Comparison of Voxel-Wise Tumor Perfusion Changes Measured With Dynamic Contrast-Enhanced (DCE) MRI and Volumetric DCE CT in Patients With Metastatic Brain Cancer Treated with Radiosurgery

Dynamic contrast-enhanced (DCE)-MRI metrics are evaluated against volumetric DCE-CT quantitative parameters as a standard for tracer-kinetic validation using a common 4-dimensional temporal dynamic analysis platform in tumor perfusion measurements following stereotactic radiosurgery (SRS) for brain metastases. Patients treated with SRS as part of Research Ethics Board-approved clinical trials underwent volumetric DCE-CT and DCE-MRI at baseline, then at 7 and 21 days after SRS. Temporal dynamic analysis was used to create 3-dimensional pharmacokinetic parameter maps for both modalities. Individual vascular input functions were selected for DCE-CT and a population function was used for DCE-MRI. Semiquantitative and pharmacokinetic DCE parameters were assessed using a modified Tofts model within each tumor at every time point for both modalities for characterization of perfusion and capillary permeability, as well as their dependency on precontrast relaxation times (TRs), T10, and input function. Direct voxel-to-voxel Pearson analysis showed statistically significant correlations between CT and magnetic resonance which peaked at day 7 for Ktrans (R = 0.74, P ≤ .0001). The strongest correlation to DCE-CT measurements was found with DCE-MRI analysis using voxel-wise T10 maps (R = 0.575, P < .001) instead of assigning a fixed T10 value. Comparison of histogram features showed statistically significant correlations between modalities over all tumors for median Ktrans (R = 0.42, P = .01), median area under the enhancement curve (iAUC90) (R = 0.55, P < .01), and median iAUC90 skewness (R = 0.34, P = .03). Statistically significant, strong correlations were found for voxel-wise Ktrans, iAUC90, and ve values between DCE-CT and DCE-MRI. For DCE-MRI, the implementation of voxel-wise T10 maps plays a key role in ensuring the accuracy of heterogeneous pharmacokinetic maps.


INTRODUCTION
Dynamic contrast-enhanced (DCE) imaging can be useful for evaluating vascular injury and endothelial permeability changes following radiation therapy, including ablative therapy such as stereotactic radiosurgery (SRS) or when combined with antiangiogenic therapy (1). Preclinical work in a murine intracranial glioma model showed that early quantitative changes in diffu-sion and perfusion magnetic resonance imaging (MRI) metrics reflect treatment responses soon after initiating combinatorial radiation and antiangiogenic therapies (2).
The development of DCE-MRI techniques has seen a rapid growth in its translation into the field of radiation therapy clinical trials (3,4), but DCE-MRI measures of tumor vascular physiology have shown heterogeneous results across studies, and this may reflect the variability in the magnetic resonance (MR) acquisition and analysis approaches across different studies and institutions, as well as MR vendors (5)(6)(7). Given the potential for DCE-MRI imaging metrics to provide early indicators of therapy-induced changes in the tumor microenvironment, it is imperative to obtain a better understanding of these imaging biomarkers for guiding adaptive and potentially individualize therapy approaches in the future. A recent 4-dimensional temporal dynamic analysis (TDA) method, which enables voxel-based, parametric analysis based on patient-specific dynamic behavior of contrast flow, may provide a more standardized approach for DCE-MRI analysis, including its validation against DCE-computed tomography (CT) (8). DCE-CT is a gold standard based on its high spatial and temporal resolution acquisition, and its highly linear and accurate relation between signal and contrast agent concentrate. However, reproducibility of either DCE-CT or DCE-MRI alone has been low (9,10), and output parameters from either of the imaging techniques have not correlated well, with very few publications reporting on direct in vivo comparison in the same tumor using both imaging techniques. One preclinical paper studied reproducibility and absolute values of DCE-MRI and DCE-CT biomarkers in a C6 glioma model, highlighting that the techniques may have similar reproducibility but that their derived absolute parameter values are not equivalent (11). Two different kinetic models were used for the DCE-CT and DCE-MRI analyses in 2 different software applications. Korporaal et al. (12) and Kallehauge et al. (13) reported on in vivo comparisons of DCE-CT and DCE-MRI in patients with prostatic and cervical cancer, respectively. Despite voxel-based acquisitions, the analyses were reported in median values and gamma analysis used to assess spatial variation in kinetic parameters in Kallehauge et al. This study reports our early clinical experience with tumor perfusion measurements following SRS for brain metastases using both volumetric DCE-CT and DCE-MRI in the same patients supported by a common TDA framework. It is hypothesized that analyzing contrast enhancement data from both modalities in a unified and voxel-based approach will strengthen the correlations between their parametric output values. This supports the concept that low-molecular-weight contrast agents can indeed help derive tumor permeability and perfusion heterogeneity independent of the imaging modality provided the image analysis methods are standardized.

Patients and Treatment
Serial volumetric DCE-CT and DCE-MRI data from patients enrolled in a Research Ethics Board-approved clinical trial evaluating multiparametric imaging biomarkers of response to single-fraction radiosurgery (SRS) for brain metastases were included.

Correlation Between DCE-CT and DCE-MRI Pharmacokinetic Parameters
To clinically validate the voxel-based TDA algorithm on DCE-MRI data against the DCE-CT data, the functional parametric maps of K trans and area under the enhancement curve (iAUC 90 ) from TDA DCE-CT and DCE-MRI were compared for their capa-bilities to evaluate tumor perfusion and permeability in this subset of patients with metastatic brain cancer for a total of 40 cases, that is, 14 tumors from 9 patients scanned at 3 time points. The following 3-step validation approach was taken: (1) Evaluate the stability of arterial input function (AIF) versus vascular input function (VIF) measurements with volumetric DCE-CT and its impact on resulting CT perfusion parameters to create a gold standard benchmark. (2) Pearson correlation (regression analysis) and Bland-Altman analyses were completed to evaluate relationships between perfusion parameters obtained from DCE-CT and DCE-MR and identify any systematic bias between the 2 imaging modalities. This was done for tumor functional histograms and for direct voxel-to-voxel comparison within the target volume. (3) Evaluate the effect of using individualized voxel-based T 10 measurements versus a fixed TR on resulting DCE-MRI perfusion metrics and compare both with DCE-CT.

Volumetric DCE-CT Image Acquisition
DCE-CT data were acquired on a 320-section CT scanner (Aquilion ONE™, Toshiba, Japan) that has previously been extensively characterized for its use in radiation oncology as a radiotherapy simulator (14,15) and DCE measurements (16). Scan parameters were as follows: 80 kV, 100 mA, 1-second rotation, and 0.468-ϫ 0.468-ϫ 1-mm reconstruction resolution. A total of 60 mL of iodixanol (Visipaque® 320) was injected intravenously at 5 mL/s synchronized with the start of scanning. The brain tumor DCE time sequence consisted of different sampling frequencies as follows: every 1.5 second for the first 30 seconds, every 5 second for the next 90 seconds, and every 10 second up to 180 seconds to allow for permeability modeling while balancing the scan dose with the measurement sampling rate. A noncontrast volume was acquired before contrast injection for baseline corrections and image registration. The associated volumetric dose index (CTDI-vol) was ϳ100 mGy compared with 60 mGy typically reported for a routine adult head scan in this scan mode (17).

DCE-MRI Acquisition
On the same day as the DCE-CT, each patient underwent MRI on a 3 Tesla Verio system (Siemens Medical Systems, Erlangen, Germany) with VQ gradients (peak amplitude, 40 mT/m; peak slew rate, 200 T/m/s) and a 12-channel head coil. MR scanning included endogenous T 1 mapping using the variable flip angle (VFA) technique ( and section thickness ϭ 1.5 mm). VFA used flip angles of 2, 10, 20, and 30°, and the scan time was 50 second per flip angle. For DCE-MRI, the temporal resolution was 5.8 seconds, and 45 repetitions were acquired (scan time ϭ 4 minute and 19 seconds). A weight-based bolus of gadolinium (Gd) contrast (Magnevist, Bayer AG, Leverkusen, Germany) was injected intravenously at 4 mL/s after a 20-s delay from the start of scanning followed by a 20-mL saline injection.

Common Parametric Perfusion Analysis Framework
Image Registration. Tumor regions of interest for each time point were delineated using semiautomated segmentation on the T 1 -weighted Gd-enhanced MR image by an expert observer and registered to the baseline DCE-MRI and DCE-CT images in GammaPlan v9 (Elekta, Sweden). Patients undergoing DCE-CT were imaged using a thermoplastic S-frame immobilization mask (QFix); thus, motion was not an issue. Immobilization could not be used in MRI, and varying degrees of motion were observed during the 3D-FLASH acquisitions, both across DCE-MRI frames and between the VFA images used for endogenous T 1 mapping. Compensatory image registration was performed with a custom script in MATLAB (The MathWorks Inc., Natick, Massachusetts), to register all DCE-MRI images to the baseline image and all VFA scans to the 20°scan. Although every voxel in the brain scan was analyzed, clinical treatment contours were exported from GammaPlan so that the analysis could be correlated to the corresponding radiation clinical target volume.

Signal-to-Contrast Concentration Conversion
For DCE-CT data, the CT numbers were converted from Hounsfield units (HU) to contrast concentration based on previous calibration phantom experiments with static and dynamic concentrations of contrast, resulting in a linear scaling of 33 HU/mgI/mL (19).
The DCE-MRI signal modeling used the standard equations for 3D-FLASH magnitude signal conversion to Gd-diethylenetriaminepentaacetic acid (DTPA) concentration as follows: where E 1 ϭ exp(ϪTR/T 1 ), is the flip angle, and S 0 and S are the relative signal enhancements before and after contrast injection, respectively (20). Magnitude signal enhancement was converted to signal using the following equation: 1 where T 10 and T 1 are the spin-lattice TRs before and after contrast injection, respectively, r 1 is the relaxivity of the contrast agent at 3 T (3.3 L/mmol/s) (21), and C is the concentration of Magnevist. For dynamic image analysis, the average of signals at the first 3 time points provided an estimate of the signal baseline. The voxel-based T 10 maps were the VFA T 1 maps derived for every patient and every imaging session using precontrast signal profiles measured at 2, 10, 20, and 30°flip angles (18).

Voxel-Based Pharmacokinetic Modeling
The TDA algorithm applies a classification scheme to each voxel on the basis of temporal characteristics of the voxel's contrast enhancement over time and then creates parametric maps within these TDA-derived masks based on a specified kinetic model to iteratively improve classification and parameter sensitivity (22). The modified Tofts model (23) is commonly used in patients with brain perfusion on the basis of the hypothesis of weak vascularization and increased permeability in tumors (24,25). It describes the arterial input, C a (t), and tissue enhancement curve, C t (t), as follows: In addition to semiquantitative measures as the integrated iAUC 90 , the following are the resulting functional parameters of interest: K trans , the transfer constant from the blood plasma into the extracellular extravascular space (EES); K ep , the transfer constant from the EES back to the blood plasma; and v e , the extravascular extracellular volume. V b is the whole blood volume per unit of tissue (mL/g). The hematocrit value, H ct , was assumed to be 0.4 for all cases. The AIF CT was chosen in the internal carotid artery for DCE-CT and compared with VIF CT in the sagittal sinus. For this study, a population-based input function (AIF MRI ) was used for DCE-MRI analysis because of variability in the flip angle between patients (26) and it allowed for a robust comparison of the impact of the analysis methodology against DCE-CT (27). A 3D voxel mask, as well as a separate sum of squared errors mask, was created for each functional parameter to show the quality of fit of the transport model. Finally, a histogram moment analysis was performed for each parameter inside the tumor mask for assessing the standard deviation, skew, and kurtosis of the histogram shape.

Statistical Analysis
Statistical analysis was performed in Matlab. The histogram parameters estimated by volumetric DCE-CT and DCE-MRI were compared via Bland-Altman analysis, in which differences in perfusion parameter values between the modalities were plotted against the mean of the pair of values, and Student t test and Pearson correlation. All statistical analyses were 2-sided and values with P Ͻ .05 were deemed statistically significant. Direct voxel-to-voxel comparison was done in a similar way with the addition of a variance component analysis to estimate inter-day variance (28).

Patient Demographics
A cohort of 9 patients with a total of 14 metastatic brain tumors (lung cancer, 3; breast cancer, 3; melanoma, 2; and squamous cell carcinoma, 1) underwent imaging at baseline (day 0 -before radiosurgery), day 7, and day 20 after radiosurgery (total number of data sets ϭ 40). One patient with 2 tumors missed the day 7 appointment. SRS with a mean dose of 20.5 Gy (18 -21 Gy) using GammaKnife Perfexion (Elekta, Sweden) was performed.

VIF Selection
VIFs for voxel-based kinetic analysis were selected in the internal carotid artery (AIF CT ) and sagittal sinus (VIF CT ). The AIF CT curves DCE-MRI and DCE-CT in Patients With Metastatic Brain Cancer Treated With Radiosurgery for patient 1 are shown in Figure 1A, together with the populationbased AIF MRI curve, whereas Figure 1B highlights the (small) variations in input curves for all DCE-CT measurements. The mean AIF CT peak (438 Ϯ 68 HU) was slightly higher than VIF CT (382 Ϯ 100 HU), with the corresponding AIF CT onset time (6.7 Ϯ 3.5 seconds) earlier than VIF CT (12.2 Ϯ 4.1 seconds). The bottom panels of Figure 1 show examples of individual-phase MRI measurements of AIF for 2 different flip angles. Given the variation in flip angle acquisition during the clinical trial, it was decided to use a population-based AIF for the DCE-MRI analysis since they are quite similar to the CT-based AIF. The impact of using AIF CT or VIF CT as the input function for parametric modeling is shown to provide equivalent performance in Figure 2, indicating the high correlation and interchangeability between the uses of either input function.

Tumor Perfusion Evaluation: Volumetric TDA CT vs MRI
Median Histogram Correlation. Statistically significant moderate correlations between MR and CT were found for median K trans (R ϭ 0.42, P ϭ .01) and median iAUC 90 (R ϭ 0.40, P ϭ .01) over all time points (n ϭ 40; 1 patient was removed because of blooming leakage effect seen on MRI). The variation over the different imaging days is listed in more detail in Table 1 with a strong correlation at baseline for median K trans (R ϭ 0.513, P ϭ .008) and median v e (R ϭ 0.58, P ϭ .03). The percentage change in median K trans and v e over time showed a statistically significant correlation for early (day 7) change relative to that for the baseline (R ϭ 0.64, P ϭ .02) but not at day 20 (R ϭ 0.07, P ϭ .81), which is likely related to the small brain metastases volumes at day 20 (mean of volumes at day 20 ϭ 1.4 cc). The percentage change in median iAUC 90 values (Ϫ25% Ϯ 67%) did not significantly correlate on either day 7 (P ϭ .29) or day 20 (P ϭ .94).
Bland-Altman agreement between volumetric DCE-CT and DCE-MRI is shown in Figure 3. K trans median ( Figure 3A) and iAUC 90 median ( Figure 3B) values remained largely within 0.105 min Ϫ1 Limits of Agreement (LoA) with some bias (0.05 min Ϫ1 ), suggesting that the 2 modalities may be interchangeable. The absolute K trans values from DCE-CT are slightly lower than those from DCE-MRI. A strong correlation between iAUC 90 values from CT and MRI is shown in Figure 3B, which exhibits a linear trend, as the bolus volumes of Visipaque® and Gd-DTPA contrast agent injected were different but consistent, and because the iAUC 90 parameter is a cumulative measurement of  enhancement, it appears as a proportional bias. This bias disappears when normalizing to the bolus amount injected, and the normalized iAUC 90 shows little bias (0.025) and small LoA (0.103) values ( Figure 3C). In terms of other histogram metrics, only the skewness change in iAUC 90 has any significant correlation between the 2 modalities (P ϭ .03).

Voxel-Wise Correlation
Modalities were also compared on a direct voxel-to-voxel basis within the tumor regions of interest. With a resolution of 128 ϫ 128 ϫ 40 matched to both CT and MRI, moderate-to-strong correlations in K trans values were found for day 0 (R ϭ 0.37, P Ͻ .001), day 7 (R ϭ 0.74, P Ͻ .001), and day 20 (R ϭ 0.52, P Ͻ .001). The correlation over all imaging days combined is shown in Figure 4. The color gradient highlights the density (frequency) of K trans values from low incidence (blue) to high incidence (red).
The Bland-Altman plot shows excellent agreement for K trans and normalized area under the curve values between modalities over all days, with very little evidence of K trans bias (0.009 min Ϫ1 , LoA 0.16 min Ϫ1 ). Correlation of voxel-wise v e was statistically significant but relatively low with an R value of 0.22.

Sensitivity of DCE-MRI Perfusion Modeling Parameters to Individualized T10 Values
It was hypothesized that individual precontrast TRs' T 10 values would make a significant difference in resulting perfusion parameters as per a prior study by Heye et al. (6) in patients with cervical cancer. Figure 5 shows the qualitative impact of using a global T 10 of 1600 milliseconds or 2400 milliseconds versus individual voxel-based T 10 maps on K trans values for 1 patient at the different imaging days (29). The higher T 10 value was based on experimentally probing the highest voxel-based values in a number of tumors. The lower value is based on the measured median tumor T 10 value over all available data, which was 1572 Ϯ 594 milliseconds (n ϭ 41). Using a constant T 10 value of 2400 milliseconds over voxel-wise T 10 resulted in significantly higher K trans (0.3 Ϯ 0.14 min Ϫ1 ) and iAUC 90 values (P Ͻ .0006) compared with CT. Use of a static T 10 value of 1600 milliseconds produced a regression correlation between CT and MRI K trans values that was closer to the voxel-based T 10 results. This is reflected in Figure 6 which shows the highest voxel-wise correlation between K trans values from CT and MRI T 10 (R ϭ 0.575, P Ͻ .0001) for all imaging days and patients, including good interchangeability as can be seen in the Bland-Altman plot.

DISCUSSION
This work investigated the use of a unified analysis platform (based on the TDA implementation) to compare DCE-MRI against DCE-CT parameters in patients with brain metastases treated with SRS. The use of volumetric DCE-CT was considered a gold standard given its linear signal-to-contrast concentration relationship and proven robustness (22). Based on DCE-CT data, AIF and VIF appear to be interchangeable in generating similar K trans values. This confirms that the use of individual VIF in DCE MRI analysis is a reasonable approach. In contrast, the application of different T 10 values considerably impacted the resulting K trans values ( Figures 5  and 6), as suggested by Heye et al. (6). The use of individual T 1 mapping with voxel-wise precontrast TRs for each DCE-MRI image set in the pharmacokinetic analysis resulted in the highest voxel-wise correlation between K trans values from CT and MRI T 10 (R ϭ 0.575, P Ͻ .0001) across all imaging days and in all patients. This approach is likely to provide more accurate quantitative evaluation of parametric tumor heterogeneity (6).
Our results show that the use of the TDA approach for both DCE-MRI and DCE-CT data results in well-correlated (R ϭ ϳ0.5) median DCE parameter values in the tumor (K trans , iAUC 90 , and V e ). This correlation increased even more (R ϭ 0.77) when performing a direct voxel-wise analysis (K trans and iAUC 90 ) and, in doing so, capturing tumor heterogeneity. The voxel-wise correlation of extravascular volume fraction, v e , was low between the 2 modalities, but it was highly statistically significant (R ϭ 0.22, P Ͻ .001). Because this parameter is highly dependent on K ep , the transfer constant from the EES goes back to the blood plasma; this discrepancy can be explained by the differences in the molecular weight and the composition of the 2 contrast agents affecting their diffusion and extraction fraction in the interstitial space. In the evaluation of median 3D tumor volume histogram, the Bland-Altman plots showed significant interchangeability, but there was some bias (0.05 min Ϫ1 ) toward higher MRI perfusion values. Using the 0.02 absolute differences in K trans values between the AIF CT and VIF CT correlation as a standard error, this offset may or may not be statistically relevant.
This bias disappeared with voxel-wise Bland-Altman analysis, which suggests that the 2 modalities may be interchangeable when assessing the vascular permeability of brain metastases in a voxel-based approach. Other histogram values such as the skew or kurtosis of the parametric distributions were not statistically significant. This further suggests that voxel-based analysis is required to capture tumor heterogeneity and that this is not necessarily a normal distribution.
The correlations of DCE parameters were consistently lower at day 20 than at day 7 for both absolute values and their relative change from baseline. This is likely due to the very small tumor volumes (1.4 cc, range: 0.1-5.3 cc) at a later time point, which, consequently, will result in a lower number of voxels available for reliable correlative statistics.
The correlations found in this work are significantly higher than those previously reported using different analysis methods with nonvolumetric DCE-CT measurements and/or limited image registration between the different modalities (30)   comparing DCE-CT and DCE-MRI in locally advanced cervical cancer (13) using a gamma similarity measure and scaling the DEC-CT results based on the amount of injected contrast compared with DCE-MRI.
At the time this imaging study protocol was developed and the trial was started, the QIBA profile for DCE-MRI recommended a VFA technique for T 1 mapping. As the study had started with the VFA technique, the same technique was continued for the duration of the study.
Other T 1 mapping techniques-such as the use of inversion pulses or incorporation of time-efficient RF mappingmay have greater accuracy and could be explored for future studies.
Some of the remaining differences may be inherent to the type and molecular weight of the injected contrast agent, with iodexol (CT) being larger and heavier than Gd-DTPA (MRI). This is expected to affect its transport capabilities across the same capillary network and could explain the lower correlation in v e fraction.
In summary, our results validated the use of the TDA method as a common analysis approach for both DCE-MRI and DCE-CT data through strong voxel-wise histogram correlation across modalities and highlighted the need for voxel-wise, individualized T 10 mapping in patients to derive meaningful DCE metrics.

Advances in Knowledge
(1) High correlation between DCE-MRI and DCE-CT using TDA validates the use of this common platform for both quantitative and semiquantitative parameters across imaging modalities, which will enable standardized functional analysis methods. (2) Voxel-wise histogram analysis of perfusion and permeability values better elicits tumor heterogeneity and results in significantly higher correlations between modalities compared with region of interest-based (mean/median) values. (3) The assumed value for T 10 precontrast relaxation significantly impacts the accuracy of heterogeneous pharmacokinetic maps, and the strongest correlation between DCE-MRI and DCE-CT was observed when individually measured voxel-wise T 10 maps were implemented. (4) Volumetric DCE-CT analysis showed that the input to pharmacokinetic calculations in the CT could be an AIF or VIF measurement, as the resulting parameters showed significant agreement with little to no bias.

Implications for Patient Care
(1) The implementation of individually measured voxel-wise precontrast relaxation maps is strongly recommended when quantitative pharmacokinetic analysis with DCE-MRI is planned given its impact on the resulting parameter accuracy.