All patients had recurrent histologically proven glioblastoma or gliosarcoma (pre-WHO 2016 classification). Detailed inclusion and exclusion criteria have been published (16). Patients received bevacizumab (10 mg/kg IV, days 1 and 15 of a 28-day cycle) and were randomized to receive either temozolomide (75 mg/m² p.o., days 1-21 during the first 28-day cycle; 100 mg/m² for cycle 2 and beyond) or irinotecan (125 mg/m² IV, days 1 and 15 of a 28-day cycle); both treatment arms were pooled for this study due to small sample size.

Imaging was performed at 1.5T (Siemens Espree, Siemens Avanto, GE Signa Excite, GE HDx) or 3T (GE HDx, GE Excite). Conventional MRI included pre-contrast T1-weighted, T2-weighted, FLAIR, and diffusion-weighted imaging. After intravenous injection of 0.1 mmol/kg of standard gadolinium-based agent, axial 2D spin-echo (2D-T1) and 3D gradient-echo (3D-T1) T1-weighted images were acquired. Contrast agent administered for conventional post-contrast imaging served as “pre-load” for subsequent DSC-MRI, potentially diminishing contrast extravasation-associated T1 contribution to DSC-MRI signal (21–23). Echo-planar gradient-echo (TE=30-40ms) DSC-MRI was performed with TR=1.3-1.5s (120 repetitions), flip angle=90°, slice thickness=5mm (0-2.5mm gap), matrix=128x128, and FOV=22-24cm. Images were acquired for 1 min before and 2 min after bolus injection of 0.1 mmol/kg gadolinium-based contrast agent. Imaging was performed at baseline, 2 weeks after bevacizumab initiation, and after 2 and 4 treatment cycles (8 and 16 weeks). Complete MRI parameters are on the ACRIN website (https://www.acr.org/Research/Clinical-Research/ACRIN-Legacy-Trials).

Central reader methods were previously described (16). All local imaging was retrospectively transmitted to ACRIN for central review. Two primary readers independently measured 2D-T1 largest diameter of contrast enhancement and maximum perpendicular diameter for each target lesion. Time of 2D-T1 progression, and radiologic response status at each time point were determined using Macdonald (24) and RANO (25) threshold criteria. Steroid dosage and clinical status were unavailable to ACRIN readers. An adjudicator settled discordant times to progression.

Normalized and standardized rCBV (nRCBV, sRCBV) maps were computed using OsiriX open-source software with the IB Neuro™ plug-in (Imaging Biometrics LLC, Elm Grove, WI). On a voxel-wise basis, baseline pre-bolus mean signal intensity was determined, omitting the five initial time points; the truncated signal-time series was converted to a relaxivity-time series, ΔR2*(t); rCBV was estimated using trapezoidal integration of ΔR2*(t) over post-bolus time points (first-pass plus post-bolus tail) and a post-processing leakage correction algorithm (21–23). rCBV maps were normalized to mean rCBV (nRCBV) in contralateral normal-appearing white matter consistently located across all longitudinal studies, and standardized rCBV (sRCBV) maps were produced using a published technique (26).

For semi-automatic lesion segmentation, we manually defined the region of lesion enhancement, excluding hemorrhage and macrovessels, on difference images computed from co-registered standardized pre- and post-contrast T1-weighted images using IB Delta Suite™ (Imaging Biometrics LLC, Elm Grove, WI). We further constrained regions of enhancement using empirical thresholds, excluding central necrosis, and edited segmentations to exclude non-lesion voxels (27). These ROIs were applied to nRCBV and sRCBV maps co-registered to the 2D-T1 images, from which mean values were extracted. This process was repeated with new ROIs at each time point.

At weeks 2, 8 and 16, we analyzed patients who did not progress on 2D-T1. Patients were classified as decreasing enhancement if both readers rated partial or complete response by the specified time point; otherwise, patients were classified as stable enhancement. Percent change in nRCBV and sRCBV were calculated from baseline for each time point, and summary statistics including mean, standard deviation, median and range were computed.

To assess predictive ability and discrimination, receiver operating characteristic (ROC) curves were derived for continuous percent change in nRCBV and sRCBV at each time point, using OS status at 1 year (OS-1) and PFS status at 6 months (PFS-6) as reference standards, where OS-1 and PFS-6 were defined from patient registration. Area under the ROC curve (AUC) and associated 95% confidence intervals were computed empirically. As 2D-T1 response status is not a continuous marker, ROC analysis cannot be performed; instead, association with OS-1 and PFS-6 was examined using Fisher’s exact test.

To assess the association with PFS and OS generally, and to compare survival times, Kaplan-Meier estimates and the log-rank test were used to compare groups defined using percent change in nRCBV and sRCBV (increase, ≥0, vs. decrease, <0) and 2D-T1 response status (decreasing enhancement vs. stable enhancement). For the time-to-event Kaplan-Meier analysis, OS and PFS were calculated from the date of the respective scan for each of week 2, week 8 and week 16.

For week 16 data, an exploratory analysis was also conducted, using both 2D-T1 response status and percent change in nRCBV to create a three-tier patient stratification: 2D-T1 progression, 2D-T1 non-progression with increased nRCBV, and 2D-T1 non-progression with decreased nRCBV. Kaplan-Meier estimates were computed, and the log-rank test was used to compare time to death.

Statistical computations were performed using SAS 9.4 (SAS Institute, Cary, NC) or R version 4.1.3 (R project: http://www.r-project.org/). All statistical tests were two-sided, with p-values <0.05 considered statistically significant. As the reported analyses are post hoc and were considered hypothesis-generating, no adjustment was made for multiplicity of inference.

The entire ACRIN 6677 study cohort with DSC-MRI results was described previously (16). A total of 37/123 enrolled patients had DSC-MRI, 21 of whom had data sufficient for analysis, defined as having both a baseline and at least one post-baseline scan with interpretable DSC-MRI. After excluding 2D-T1 progressors, there were 13, 15, and 8 2D-T1 non-progressors available for analysis at weeks 2, 8, and 16, respectively. There were no 2D-T1 progressors at week 2. Table 1 provides salient demographic and clinical data for analyzed patients, including age, sex, surgery, location of primary tumor, treatment arm, and indicators for availability of week 2, week 8 and week 16 DSC-MRI data. Table 2 provides summary statistics of the percent change in nRCBV and sRCBV at each time point.

We evaluated the ability of DSC MRI to predict PFS among 2D-T1 non-progressors at weeks 2, 8 and 16 separately using two different approaches. First, we determined whether percent change in nRCBV and sRCBV predict PFS-6 using ROC analysis. There were 12 and 14 patients at week 2 and 8, respectively, whose progression status was confirmed after the respective scan and before 6 months. Of those patients, 6/12 (50%) and 10/14 (71%) were progression-free at 6 months; none of the 2D-T1 non-progressors at week 16 had progressed at 6 months, precluding ROC analysis ( Table 2 ). Figure 1 shows the distribution of percent change in nRCBV and sRCBV at week 2 (A), week 8 (B), and week 16 (C) by PFS-6 status. Table 3 provides estimates of ROC AUC using PFS-6 as the reference standard, along with associated 95% confidence intervals. Percent change in nRCBV and sRCBV at week 2 exhibited high AUC values (AUC [95% CI] = 0.94 [0.82–1] and 0.92 [0.74–1], respectively), indicating very good discrimination. Corresponding values at week 8 were lower (AUC [95% CI] = 0.83 [0.59–1] and 0.75 [0.47–1], respectively). By contrast, 2D-T1 response status (stable vs. decreasing enhancement) at weeks 2 and 8 was not associated with PFS-6 in the subset of 2D-T1 non-progressors (p=1.0 and p=0.58, respectively; cross tabulations not shown).

We also compared time to progression/death after dichotomizing patients with increase (≥0) vs. decrease (<0) in nRCBV and sRCBV. At week 2, patients with increased blood volume had significantly shorter PFS compared to patients with decreased blood volume for both nRCBV (median 1.7 months post week 2 scan vs. 12.6 months, p=0.003) and sRCBV (median 1.7 months post week 2 scan vs. 12.6 months, p=0.03) ( Figure 2 ). By comparison, PFS did not differ significantly between groups based on 2D-T1 response status (median 3.7 months post week 2 scan for decreasing enhancement vs. 6.4 months for stable enhancement, respectively, p=0.44) ( Figure 2 ). At week 8, neither percent change in nRCBV (p=0.40), percent change in sRCBV (p=0.19), nor 2D-T1 response status (p=0.33) were associated with PFS ( Figure 3 ). At week 16, there were no patients with increase in sRCBV which precluded comparison of PFS. nRCBV was not predictive of PFS at week 16 (p=0.09), but 2D-T1 response status was borderline significant (p= 0.05) ( Figure 4 ); however, PFS did not differ by 2D-T1 response status among the entire cohort of 2D-T1 non-progressors at week 16 (p=0.66, n=45), of which this small number of patients also with DSC-MRI data (n=7) was a subset.

These results suggest that at week 2, among 2D-T1 non-progressors, percent change in nRCBV and sRCBV are good prognostic markers for both PFS-6 (evaluated as continuous markers) and PFS in general (dichotomized as increase vs. decrease in rCBV) and outperform the 2D-T1 response criteria. At week 8, neither nRCBV nor sRCBV could further distinguish PFS in general (dichotomized as increase vs. decrease in rCBV), and the 2D-T1 response criteria are not useful. At week 16, nRCBV did not yield a statistically significant difference in PFS (dichotomized as increase vs. decrease in rCBV), although the sample size at this time point was small.

The ability of percent change in nRCBV and sRCBV to predict OS at week 2 and week 16 in the entire patient cohort was previously described (20); here, we test whether it can predict OS in the subset of 2D-T1 non-progressors. Because no patients progressed at week 2, our week 2 cohort is identical to that published previously (20), where only OS data were presented. Schmainda et al. showed that percent change in nRCBV at week 2 differed significantly between patients who died by 1 year versus those who did not (p=0.045, ROC AUC [95% CI] = 0.85 [0.62–1]), and that patients with increased nRCBV at week 2 had poorer OS compared to those with decreased nRCBV (p=0.002), with a similar finding at week 2 for sRCBV (p=0.03). For clarity and ease of presentation, these results are re-incorporated into the findings for the subset of 2D-T1 non-progressors below.

Figure 1 shows the distribution of percent change in nRCBV and sRCBV at week 2 (A), week 8 (B), and week 16 (C) by OS-1 status. Estimates of ROC AUC using OS-1 as the reference standard, along with associated 95% CIs, are shown in Table 3 . Percent change in nRCBV (AUC [95% CI] = 0.85 [0.62–1]) and sRCBV (AUC [95% CI] = 0.83 [0.58–1]) at week 2 are good prognostic markers of OS-1. AUC estimates at week 16 were similar but exhibited wider confidence intervals due to the smaller number of available patients; AUC estimates at week 8 were poor. Response status as determined by 2D-T1 at weeks 2, 8 and 16 were not associated with OS-1 (p=1.0, p=1.0, and p=1.0, respectively; cross tabulations not shown).

At week 2, patients with increased blood volume had significantly shorter survival compared to patients with decreased blood volume for both nRCBV (median 3.4 months post week 2 scan vs. not reached, p=0.002) and sRCBV (median 5.0 months post week 2 scan vs. not reached, p=0.03), whereas OS did not differ significantly based on 2D-T1 response status (p=0.86) ( Figure 2 ). At week 8, neither percent change in nRCBV (p=0.74), percent change in sRCBV (n=0.56), nor 2D-T1 response status (p=0.69) were associated with OS ( Figure 3 ). At week 16, patients with increased sRCBV had significantly shorter survival compared to patients with decreased sRCBV (median 3.6 months post week 16 scan vs. not reached, p=0.008), although the same comparison did not reach statistical significance for nRCBV (p=0.08) ( Figure 4 ). OS did not differ significantly at week 16 based on 2D-T1 response status (p=0.21) ( Figure 4 ).

These results suggest that at week 2, among 2D-T1 non-progressors, percent change in nRCBV and sRCBV are good prognostic markers for both OS-1 (evaluated as continuous markers) and OS in general (dichotomized as increase vs. decrease in rCBV), and again outperform the 2D-T1 response criteria. At week 8, neither percent change in nRCBV or sRCBV nor the 2D-T1 response criteria were predictive of survival. At week 16, percent change in sRCBV is associated with OS, although patient counts are small; 2D-T1 response status at week 16 was not predictive of survival.

As an exploratory analysis, we further analyzed the week 16 data to compare OS using both 2D-T1 response status and percent change in nRCBV to create a three-tier patient stratification: progression by 2D-T1 forms one patient group, and percent change in nRCBV is used to further divide the 2D-T1 non-progressors. As there were no 2D-T1 progressors at week 2 among patients with available DSC-MRI data, week 16 was selected for this analysis due to the larger number of 2D-T1 progressors (n=33), and a modest number of non-progressors with available nRCBV data (n=8, 4 with positive change from baseline). The survival curves for the stratification are shown in Figure 5 . Overall, a statistically significant difference in OS was observed between the three defined groups (p=0.005). OS did not differ significantly between 2D-T1 progressors and 2D-T1 non-progressors with positive change in nRCBV (median survival 3.3 months post week 16 scan vs. 9.2 months, respectively; p=0.13), suggesting that 2D-T1 non-progressors with increasing nRCBV may have a prognosis like that of 2D-T1 progressors. However, this result would require validation with a larger sample size.