Assessment of tumor hypoxia and perfusion in recurrent glioblastoma following bevacizumab failure using MRI and 18F-FMISO PET

Tumoral hypoxia correlates with worse outcomes in glioblastoma (GBM). While bevacizumab is routinely used to treat recurrent GBM, it may exacerbate hypoxia. Evofosfamide is a hypoxia-targeting prodrug being tested for recurrent GBM. To characterize resistance to bevacizumab and identify those with recurrent GBM who may benefit from evofosfamide, we ascertained MRI features and hypoxia in patients with GBM progression receiving both agents. Thirty-three patients with recurrent GBM refractory to bevacizumab were enrolled. Patients underwent MR and 18F-FMISO PET imaging at baseline and 28 days. Tumor volumes were determined, MRI and 18F-FMISO PET-derived parameters calculated, and Spearman correlations between parameters assessed. Progression-free survival decreased significantly with hypoxic volume [hazard ratio (HR) = 1.67, 95% confidence interval (CI) 1.14 to 2.46, P = 0.009] and increased significantly with time to the maximum value of the residue (Tmax) (HR = 0.54, 95% CI 0.34 to 0.88, P = 0.01). Overall survival decreased significantly with hypoxic volume (HR = 1.71, 95% CI 1.12 to 12.61, p = 0.01), standardized relative cerebral blood volume (srCBV) (HR = 1.61, 95% CI 1.09 to 2.38, p = 0.02), and increased significantly with Tmax (HR = 0.31, 95% CI 0.15 to 0.62, p < 0.001). Decreases in hypoxic volume correlated with longer overall and progression-free survival, and increases correlated with shorter overall and progression-free survival. Hypoxic volume and volume ratio were positively correlated (rs = 0.77, P < 0.0001), as were hypoxia volume and T1 enhancing tumor volume (rs = 0.75, P < 0.0001). Hypoxia is a key biomarker in patients with bevacizumab-refractory GBM. Hypoxia and srCBV were inversely correlated with patient outcomes. These radiographic features may be useful in evaluating treatment and guiding treatment considerations.

Image acquisition. MRI scans were performed on 3 T MRI scanners (Philips, GE, or Siemens). Each scanning session consisted of 3D pre-and post-contrast T1 weighted, FLAIR, diffusion-weighted MRI (DWI), dynamic contrast enhanced (DCE), and DSC MR images. T1 pre-contrast and FLAIR images were acquired before contrast injection. DCE-MRI and DWI were acquired after the first intravenous injection of 0.1 mmol/ kg of a standard gadolinium-based contrast agent. For DCE-MRI, injection took place after 10 baseline frames were obtained. The second injection was for DSC-MRI and T1 post contrast images.
PET scans were performed on two devices, both of which were calibrated. On a CTI EXACT HR + scanner (Knoxville, TN), acquisition parameters were 63 slices; 2.4-mm thickness; and image size: 128 × 128 × 63. The images were reconstructed using 3D iterative reconstruction with four iterations, 16 subsets, zoom = 2; Gaussian kernel FWHM 5.0 mm, 2D measured attenuation correction, axial filtering, and scatter correction). On a Siemens Biograph40 mCT scanner, acquisition parameters were 75 slices; 3-mm thickness; and image size: 128 × 128 × 75. The images were reconstructed using 3D iterative ordered-subset expectation maximization with two iterations and 21 subsets, time of flight, point-spread function-correction, slice thickness 3 mm, matrix size 128 × 128, in-plane reconstruction pixels size 6.3638 mm × 6.3638 mm, and a Gaussian post-reconstruction convolution kernel with full width at half maximum of 5 mm. A CT scan was used for attenuation correction.
In all cases, patients were injected intravenously with 3.7 MBq/kg of 18 F-FMISO. A 20-min static 18 F-FMISO PET emission image was acquired at about 120 min after injection of 18 F-FMISO.

Survival analyses.
Overall survival time was counted from enrollment until death or end of follow-up.
Progression time was counted from enrollment until progression or death or end of follow-up. The enrollment date was used as start date, and either the death date or last follow-up date was used as the endpoint. Patients who were lost to follow-up or survived to the end of follow-up were considered censored. www.nature.com/scientificreports/ Image analyses. All imaging data were post-processed using MATLAB2016 (Math Works) and OsiriX (Pixmeo, Geneva, Switzerland) with the IB Rad Tech plugin (Imaging Biometrics LLC, Elm Grove, WI). Briefly, using the IB Rad Tech tool, the following processing steps were performed with minimal user intervention. First, T1-pre contrast images were co-registered to T1-post contrast images. Next, delta T1 images were generated based on differences between the standardized T1-post contrast and standardized T1-pre contrast images 23 . Likewise, the DSC images were registered to T1-post contrast images. IB Rad Tech then directs the user to manually draw a reference region of interest (ROI) in normal-appearing white matter to create normalized rCBV and CBF maps (nrCBV, nrCBF). In addition, using the registered DSC-MRI data, IB Rad Tech computed srCBV 22 , rCBF, MTT, TTP and Tmax maps, which did not require a reference ROI. Tumor ROIs were manually drawn on the T1-post contrast images which, according to convention, include the radiologic necrotic region. The total T1 weighted tumor volume was calculated. In addition, an empirically determined threshold of 3000 (IB Rad Tech plugin) was applied to the delta T1 maps within the tumor ROIs to extract the enhancing tumor ROIs (without necrosis) and the subset ROIs that include radiological necrosis only. The enhancing and radiologic necrotic tumor volumes (T1_Vol_et and T1_Vol_nt) were calculated.
The FLAIR ROIs, which typically include both the tumoral and peritumoral regions, were manually drawn on the FLAIR images. The FLAIR tumor volume, FLAIRΔT1 (FLAIR_Vol excluded T1_Vol) and the Volume ratio (T1_Vol/ FLAIR_Vol) were calculated. These ROIs were then applied to the different perfusion parameter maps. The mean, median, and the maximal value of the MTT, TTP, nrCBV, srCBV, nrCBF and Tmax within tumor ROIs were calculated.
FLAIR images were co-registered to the 18 F-FMISO PET images and the FLAIR tumor ROIs were used to determine the tumor ROIs on the 18 F-FMISO PET images. Because FMISO is a freely diffusing tracer, the tumor ROIs on the 18 F-FMISO PET images were determined by expanding the FLAIR ROIs to include all regions where there was FMISO uptake, which extended beyond the original FLAIR tumor ROI 24 . Two 2 cm diameter ROIs on both sides of the cerebellar cortex were used as the image derived blood surrogate to determine the surrogate of tissue to blood ratio (TB ratio) 9 , the mean value of hypoxia volume (HVmean) and the mean of the top 5% of TB pixels (TB5percent). HV was determined by the number of pixels with values above 1.2 on the TB image 9 .
Statistical methods. OS and PFS were graphically described with Kaplan-Meier curves. Imaging parameters were summarized in original units with the median, minimum, and maximum. The significance of variations in OS and PFS with PET and MRI brain imaging parameters were assessed with univariate and multivariate proportional hazards models in log units and standardized to mean zero and unit variance. Results were summarized as hazard ratios (HR) and 95% confidence intervals (CI). The significance of variation in imaging parameters, OS, and PFS with clinic site was assessed with principal component analysis (PCA) and univariate and multivariate proportional hazards models in log units. Multivariate proportional hazards models were fit on the first two principal components, clinic site, and interactions with clinic site. Imaging data were stratified in original units at the median and the resulting survival distributions in each stratum were described with Kaplan-Meier curves and compared with log-rank tests. Relationships between OS and PFS and all log-transformed and standardized imaging parameters were assessed with proportional hazards models and stepwise forward selection; the resulting reduced models were described with stratification, stratum-specific Kaplan-Meier curves and log-rank tests. Spearman correlations between imaging parameters in original units were graphically described with the corrplot R package and the hclust option. SAS Version 9.4 for Windows (SAS Institute, Cary, North Carolina) and R were used throughout (Supplementry Information).
Previous presentation. Some of the data in this article were previously presented at the 2016 Annual Meeting of the American Society for Clinical Oncology, the 22nd annual meeting of the Society for Neuro-Oncology (SNO), San Francisco, California, in 2017 and the SNO 23rd annual meeting, Louisiana, New Orleans, in 2018.

Results
Forty-one subjects were screened for this study; 6 were screen failures (Dana Farber Elevated LFT 1, Declining performance status 1, UT Health Declining performance status 3, Thrombocytopenia and anxiety 1), 35 were enrolled (Dana Farber 17, UT Health 18) and two withdrew consent (Dana Farber 1, UT Health 1). Of the remaining 33, the mean patient age was 46 years (range, 19-76 years). 22 were male and 11 were female, and the ECOG performance status was 0 or 1 in 82.9% of patients. All of the 33 enrolled patients failed standard treatment (randomized in pre-surgery cohorts 1-3 with 9 proceeding to Evo/Bev after surgery and the remainder proceeding directly to Evo/Bev). Of the 33, a total of 28   We extracted the mean imaging parameters within the different ROIs. The median, minimum, and maximum of these mean parameters at baseline are shown in Table 2 www.nature.com/scientificreports/ HV was 28.5 cm 3 (range 0-155.1 cm 3 ). The median srCBV within T1_Vol ROI was 1.2 (range 0.5-5.1). The associations between imaging parameters among the 33 patients are shown in Fig. 2A.
In a proportional hazards model of PFS in terms of clinic site, PCA1 and PCA2, and pairwise interactions, all interactions were non-significant and were removed from the model. In the resulting main effects model, clinic site did not contribute significantly (P = 0.32) and was removed; in the reduced model PCA1 and PCA2 did not contribute significantly (PCA1 HR = 1.019, 95% CI 0.994 to 1.044, P = 0.15, PCA2 HR = 1.002, 95% CI 0.982 to 1.023, P = 0.84). After removing PCA2, PCA1 contributed significantly to the results (HR = 1.020, 95% CI 1.002 to 1.039, P = 0.03).
After forward stepwise reduction, the final OS model contained Tmax, SUVmax, T1_Vol_nt, and nrCBV_ et (Tmax HR = 0.11 P < 0.001, SUVmax HR = 2.14 P = 0.01, T1_Vol_nt HR = 2.07 P = 0.007, nrCBV HR = 1.55 P = 0.01). Restriction to Tmax and nrCBV_et revealed a significant decrease in time to death among those with nrCBV_et above the median and Tmax below the median (n = 8) relative to those with nrCBV_et below the median and Tmax above the median (p = 0.04) (Fig. 3F).

Discussion
This is the first study, to our knowledge, to use multimodality imaging to explore tumor hypoxia, vasculature, and the correlation between radiographic features and outcomes in patients with Bev-refractory recurrent GBM.
The potential prognostic value of anatomic imaging of tumor volumes for OS and PFS in Bev-treated recurrent GBM remains controversial 25,26 . In this study, we sought to assess whether T1 weighted and FLAIR tumor volumes and their ratio are useful predictors of OS and PFS in these patients. Our results show that T1_Vol, T1_Vol_et, T1_Vol_nt and FLAIR_Vol are significantly associated with OS, but not PFS. The Vol_Ratio did not predict either OS or PFS.
Huang et al. assessed associations between tumor volumes and outcomes for recurrent GBM patients treated with Bev 27 ; posttreatment enhancing volume and posttreatment FLAIR volume were significantly associated with OS and PFS. Ellingson et al. explored the relationship between conventional MRI tumor volume and survival for recurrent GBM patients treated with Bev 28 . Bev significantly reduced T1 weighted and FLAIR tumor volumes; both pretreatment and posttreatment FLAIR tumor volumes were not significantly correlated with OS and PFS. Further, both pretreatment and posttreatment T1 weighted enhancing tumor volumes were significantly correlated with PFS, but not OS. The pretreatment ratio of FLAIR to contrast enhancing tumor volume was a predictor of OS and PFS, unlike the posttreatment ratio of FLAIR to contrast enhancing tumor volume.
We calculated the ratio of FLAIR to contrast enhancing tumor volume; this ratio was not significantly correlated to OS and PFS in our cohort. We acquired data within 3 days prior to treatment. In contrast, Huang et al. acquired their data between 3 to 6 weeks after treatment while Ellingson et al. acquired their posttreatment data between 6 to 8 weeks after treatment. These inconsistencies may have affected the results 27 . As the antipermeability effect of Bev may change over time, and thus potentially cause tumor volumes to also change with time, associations between tumor volumes and outcomes may vary accordingly. In another study, Schmainda et al. found that percent changes of rCBV relative to baseline at 2 and 16 weeks were significantly related to OS, www.nature.com/scientificreports/ but not at 8 weeks 16 . This suggests that correlations of OS and PFS to imaging biomarkers may be sensitive to when imaging is done. In vivo tumor hypoxia and vasculature imaging could enhance our understanding of the pathophysiologic mechanisms of GBM, and in turn optimize timing and doses of radiotherapy and chemotherapy. Our results indicate that the median time to GBM progression increased significantly relative to historical data 29 ; specifically, median PFS in the current study (53 days; 95% CI 42 to 113) was higher than historical controls (37.5 days; 95% CI, 34 to 42 days; P < 0.001). However, the median time to death in our study, 129 days (95% CI 86 to 199 days) or 4.3 months (95% CI 2.9 to 6.6 months), was not significantly different (P = 0.10) from the median time to death of 5.9 months (95% CI 4.4 to 7.6 months) more recently observed in 55 patients with recurrent GBM taking Bev beyond initial Bev progression 4 . PFS at 4 months in our study (31%) was not significantly different from PFS-4 (38%) in 99 patients who received subsequent therapy after progression on one of five consecutive, single-arm, phase II clinical trials evaluating bevacizumab regimens for recurrent GBM (P = 0.40).
Perfusion (srCBV) was tightly associated with hypoxia (HV, HVmean, TBpeak, TBmax, TB5percent), perhaps because most of the hypoxia and hypervascularization zones were in contrast enhancing areas 30 . Hypoxia (HV, HVmean, TBpeak, TBmax, TB5percent) and tumor volumes (T1_Vol, T1_Vol_et, Vol_Ratio) were also tightly associated. Whereas other PET and MRI parameters were weakly associated, as previously noted 24 , those results  www.nature.com/scientificreports/ may be related to locating hypoxia zones in different ROIs 30 , or may indicate that these parameters have a unique and complementary role relative to tumor status 24 . Normalized and standardized rCBV are the most common DSC-MRI metrics used for evaluating tumors 19 . We used spin echo (Dana Farber Cancer Institute) and gradient echo (UT Health at San Antonio) based methods to acquire our DSC data. The former is sensitive to capillary-sized vessels, whereas the latter is sensitive to a broad range of vessel sizes 19,31 . The value of the rCBV obtained with the gradient echo method is much larger than that obtained with the spin echo method in high-grade gliomas 32 . Some studies have confirmed that these two methods are comparable in terms of related parameters in in vivo studies 33,34 . However, srCBV has greater reproducibility 17 than nrCBV and does not require the manual step of drawing a reference ROI, as does normalized rCBV. This may also explain why significant correlations were found with srCBV but not nrCBV. We also confirmed this finding with PCA, including all spin echo-and gradient echo-based DSC parameters. The mean within-subject difference for the first two principal components did not vary significantly with clinic site-or, therefore, with method (Fig. 2B).
Higher srCBV values were associated with poor outcome, which suggests that more vascular tumors convey a poorer prognosis. These results are consistent with several previous studies 14,36 . Patients with a within-subject mean nrCBV greater than 1.75 have a significantly shorter PFS than those with nrCBV less than 1.75 37 . We found that patients with a mean srCBV greater than 0.99 had a significantly shorter PFS than those with srCBV less than 0.99 (HR = 3.4 P = 0.02). This threshold is consistent with the tissue-validated thresholds previously determined by Hu et al 38 and Prah et al 39 .
Tmax is considered a promising prognostic parameter 40 , although it is known to be complex, and may be affected by different factors, such as the arrival delay between the arterial input function and the tissue contrast agent concentration, arterial abnormalities that cause bolus temporal dispersion, and the MTT of the contrast agent, that mainly reflect the characteristics of microvascularization 40 . A reduction in Tmax may suggest high vascularity in a tumor 41 . We found that Tmax, Tmax_et, and Tmax_nt were significantly associated with OS and PFS, again suggesting that high vascularity in a tumor is associated with worse survival.
Hypoxia is exacerbated by antiangiogenic treatment and is important in tumor development, angiogenesis, and growth, and in treatment resistance. In our study, nrCBF and HV were positively correlated, suggesting that the larger the baseline HV value, the higher the CBF value in the tumor area after Bev treatment. A recent multicenter study of patients with newly diagnosed GBM reported a positive correlation between elevated nrCBF and HV 24 . Based on these data, we hypothesize that abnormal vascular anatomy and compromised vascular function within the tumor area contribute to heterogeneous blood flow and tumor hypoxia.
Evo is an investigational hypoxia-activated prodrug designed to be activated under hypoxic conditions in tumors. Evo plus Bev treatment in Bev-refractory GBM could therefore reduce HV and hypoxia-induced resistant tumors in Bev-refractory GBM 20 . In our study, HV was associated with worse OS and PFS. As shown in Table 3, higher HV reduced OS and PFS in the univariate proportional hazards regressions (P = 0.009 for PFS and P = 0.01 for OS). Spence et al. studied the HV and TBmax in newly diagnosed GBM before chemo-and radiation therapy with 18 F-FMISO PET to assess their impact on OS and PFS; volume and intensity of hypoxia in GBM before radiotherapy were strongly associated with poor OS and PFS 42 . These results indicate that HV is a meaningful biomarker in GBM assessment and could provide more information than conventional anatomic imaging. In another study, Kawai et al. found a correlation between 18 F-FMISO uptake in tumors and the expression of vascular endothelial growth factor, and confirmed that the volume and intensity of hypoxia were associated with OS 43 . Although their patients were newly diagnosed and ours were recurrent, these findings suggest that HV may be a reliable biomarker in tumor detection and treatment assessment in both settings. Additionally, based on our results, the timing of HV assessment does not change its reliability as an accurate marker.
Necrosis is an important characteristic of GBM tumors 44,45 thought to be related to hypoxia 46,47 , and may affect treatment outcomes. Our data support a negative correlation between tumor necrosis and OS in our cohort of patients with recurrent GBM (Table 3), consistent with some studies of newly diagnosed GBM. Hammoud et al. reported negative correlations between degree of necrosis and OS 48 . Lacroix et al. found that smaller areas of necrosis were significantly associated with OS 49 . Taken together, these findings suggest necrosis might be an important biomarker.
We found that a negative change in HV was associated with longer OS and PFS and a positive change in HV was associated with shorter OS and PFS, although these results did not reach statistical significance. Yamaguchi et al. evaluated the performance of BEV treatment based on 18 F-FMISO accumulation. They found that the 18 F-FMISO responders had significantly longer OS than that of 18 F-FMISO non-responders 50 . These results suggest that some patients with refractory GBM may benefit from combined Evo and Bev treatment and that HV may identify a subpopulation of Bev-refractory patients who would most benefit from this treatment. Further studies in larger populations are needed to confirm our results.
There were some limitations to our study. Some variance in acquisition parameters and scanners were present at the two sites for perfusion data. Although the parameters calculated by these two methods are comparable in in vivo studies, there may be some differences not accounted for. We acquired our 18 F-FMISO data at 2 h post 18 F-FMISO injection. The time between 18 F-FMISO injection and data acquisition may affect 18 F-FMISO uptake 51 , with some studies suggesting 18 F-FMISO data should be acquired at 4 h post 18  www.nature.com/scientificreports/ Our study was also limited by the numbers of patients and time points. P-values are presented without correction for multiple testing, possibly increasing the Type 1 error, requiring interpretation considering correlation, biological plausibility, and consistency with other results in this study and with results published in other studies.

Conclusions
In this cohort of 33 patients with bevacizumab-refractory GBM, hypoxia is a key biomarker for therapeutic efficacy. High hypoxia volume, enhancing and non-enhancing tumor volumes, and srCBV were all inversely correlated to patient outcomes. If validated in larger studies, these imaging biomarkers may be useful for detection of GBM, and for planning treatments and assessing responses.