 0.2093, median OS = 5 vs. 10.5 months). Interestingly, a much less dramatic, but still significant, difference in survival was observed between risk I and III patients treated with lomustine (Fig. 4H; P = 0.0390, HR = 0.4214, median OS = 6.2 vs. 8.3 months), likely driven primarily by differences in tumor volume, as risk II patients had a slightly longer median OS compared with other risk categories (median OS = 11.2 months in risk II patients treated with lomustine).

Results from the current study confirm that pretreatment diffusion MR characteristics within contrast enhancing tumor are a predictive imaging biomarker for overall survival benefit in recurrent GBM patients treated with anti-VEGF therapies. Although these results support previous studies demonstrating the prognostic significance in bevacizumab therapy (18–20), findings from the current study further refine this hypothesis, suggesting that diffusion MR measurements may predict OS during VEGFR-2 inhibition in recurrent GBM, because cediranib, bevacizumab, aflibercept, and cabozantinib all commonly inhibit activity of this receptor. Because VEGFR-2 is thought to be the primary mediator of the proangiogenic effects of VEGF on endothelial cells (31–33) and is one of the primary targets for antiangiogenic therapies in glioblastoma (34), a pretreatment imaging biomarker that could identify patients who will receive a significant survival benefit from VEGFR-2 inhibition would be valuable both for guiding clinical practice and for use in future trials for patient cohort enrichment trials involving anti-VEGF combination strategies.

Data from the current study suggest an ADC_L threshold of 1.24 μm²/ms is the optimal threshold for the best prediction of OS when evaluating data pooled from four different anti-VEGF agents that commonly target VEGFR-2, and this threshold for classification maintained similar groupings in more 85% of patients during repeated evaluations 1 week apart (Fig. 3A). This optimized threshold is consistent with the previous work by Pope and colleagues (19, 20), who empirically identified a threshold of 1.2 μm²/ms for stratifying OS in patients treated with bevacizumab based primarily on the median ADC_L within the datasets evaluated and not based on optimizing prediction of OS. While data from the current study suggest that an ADC_L threshold of 1.24 μm²/ms may be the optimal threshold for the best prediction of OS when pooling data from all three anti-VEGF agents, Fig. 3E and F implies each therapeutic targets may have their own unique signatures. For example, an ADC_L threshold of 1.46 μm²/ms appears to result in the largest difference in OS in patients treated with cabozantinib. It is conceivable that the specific mechanisms or domains in which these anti-VEGF agents act may result in different sensitivity to diffusion MR characteristics. Alternatively, these differences may be related to the activity of secondary therapeutic targets in cediranib (e.g., cKIT, PDGFRβ) and cabozantinib (e.g., MET, AXL, RET). Although provocative, future studies aimed at further eliciting the possible mechanisms for these observations are necessary to properly test this hypothesis.

The particular mechanism underlying the association between diffusion MRI measures of ADC_L and therapeutic efficacy of anti-VEGF drugs remains speculative. A reasonable explanation for this association may be that tumors with high ADC_L have less total tumor burden, because high diffusivity implies more water content and/or less tumor cellularity (i.e., less tumor cells and more extracellular water). Thus, treatments like anti-VEGF therapy that are known to reduce vascular permeability, and subsequently vasogenic edema, may simply consolidate the total tumor burden by removing excess water. While feasible, this explanation implies pretreatment ADC_L may be prognostic for all therapies; however, previous studies have failed to identify similar relationships with OS in recurrent GBM treated with chemotherapies, including temozolomide or lomustine (18). Additionally, isolation of the lower component of the ADC histogram (i.e., ADC_L) may act to “filter out” the edematous and necrotic portions of the enhancing tumor known to have higher ADC measurements. Hence, ADC_L measurements may be more specific to the composition of tumor and/or stroma within the lesion.

Consistent with this hypothesis, a recent differential gene expression study from our laboratory has identified overexpression of decorin (DCN) as a possible mechanism for this altered water diffusivity (35). DCN may increase water diffusivity through direct modulation (softening) of the extracellular matrix (ECM), as DCN acts to modulate the rigidity and stiffness of the ECM by binding with various ECM macromolecules and activating specific matrix metalloproteinases (MMPs; ref. 36). The protein core of DCN is known to bind with a variety of collagen molecules, fibrils, and other macromolecules, acting to significantly increase interfibrillary spacing (37). Because ADC is inversely correlated with fluid viscosity (38, 39) and tortuosity of the ECM (40–44), it is conceivable that high DCN expression results in softer, less viscous tumors with decreased boundaries to fluid within the extracellular space from DCM remodeling, ultimately resulting in a higher measured ADC_L. Future studies aimed at exploring the relationship between diffusion MRI and DCN expression, as well as any relationship between DCN expression and OS benefit when treated with anti-VEGF agents, are warranted in order to link these observations and highlight the underlying mechanism behind both water diffusivity changes and therapeutic benefit from anti-VEGF agents.

The retrospective, multicenter, and international nature of the data in the current study inevitably leads to heterogeneity in MRI acquisition and image quality. However, one may argue that this level of heterogeneity more accurately reflects the variability encountered in the real world, implying diffusion MR may be a robust biomarker suitable for use in clinical practice. We posit a standardized protocol that includes identical T1-weighted pre- and post-contrast scans as well as standardized DWI acquisition, ideally in compliance with the recently outlined International Standardized Brain Tumor Imaging Protocol (45), would provide increased consistency and lower variability in T1 subtraction map-defined enhancing tumor volume and ADC estimation. Additionally, image distortions on diffusion MR images obtained using echo planar imaging may have led to inaccuracies when estimating the ADC characteristics within the contrast-enhancing lesion. Estimates of ADC_L, however, have been shown to be relatively robust even in the presence of distortion (30) and results from the current study suggest a COV of only 2.5%, suggesting image distortion may not have been a critical limitation. Additionally, more in-depth baseline characteristics, including information about post-progressive treatment and molecular information, were not available for the trials used in the current study. Recent evidence suggests that MGMT promoter methylation status may predict response to bevacizumab (23, 46); however, this information was not available for most trials and this trend has not been confirmed in other anti-VEGF therapies. Also, despite there being little evidence that subsequent treatments have an effect on post-progression survival after failure of anti-VEGF therapies, this is nevertheless a limitation of the current study. Lastly, post hoc correction for multiple statistical comparisons was not performed, in part due to the exploratory nature of the current study. Despite a relatively large number of comparisons both within and across treatment paradigms, this lack of correction likely did not lead to substantial type II error, because the effect sizes observed for both volume and ADC_L tended to be large and consistent. Nevertheless, this is a possible limitation to the current study and should be acknowledged.

In summary, the current study suggests that diffusion MR characteristics are independent predictors of long-term therapeutic benefit in recurrent GBM treated with anti-VEGF therapies, including cediranib, bevacizumab, aflibercept, and cabozantinib. Results suggest diffusion MRI signatures can further improve patient stratification, which may be beneficial for future cohort enrichment trials involving anti-VEGF combination strategies and understanding the therapeutic mechanisms that underscore anti-VEGF efficacy and failure.

B.M. Ellingson reports receiving commercial research grants from and is a consultant/advisory board member for Roche/Genentech and Siemens. M. Smits is an employee of Parexel. D.T. Aftab and G.M. Schwab hold ownership interest (including patents) in Exelixis, Inc. W.B. Pope is a consultant/advisory board member for Genentech/Roche. J.F. de Groot is an employee of Ziopharm Oncology, reports receiving commercial research grants from AstraZeneca, Deciphera Pharmaceuticals, Eli Lilly, EMD-Serono, Mundipharma, Novartis, and Sanofi-Aventis, other commercial research support from DSMB, Novella, and VBL Therapeutics, holds ownership interest (including patents) in Gilead and Ziopharm Oncology, and is a consultant/advisory board member for AbbVie, AstraZeneca, Celldex, Deciphera Pharmaceuticals, FivePrime Therapeutics, Inc., Foundation Medicine, Inc., Genentech, Insys Therapeutics, Kadmon, Merck, Novartis, and Novogen. P.Y. Wen reports receiving other commercial research support from Acerta, Agios Pharma, Angiochem, AstraZeneca, Genentech/Roche, GlaxoSmithKline, Karyopharm, Merck, Novartis, Oncoceutics, Sanofi-Aventis, and Vascular Biogenics, speakers bureau honoraria from Merck, and is a consultant/advisory board member for Cavion, Crotice Biosciences, Foundation Medicine, Genentech/Roche, Insys, Monteris, Novartis, and Vascular Biogenics. T. Batchelor reports receiving commercial research grants from Pfizer, and speakers bureau honoraria from Amgen, Merck, and NXDC. M.J. van den Bent reports receiving commercial research grants and other commercial research support from Roche. T.F. Cloughesy is a consultant/advisory board member for AbbVie, Agios, Boston Biomedical, Bristol-Myers Squibb, Celldex, Cortice, Cytrx, INSYS, MedQIA, Merck, NewGen, Novartis, Pfizer, ProNAI, Roche, Sunovion, Tocagen, Upshire Smith, VBL, and Wellcome Trust. No potential conflicts of interest were disclosed by the other authors.

Conception and design: B.M. Ellingson, E.R. Gerstner, W.B. Pope, P.Y. Wen, T.F. Cloughesy

Development of methodology: B.M. Ellingson, R.J. Harris, W.B. Pope, K. Leu, T.F. Cloughesy

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): B.M. Ellingson, E.R. Gerstner, M. Smits, R.Y. Huang, R. Colen, L.E. Abrey, D.T. Aftab, G.M. Schwab, C. Hessel, R. Gahrmann, J. de Groot, P.Y. Wen, T.T. Batchelor, M.J. van den Bent, T.F. Cloughesy

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): B.M. Ellingson, E.R. Gerstner, R.J. Harris, A. Chakhoyan, W.B. Pope, K. Leu, D.C. Woodworth, T.F. Cloughesy

Writing, review, and/or revision of the manuscript: B.M. Ellingson, E.R. Gerstner, M. Smits, R.Y. Huang, R. Colen, L.E. Abrey, D.T. Aftab, G.M. Schwab, C. Hessel, W.B. Pope, K. Leu, D.C. Woodworth, J. de Groot, P.Y. Wen, T.T. Batchelor, M.J. van den Bent, T.F. Cloughesy

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): B.M. Ellingson, L.E. Abrey, R.J. Harris, C. Raymond, D.C. Woodworth

Study supervision: B.M. Ellingson, E.R. Gerstner, T.F. Cloughesy

Other (pre-processing of data): R. Gahrmann

This study was financially supported by the American Cancer Society (ACS) Research Scholar Grant (RSG-15-003-01-CCE; B.M. Ellingson), the American Brain Tumor Association (ABTA) Research Collaboration Grant supported by Humor to Fight the Tumor (ARC1700002; B.M. Ellingson), the National Brain Tumor Society (NBTS) Research Grant (B.M. Ellingson, T.F. Cloughesy), the Art of the Brain (T.F. Cloughesy), the Ziering Family Foundation in memory of Sigi Ziering (Cloughesy), the Singleton Family Foundation (T.F. Cloughesy), NIHR01CA129371 (T.T. Batchelor), NIHR21CA117079 (T.T. Batchelor), and NIHK24CA125440 (T.T. Batchelor).

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.