Automated Processing of Dynamic Contrast-Enhanced MRI: Correlation of Advanced Pharmacokinetic Metrics with Tumor Grade in Pediatric Brain Tumors

BACKGROUND AND PURPOSE: Pharmacokinetic parameters from dynamic contrast-enhanced MR imaging have proved useful for differentiating brain tumor grades in adults. In this study, we retrospectively reviewed dynamic contrast-enhanced perfusion data from children with newly diagnosed brain tumors and analyzed the pharmacokinetic parameters correlating with tumor grade. MATERIALS AND METHODS: Dynamic contrast-enhanced MR imaging data from 38 patients were analyzed by using commercially available software. Subjects were categorized into 2 groups based on pathologic analyses consisting of low-grade (World Health Organization I and II) and high-grade (World Health Organization III and IV) tumors. Pharmacokinetic parameters were compared between the 2 groups by using linear regression models. For parameters that were statistically distinct between the 2 groups, sensitivity and specificity were also estimated. RESULTS: Eighteen tumors were classified as low-grade, and 20, as high-grade. Transfer constant from the blood plasma into the extracellular extravascular space (Ktrans), rate constant from extracellular extravascular space back into blood plasma (Kep), and extracellular extravascular volume fraction (Ve) were all significantly correlated with tumor grade; high-grade tumors showed higher Ktrans, higher Kep, and lower Ve. Although all 3 parameters had high specificity (range, 82%–100%), Kep had the highest specificity for both grades. Optimal sensitivity was achieved for Ve, with a combined sensitivity of 76% (compared with 71% for Ktrans and Kep). CONCLUSIONS: Pharmacokinetic parameters derived from dynamic contrast-enhanced MR imaging can effectively discriminate low- and high-grade pediatric brain tumors.

P ediatric brain tumors are the most common cause of death from solid tumors, with an incidence rate of 5.57 cases per 100,000. 1 Recent advances in the molecular characterization and treatment of brain tumors 2 have made their proper classification by using imaging techniques critical. Conventional MR imaging is the technique of choice for preoperative diagnosis and evaluation of the child with an intracranial neoplasm because of its multipla-nar capability and superior anatomic detail and resolution. Advanced imaging techniques such as MR perfusion are used to complement structural imaging, providing further insight into tumor physiology. In adults, dynamic contrast-enhanced (DCE) MR perfusion has been used to determine tumor grade [3][4][5] and to distinguish pseudoprogression from tumor recurrence, 6 thus affecting treatment.
While dynamic susceptibility contrast perfusion and DCE-MR perfusion in adult brain tumors have been extensively studied in the literature, particularly for monitoring tumor antiangiogenesis treatments, 7-11 DCE-MR imaging studies in pediatric brain tumors have been scarce [12][13][14][15][16][17][18] and have not focused on tumor grading.
Multiparametric methods to characterize and monitor brain tumors have also shown great promise. 19,20 DCE-MR imaging is particularly suited to multiparametric analyses that require image registration between modalities because it does not have geometric distortion due to susceptibility effects, unlike other advanced MR imaging modalities such as dynamic susceptibility contrast perfusion imaging and diffusion imaging.
In this study, we retrospectively reviewed DCE perfusion data from children with newly diagnosed brain tumors during a 2-year period at our institution and analyzed the pharmacokinetic tumor permeability perfusion parameters correlating with tumor grade.

Subjects
The study was performed with the approval of the institutional review board at the Dana Farber Cancer Institute. Children who presented with a brain mass and had undergone DCE perfusion studies were included. Of 52 patients identified with brain masses who had undergone DCE imaging, 6 patients had final diagnoses that were not brain tumors, 6 had nonenhancing tumors and therefore were not eligible for DCE-MR imaging analysis, and 2 patients were excluded due to motion. Thirty-eight patients were included in this study: 14 girls and 24 boys; age range, 0.30 -18.14 years (median age, 6.01 years; mean age, 7.83 years).
2) DCE-MR imaging sequence consisting of 50 phases, 7 seconds apart, with flip angle ϭ 15°, TR ϭ 4 seconds, TE ϭ minimum. FOV, section thickness, and scan locations were identical to those in the T1 mapping sequences. A single bolus of gadobutrol (Gadavist, 0.1 mmol/kg body weight; Bayer Schering Pharma, Berlin, Germany) was injected 20 seconds after the start of scanning at an injection rate of 2 mL/s.

MR Imaging Postprocessing
MR images were transferred to a VersaVue workstation (iCAD, Nashua, New Hampshire) for automated processing by using OmniLook software (iCad). T1 maps were automatically calculated from the variable flip angle images 21 to yield native T1 of the tissue. The 2-compartment Tofts model 22 was used for the voxelwise calculation of advanced pharmacokinetic parameters such as the transfer constant from the blood plasma into the extracellular extravascular space (K trans ), rate constant from extracellular extravascular space back into blood plasma (K ep ), extracellular extravascular volume fraction (V e ), fractional blood plasma volume (V p ), and initial area under gadolinium curve at 60 seconds (IAUGC 60 ). The model of Weinmann et al 23 for blood plasma concentration was used along with a relaxivity of 5.1 L ⅐ mmol Ϫ1 ⅐ s Ϫ1 for the contrast agent.
ROIs were drawn on each section of tumor around contrastenhancing portions of the tumor by an imaging data analyst or by a PhD scientist and verified by a Certificate of Added Qualification-certified neuroradiologist, and the mean (over voxels) and SDs of each of the variables were recorded for statistical analysis. We included only voxels that could be fit to the model in the computation of mean and SD, excluding areas of cyst, and we took care to exclude vessels from the ROI.

Statistical Analysis
Subjects were categorized into 2 groups based on pathologic analyses consisting of low-grade (World Health Organization I and II) and high-grade (World Health Organization III and IV) tumors. All the pharmacokinetic parameters described above, along with T1 of the tissue, were compared between the 2 groups by using linear regression models with each parameter as a dependent variable (the outcome) and tumor grade as a categoric independent variable (low-grade ϭ 0, high-grade ϭ 1). For parameters significantly distinct between the 2 groups, sensitivity and specificity were also estimated.
Given the non-normal distribution of all parameters, summary statistics reported throughout included median and interquartile ranges. In addition, confidence intervals were estimated via bootstrapping with replacement (2000 draws).
Sensitivity and specificity were estimated as follows: First, the CIs for individual parameter medians were used for thresholding. For each parameter, there were 2 confidence intervals, 1 for the median of high-grade tumors and 1 for the median of low-grade tumors. The lower CI for intervals of statistically higher values and the upper CI for intervals of statistically lower values were used as thresholds. For example, if a parameter median was significantly higher for high-grade than low-grade tumors, then any high-grade parameter value at or above the lower CI for the group median was considered a true-positive and any value below this CI was considered a false-negative (or a false-positive for lowgrade). Similarly, any low-grade parameter value at or below the upper CI for the group median was considered a true-positive, and any value above this CI was considered a false-negative (or a false-positive for high-grade).
There was no statistically significant difference (P ϭ .8) between patient age and tumor grade. For low-grade tumors, the median patient age was 5.52 years (25th to 75th quartiles ϭ 2.62-12.97 years), and for high-grade tumors, the median patient age was 6.88 years (25th to 75th quartiles ϭ 3.72-19.38 years).
The linear regression model results of the pharmacokinetic parameters are summarized in Table 1. The regression coefficient corresponding to tumor grade, its confidence intervals, standard error, significance (P value), and Wald statistics are included for parameters that were found to be significantly correlated with tumor grade. These included K trans , K ep , and V e . Specifically, K trans was statistically higher for high-grade tumors (median ϭ 0.89, 25th to 75th quartiles ϭ 0.46 -2.67) than for low-grade tumors (median ϭ 0.09, 25th to 75th quartiles ϭ 0.04 -0.13). K ep was statistically higher for high-grade tumors (median ϭ 6.76, 25th to 75th quartiles ϭ 3.77-16.88) than for low-grade tumors (median ϭ 0.66, 25th to 75th quartiles ϭ 0.29 -1.04). V e was statistically lower for high-grade tumors (median ϭ 0.12, 25th to 75th quartiles ϭ 0.11-0.15) than for low-grade tumors (median ϭ 0.23, 25th to 75th quartiles ϭ 0.19 -0.26). Information on the range, sensitivity, and specificity of these parameters is provided in Table 2.

K trans
For low-grade tumors, K trans was in the range of 0.02-0.52 (median ϭ 0.09; 95% CI for the median ϭ 0.06 -0.13). For high-grade tumors, it was in the range of 0.09 -6.19 (median ϭ 0.89; 95% CI ϭ 0.57-1.85). Based on the CI thresholds, there were 14 highgrade and 13 low-grade true-positives, resulting in a 71% (27/38) combined sensitivity of this parameter to detect high-or lowgrade tumors. Individually, the sensitivity of this parameter to detect high-grade tumors was 70% (14/20), and for low-grade tumors, it was 72% (13/18). In addition, there were 2 high-grade tumors with values below the threshold for low-grade. These were considered false-positives for low-grade. There were no low-grade tumors with values above the threshold for high-grade. Consequently, the specificity of this parameter was 100% (18/18) for highgrade tumors and 90% (18/20) for low-grade tumors.

DISCUSSION
Pediatric brain tumors encountered in a clinical setting differ significantly in tumor type from those seen in adults; therefore, predicting tumor grade by using MR imaging in a pediatric clinical setting presents a unique set of issues. While vessel permeability metrics derived from DCE-MR imaging have been associated with tumor grade in adult populations, [24][25][26] such studies in pediatric brain tumors have been lacking.
Dynamic susceptibility contrast perfusion MR imaging has been studied in children by Ho et al 27 to associate tumor grade with maximal relative cerebral blood volume and with the postbolus shape of the enhancement curve. 28 Koob et al 19 used a multiparametric approach to show that the highest grading accuracy was achieved by using a combination of parameters derived from diffusion and DSC perfusion imaging. Yeom et al 29 used arterial spin-labeling to measure perfusion and found that maximal relative tumor blood flow of high-grade tumors was significantly higher than that of low-grade tumors.
Our results suggest that the transfer constants, both K trans and K ep , are significantly distinct between the low-grade and highgrade groups. Several studies have examined the role of K trans and have shown K trans correlates well with tumor grade, particularly in gliomas in adults. [24][25][26][30][31][32] The role of angiogenesis in promoting leakiness of the tumor vasculature and development of new vessels is well-documented, and our findings of increased K trans in higher grade tumors supports that hypothesis. K trans in gliomas has also been shown to be a marker of progression 31,33 in adults.
Our study shows that pediatric low-grade tumors in fact have a higher V e compared with high-grade tumors, contrary to findings in adult tumors showing lower V e in low-grade adult tumors. [24][25][26] In fact, the optimal sensitivity appears to be achieved for V e , with a combined sensitivity of 76% (compared with 71% for K trans and K ep ) and individual sensitivities of 75% and 78%, respectively, for high-and low-grade tumors. The role of V e , which is an indicator of extracellular extravascular space, is poorly understood in the brain tumor literature. Our findings concur with the theory that the higher cellularity in high-grade tumors would lead .34 Note:-T 10 indicates T1 of tissue.  FIG 2. A 3-year-old boy with posterior fossa pilocytic astrocytoma is shown. Axial T2 image shows a T2 hyperintense mass in the vermis, which shows enhancement and increased diffusion. Permeability images show that though there is marked enhancement typical of these tumors, K trans and K ep are considerably lower, whereas V e is higher throughout the tumor compared with the high-grade tumor shown in Fig 1. to a decreased extracellular space due to the closely packed tumor cells, and hence lower V e . As seen in Figs 1 and 2, the areas of decreased V e also correlate with areas of decreased ADC, further confirming our hypothesis. Mills et al 34 however failed to find the expected correlation in a voxelwise analysis between V e and ADC in adult glioblastoma multiformes, possibly due to the confounding effects of the heterogeneous nature of those tumors. All 3 parameters had high specificity, in the range 82%-100%. For low-grade tumors, their specificity was 90%-100%, and for high-grade tumors, the specificity was 82%-100%. K ep had the highest specificity (100%) for both grades.
One of the limitations of this study is that DCE-MR imagingderived pharmacokinetic parameters are heavily dependent on the model and input parameters used 12,22 and are thought to be difficult to standardize. Some of these parameters may not be as critical as previously thought. For example, Larsson et al 35 recently found that there was no significant difference between using T1 derived from a mapping sequence and using a fixed T1 in high-grade gliomas in adults. Because all our subjects were analyzed by using identical model parameters, this finding may not be that critical in this study. Last, the heterogeneity of tumor types and the relatively small sample in this study are also a limitation. Previous studies, however, have investigated smaller samples, so our findings are based on a comparatively larger sample. Nevertheless, this work may be validated in a larger cohort of children with pediatric brain tumors in future studies.

CONCLUSIONS
Dynamic contrast-enhanced perfusion MR imaging is useful in a clinical setting for the differential diagnosis and grading of pediatric brain tumors. Pharmacokinetic parameters such as V e , K trans , and K ep can be used to differentiate low-and high-grade tumors to facilitate treatment planning and determine prognosis and have comparable specificities for tumor grade. In our study, the parameter K ep had the highest specificity for both grades. Of the pharmacokinetic parameters studied, V e offers the highest sensitivity (overall 76%) for determining tumor grade.