Prediction of survival and progression in glioblastoma patients using temporal perfusion changes during radiochemotherapy

The aim of this study was to investigate changes in structural magnetic resonance imaging (MRI) according to the RANO criteria and perfusionand permeability related metrics derived from dynamic contrast-enhanced MRI (DCE) and dynamic susceptibility contrast MRI (DSC) during radiochemotherapy for prediction of progression and survival in glioblastoma. LARSSON, Christopher, et al. Prediction of survival and progression in glioblastoma patients using temporal perfusion changes during radiochemotherapy. Magnetic Resonance Imaging, 2020, vol. 68, p. 106-112 DOI : 10.1016/j.mri.2020.01.012


Introduction
Glioblastoma (GBM) is the most common primary brain cancer in adults. Prognosis is dire, with an average overall survival (OS) of 12-15 months [1]. The OS range is, however, wide and some patients respond beneficial to therapy and have a two-year OS of 20-25%, making early and correct prognosis challenging [1,2]. Accurate prognosis is preferred by cancer patients [3], in addition, early prognostic biomarkers are warranted for timely change of therapy in cases showing tumor recurrence and treatment failure.
The Response Assessment in Neuro-Oncology (RANO) criteria, considered to be the gold standard in GBM assessment, are based on the visual radiological evaluation of structural image series [4]. These criteria include estimating changes in MRI-based measures of tumor size, measured on two-dimensional structural T1-weighted (T1w) and T2-weighted (T2w) or Fluid Attenuated Inversion Recovery (FLAIR)weighted images, in conjunction with corticosteroid use and clinical deterioration [4]. Disease response is grouped into four categories: complete response, partial response, stable disease, and progressive disease. Concerns about the limited ability of structural measurements to reflect pathologic heterogeneity and predict OS have led to an interest in more sophisticated MRI tools [5,6]. Perfusion-weighted MRI techniques are used increasingly in the assessment of GBMs and most tumor protocols now include either dynamic contrast-enhanced MRI, (DCE) and/or dynamic susceptibility contrast MRI (DSC) [7]. Perfusion imaging for prognostics is promising in several studies [8][9][10]. However, there is no consensus on how to best perform this assessment during conventional therapy [11]. Studies vary with respect to imaging time-point relative to treatment, statistical analysis, and the metrics investigated. Furthermore, few have systematically investigated differences and similarities between DSC and DCE in terms of their ability to predict outcome [10].
To this end, we hypothesized that perfusion derived metrics will help improve the prediction of OS in GBM patients, compared to conventional volumetric or radiologic assessment by the RANO criteria alone. Furthermore, our study aimed to assess the temporal trends of estimated parameters derived from DCE and DSC imaging and differences in their predictive value.

Patient population
Patients with histologically confirmed high-grade glioma (grade III and IV) operated at our institution between May 2010 and May 2012 were eligible for inclusion in this prospective study at start of radiochemotherapy treatment. Approval from the regional ethics committee and written informed consent was obtained from all patients before imaging start. Exclusion criteria were impaired renal function (estimated glomerular filtration rate (eGFR) <60 ml/min), claustrophobia and artifacts from surgical clips near the area of interest. A series of 23 patients were included and available for the final analysis; for key demographics see Table 1. Treatment included surgery when possible, followed by stereotactic radiotherapy approximately four weeks after surgery with concomitant and adjuvant chemotherapy with temozolomide according to the standard treatment protocol proposed by Stupp et al. [2]. Imaging was performed immediately before the start of radiochemotherapy, every second week during this treatment, as well as two weeks after treatment. At each imaging time-point, clinical assessment was done using the Karnofsky performance status (KPS) [12].
Perfusion-weighted imaging: DCE were acquired using a 3D saturation recovery (SR) GRE sequence (TE/TR/flip angle (FA) = 2.5/ 8.2 ms/26°, voxel size 2 × 2 × 4 mm 3 (interpolated to 1.8 × 1.8 × 4 mm 3 ), matrix 120 × 120, 11 slices), a SENSE factor of 2 was used [13]. Each slice was acquired after application of a nonselective saturation prepulse with a saturation time delay (TD). Centric phase ordering was used so that centre of k-space was recorded at time TD. TD/sampling interval was 80 ms/3.4 s (first seven subjects) in a double-echo sequence and 50 ms/2.1 s in the remaining subjects, giving a total of 100 or 150 dynamic images respectively [14]. For the DSC a 2D spin echo (SE) echo planar imaging (EPI) (TE/TR = 70/1349 ms, voxel size 1.88 × 1.88 × 4.0 mm 3 , matrix 120 × 120, 13 slices, sampling interval of 1.33 s) sequence was used. DCE was performed before DSC, and 0.1 mmol/kg body weight gadobutrol (Gadovist®, Bayer Schering Pharma AG, Berlin, Germany) was injected after baseline imaging for both DCE and DSC using a power injector, at a rate of 3 mL/s and 5 mL/s respectively, immediately followed by a 20 mL saline flush.

Co-registration and region-of-interest creation
Co-registration was performed using SPM8 in a hierarchic manner. Pre-treatment non-contrast T1w image sets served as reference, and non-contrast T1w images from each time-point were co-registered to the reference image. Contrast enhanced T1w-and FLAIR images were registered to already registered non-contrast T1w images. DCE-data was registered to contrast-enhanced T1w images and DSC-data to FLAIR images for optimal results. The framework is shown in Supplementary  Fig. 1. Two regions-of-interest (ROIs) were generated using a semi-automatic method previously described [15]. In short, a contrast-enhancing tumor (CET) ROI was defined from hyper-intense regions in the contrast enhanced T1w images. Thick linear enhancement in the contrast enhanced T1w images (>2 pixels thick) was included in the ROIs as this has shown the same prognostic significance as nodular enhancement early post operatively [16]. A non-enhancing tumor (NET) ROI was defined from regions of high signal intensity in the FLAIR images. The CET was subtracted from the NET to avoid overlapping ROIs. All auto-generated ROIs were edited and approved by a radiologist (4 years of experience). ROIs smaller than 0.5 mL were excluded from the analysis.

Visual assessments
Radiographic progression-free survival (PFS) was defined as time to progressive disease according to the RANO criteria [4]. Three radiologists (4-22 years of experience) made a consensus agreement for each patient case. In addition to the four RANO categories, a fifth category -pseudoprogression -was included. Pseudoprogression was defined by new or growing contrast enhancing lesions where the radiographic findings were stable for more than six months and later decreased in size or was confirmed by surgery (one patient). Volumes of CET and NET, in addition to volumetric change from baseline, were estimated from the structural images.

Table 1
Key demographics of all patients and for the two survival groups based on median split. Median values and estimated 95% confidence intervals shown in parenthesis. KPS = Karnofsky Performance Score, CET = Contrast enhanced tumor, NET = non-enhanced tumor, GTR = gross total resection, STR = sub-total resection.

Perfusion processing
The extended two-compartment Tofts model was used for the DCE analysis, yielding statistical parametric maps of the unidirectional contrast agent capillary transfer constant (K trans ) and plasma volume (v p ) [17]. A fixed baseline T 1 of 1490 ms across all patients was used for further analysis [18,19]. Patient-specific carry-on arterial input functions (AIFs) obtained from the baseline scan was used for both DCE and DSC analysis at every time-point as previously described [20]. In addition to parametric maps of K trans and v p , CBF from all time-points (CBF T1 ) was also estimated from the DCE data. An example of the model fit and the residual function from the DCE is shown in Supplementary Fig. 2The DSC series was corrected for geometric EPI distortions [21], and parametric maps representing CBF (CBF T2 ) and CBV (CBV T2 ) were generated [22]. Absolute measures of CBF T2 and CBV T2 using a patient-specific carry-on AIF from the first scan [20], as well as values relative to normal appearing grey-and white matter (rCBF T2 and rCBV T2 ) were estimated from the DSC data [23]. CBF was estimated based on the same method in the DCE and DSC series [24]. All image analysis was performed using the nordicICE analysis package (NordicNeuroLab AS, Bergen, Norway).

Statistical analysis
Median OS and PFS, from the date of radiochemotherapy start, were calculated from Kaplan-Meier plots and the subjects were divided according to a median split in two groups. Median and 90-percentile values of all perfusion parameters were investigated for absolute values and the percentage change from the baseline scan (δ), for both ROIs. Temporal change from the baseline examination in all patients was assessed using Wilcoxon signed-rank test, with Holm-Bonferroni corrections for multiple comparisons. Differences in the two survival groups were assessed by Mann-Whitney U test. Receiver operating characteristics (ROC) curves were investigated for significant parameters from the Mann-Whitney U test and dichotomized based upon Youdens index [25]. Parameter correlation was assessed using Spearman's correlation. Log-rank test and Kaplan-Meier plots for OS were then generated based on the cut-off values from the ROC curve analysis.
In an effort to accurately predict OS, in compliance with earlier published work, a survival index was produced by combining parameters from DCE and DSC [26]. Cox proportional hazards analysis using age as a time dependent covariate and changes in log-transformed K trans and rCBV T2 from baseline to week eight was performed. Using the sum of the estimated regression coefficients β for each parameter an index for each patient was calculated using the following formula and correlated to OS [26].
x Ktrans x rCBV x age 1 2 3 The predictive value of the index was assessed using a concordance index [27]. Findings were considered significant if p < 0.05. All statistical analyses were performed in Statistical Package of the Social Sciences (SPSS) v.22 (IBM Corporation, New York, USA).
Temporal changes in tumor sub-volumes (CET and NET), DCE and DSC parameters for all patients during-and two weeks past radiochemotherapy, are shown in Fig. 2a and Supplementary Fig. 3. CET decreased at all time-points (p < 0.05). NET volume showed a significant increase in absolute and relative volume at the second and third time-point respectively (p < 0.05). 90 percentile values of CBF T1 and v p decreased two weeks past the end of radiochemotherapy in both CET and NET (p < 0.05). Significant decreases in rCBF T2 and rCBV T2 were evident at four weeks in both median and 90-percentile values in CET (p < 0.05). Representative images of the evolution of structural volume and perfusion metrics in two sample patients are shown in Fig. 3.
The temporal evolution of δ-values in the two survival groups is shown in Fig. 2b and Supplementary Fig. 4. Both δ CBF T1 and K trans in CET showed increasing differences during treatment. Patients with stable or decreasing CBF T1 /K trans at two weeks after radiochemotherapy had prolonged survival of 33.1/31.5 months compared to patients with increasing values (16.8/13.9 months; p = 0.036/0.016). ROC curve analysis revealed a cut-off value for δ CBF T1 and δ K trans of 10% and 30% from the Youden index (p = 0.016 and p = 0.021). The Kaplan Meier plots of both parameters are shown in Fig. 4a along with two sample patients (Fig. 4b) and parameter histograms (Fig. 4c). δ K trans and CBF T1 in CET correlated significantly at all time-points (r 2 > 0.847; p < 0.001).
Univariate Cox regression revealed age as a significant predictor for OS (p = 0.013). Other known clinical predictors including tumor volume, KPS, and surgical resection were not found to be significant. No continuous parameter was found significant in the univariate analysis. Multiparametric Cox regression of log-transformed δ K trans and δ rCBV T2 were significant (p = 0.028 and p = 0.025) for OS. The index was significantly negatively correlated with OS (r 2 = 0.843; p < 0.001). A strong concordance with a Harrels C score of 0.881 was found for the index. All beta coefficients were positive, signifying that higher age, increasing K trans and increase in rCBV T2 from baseline were associated with shorter OS.

Discussion
This study demonstrates a decrease in CET volume and a decrease in DSC derived CBF T2, and CBV T2 in GBM patients during-and two weeks after radiochemotherapy. In addition, the prognostic value of K trans and CBF T1 from DCE in CET was superior to the DSC-derived parameters. An index of age and log-transformed K trans and rCBV T2 showed strong correlation with survival.
In agreement with previous studies, PFS and OS varied much within the patient population (Fig. 1). However, PFS is not a good marker for survival prediction [26]. According to the RANO criteria, progression can only be evaluated three months after the end of radiochemotherapy at the earliest due to pseudoprogression [4]. Pseudoprogression occurs in 20-30% of patients during radiochemotherapy, effectively making PFS unreliable as a predictive biomarker until pseudoprogression can be separated from true progression [28]. Studies distinguishing pseudoprogression from true progression using imaging biomarkers are emerging, but prospective evaluation is warranted before clinical use [29,30].
Volumetric anatomical assessment is not part of the RANO criteria due to lack of standardization and measurement tool availability [4]. Despite a decrease in CET and an increase in NET volume in all patients during radiochemotherapy, univariate analysis of volumetric change did not separate the two survival groups or predict survival. In a recent study of 125 patients with treatment naïve GBM comparing clinical parameters with MRI parameters, no association (p = 0.855) between OS and pre-treatment tumor volume was found [31]. Conversely, Li et al., in a study similar to ours, found a correlation between OS and volume before radiochemotherapy of non-enhancing lesions in 64 patients [32]. While post-surgical tumor volume is a known predictor for OS, extent of resection other than gross-total or sub-total resection is poorly understood [33].
Studies have shown that high rCBV T2 may predict shorter OS in GBM patients [34,35]. The DSC-derived parameters rCBV T2 and rCBF T2 decreased significantly during the radiochemotherapy period, but showed no differences between the two survival groups. One reason for this could be the use of a SE rather than the more commonly used GRE C. Larsson, et al. Magnetic Resonance Imaging 68 (2020) 106-112 DSC method. The choice of SE over GRE was based on pilot data showing more artifacts from surgical clips when GRE was used, expected to reduce the number of valid DSC-MRI datasets for analysis. SE generally has a poorer signal-noise-ratio than GRE DSC, due to lower contrast-agent sensitivity, but is at the same time more sensitive to microvascular perfusion [36]. The observed decline in rCBV T2 and rCBF T2 compared to CBF T1 and v p might therefore reflect a difference in radio-resistance between different-sized blood vessels and auto-regulation of blood flow following damage to capillary endothelial cells [37,38]. Tumoral CBF T1 was twice as high as that of apparently unaffected tissue and CBF T2 equal to unaffected tissue at therapy start without any group differences. Interestingly, no DSC-based perfusion measures were significantly different between the patient groups, while DCE-derived CBF T1 was higher in the subpopulation with the worst prognosis after finishing radiochemotherapy. Simulations looking at shunting in tumor vessels have shown that absence of anti-shunting mechanisms can lead to functional shunting, a condition with high average flow and substantial hypoxia [39]. A lack of anti-shunting mechanisms possibly explains why an increase in CBF T1 is associated with a lower-than-average OS. The blood-brain-barrier depends upon the integrity of gap junctions, a signaling pathway propagating the antishunting mechanism. The correlation between high K trans and CBF T1 might be explained by a higher degree of disruption of these gapjunctions in patients with a poorer prognosis. The decrease in K trans and CBF T1 in patients with greater OS could implicate a lesser degree of shunting and normalization of the endothelial integrity in the tumor vessels, a condition thought to promote better effect of anti-cancer therapy [40]. The early overall trend of CBF T1 and K trans is in line with previous investigations of perfusion-changes in GBM patients [41]. Møller et al. found, in a study of 11 patients, an increase in CBF T1 after one week of radiochemotherapy and a decline in v p after 5 weeks, similar to our findings. No difference in PFS was found, while other studies have shown that increasing values of K trans is a potential biomarker in early prognostics [8]. Earlier work by Li et al. showed a statistical decrease in rCBV T2 and normalized peak height (a biomarker similar to rCBF T2 ) after radiochemotherapy [32]. These results are in good agreement with the results presented here.
The study has limitations. In this study, the extended Tofts model was chosen as the kinetic model for the DCE analysis as this is the Fig. 1. Event chart for 23 patients sorted by time of radiographic progression. Color represents RANO category, the inner dashed line represents progressionfree survival, and the outer dashed line represents overall survival. Subjects still alive at time of analysis are marked with *. The median split was between patient eleven and twelve. In the bottom row the blue lines represent subjects with a worse overall survival (OS) and the red lines shows subjects with a better OS based upon the median split groups. 95% confidence intervals are shown at each time point. Significant change from baseline in the upper row or significant difference between the two groups are shown by a plus (p < 0.05) or an asterisk (p < 0.01). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) method most often used in GBM imaging. The chosen method is dependent on the quality of the data and can be assessed by an incremental modeling method [42,43]. This showed that 70% of all tumor voxels were best described using the extended Tofts model in this material. A 'carry-on' AIF approach was implemented, where the patient specific AIF obtained at the baseline scan was used for all subsequent scans. The use of carry-on AIF is based on the assumption that measurement errors in individual AIF measured at each time-point would be larger than the actual patient specific variation in bulk flow to the brain over the eight-week study duration. The improved reproducibility of carry-on AIF over individual AIFs is supported in studies using double baseline data for both DSC and DCE [20,44]. Although the patients received different surgical treatment, no difference in OS or in estimated parameters were found between the different groups. There was, not surprisingly, a difference in CET volume between the groups (p < 0.003). However, to remove bias from tumor location and extent of surgery we compared each parameter to the baseline exam and excluded all CET and NET smaller than 0.5 ml. Thus, only temporal evolution of each tumor regardless of these factors was included in the survival index and δ values.

Conclusion
Comparing structural and perfusion MRI in GBM patients during radiochemotherapy treatment, temporal changes are more apparent in CET and DSC-derived parameters compared to parameters from DCE. The DCE-derived metrics K trans and CBF T1 in CET were found to be the most promising parameters for early OS prediction. A multiparametric approach of change in K trans and rCBV T2 shows promise as an early predictor biomarker in this patient group.

Compliance with ethical standards
Author Atle Bjørnerud is a paid consultant of NordicNeuroLab AS. No other conflict of interest was reported from any of the authors.
No funding was received for the writing of this study. Informed consent from all participants was acquired before imaging according to the ethical region committee and the Helsinki declaration from 1964.