Diagnostic Accuracy of T1-Weighted Dynamic Contrast-Enhanced–MRI and DWI-ADC for Differentiation of Glioblastoma and Primary CNS Lymphoma

BACKGROUND AND PURPOSE: Glioblastoma and primary CNS lymphoma dictate different neurosurgical strategies; it is critical to distinguish them preoperatively. However, current imaging modalities do not effectively differentiate them. We aimed to examine the use of DWI and T1-weighted dynamic contrast-enhanced–MR imaging as potential discriminative tools. MATERIALS AND METHODS: We retrospectively reviewed 18 patients with primary CNS lymphoma and 36 matched patients with glioblastoma with pretreatment DWI and dynamic contrast-enhanced–MR imaging. VOIs were drawn around the tumor on contrast-enhanced T1WI and FLAIR images; these images were transferred onto coregistered ADC maps to obtain the ADC and onto dynamic contrast-enhanced perfusion maps to obtain the plasma volume and permeability transfer constant. Histogram analysis was performed to determine the mean and relative ADCmean and relative 90th percentile values for plasma volume and the permeability transfer constant. Nonparametric tests were used to assess differences, and receiver operating characteristic analysis was performed for optimal threshold calculations. RESULTS: The enhancing component of primary CNS lymphoma was found to have significantly lower ADCmean (1.1 × 10−3 versus 1.4 × 10−3; P < .001) and relative ADCmean (1.5 versus 1.9; P < .001) and relative 90th percentile values for plasma volume (3.7 versus 5.0; P < .05) than the enhancing component of glioblastoma, but not significantly different relative 90th percentile values for the permeability transfer constant (5.4 versus 4.4; P = .83). The nonenhancing portions of glioblastoma and primary CNS lymphoma did not differ in these parameters. On the basis of receiver operating characteristic analysis, mean ADC provided the best threshold (area under the curve = 0.83) to distinguish primary CNS lymphoma from glioblastoma, which was not improved with normalized ADC or the addition of perfusion parameters. CONCLUSIONS: ADC was superior to dynamic contrast-enhanced–MR imaging perfusion, alone or in combination, in differentiating primary CNS lymphoma from glioblastoma.

T he standard of care for glioblastoma (GBM) dictates maximum safe resection. 1 In contrast, efforts at resection in pri-mary CNS lymphoma (PCNSL) are discouraged due to lack of survival benefits and an increase in postoperative deficits. 2,3 Given the distinct prognostic implications and the differing surgical planning and treatment options for PCNSL and GBM, their preoperative differentiation is important in patients presenting with an enhancing brain tumor. MR imaging features of PCNSL and GBM are highly variable and overlapping, 4,5 rendering the differentiation difficult when based solely on conventional MR imaging.
Several studies have suggested the usefulness of diffusionweighted imaging-derived ADC maps in differentiating PCNSL from GBM. 6,7 Due to its higher cellularity compared with GBM, PCNSL has been shown to exhibit lower ADC values. 6,7 The use of dynamic MR imaging perfusion techniques has also been of growing interest. In pathologic studies, PCNSL and GBM exhibited varying degrees of increased vascular permeability and perfusion. 8,9 T2*-weighted DSC studies have suggested the discriminative value of cerebral blood volume; however, the results have been inconsistent. [10][11][12] More recent studies have suggested the effectiveness of T1-weighted dynamic contrast-enhanced (DCE) perfusion in differentiating PCNSL from GBM. 13,14 DCE-MR imaging measures fractional blood plasma volume per unit volume of tissue (Vp) and time-dependent leakage (permeability transfer constant [K trans ]), which reflect tissue perfusion and leakiness, respectively. 15,16 Compared with DSC, DCE perfusion has the advantages of higher spatial resolution, better quantification of microvascular leakiness and perfusion, and increased resistance to susceptibility artifacts. 16,17 In this study, we aimed to examine the use of DCE-MR imaging and DWI-ADC as potential discriminative tools and to define cutoff (threshold) values for DWI-ADC and DCE perfusion parameters that would be sensitive and specific for PCNSL. We hypothesized that PCNSL would have greater diffusion restriction, while GBM would have greater leakiness and perfusion.

Patient Selection
This study is an institutional review board-approved retrospective single-institution study performed under a waiver of informed consent. All Health Insurance Portability and Accountability Act regulations were followed. We queried institutional and departmental data bases for all patients with histologically confirmed newly diagnosed PCNSL between January 2011 and December 2014 who had pretreated DWI-and DCE-MR imaging scans available for analysis. As part of our hospital routine practice, all histology was verified by 1 of 2 neuropathologists, both of whom had Ͼ8 years of experience in neuropathology. We excluded patients under the following conditions: 1) systemic lymphoma, 2) nonparenchymal PCNSL, 3) having undergone chemotherapy before PCNSL diagnosis, and 4) a known history of testing positive for human immunodeficiency virus. A GBM cohort, matched for age and sex, was selected from an institutional data base of newly diagnosed patients with GBM who had histologic confirmation and preoperative DWI and DCE-MR imaging.
Patient charts were reviewed for demographic characteristics, functional status at initial tumor diagnosis, and clinical outcome data. For patients with PCNSL, we collected serum lactate dehydrogenase results obtained within 1 month of tumor diagnosis; for patients with GBM, we collected the available tumor molecular profile, including O(6)-methylguanin-DNA-methyltransferase (MGMT) methylation, isocitrate dehydrogenase (IDH) mutation, and epidermal growth factor receptor (EGFR) mutation status.

MR Imaging
MR imaging sequences were acquired with a 1.5 or 3T MR imaging scanner (Signa Excite, HDx, and Discovery 750; GE Health-care, Milwaukee, Wisconsin) and a standard 8-channel head coil. Gadopentetate dimeglumine (Magnevist; Bayer HealthCare Pharmaceuticals, Wayne, New Jersey) was injected via a venous catheter (18 -21 ga) at doses based on patient body weight (0.2 mL/kg body weight; maximum, 20 mL) at 2-3 mL/s. DWI was acquired in the transverse plane by using a spin-echo, echo-planar imaging sequence with the following parameters: TR/TE ϭ 8000/ 104.2 ms; diffusion gradient encoding in 3 orthogonal directions; bϭ1000 s/mm 2 ; FOV ϭ 240 mm; matrix size ϭ 128 ϫ 128 pixels; section thickness ϭ 5 mm; section gap ϭ 1 mm; and number of average ϭ 2. ADC values were calculated with the following parameters: ADC ϭ [ln(S / S 0 )] / b, where S is the signal intensity of the ROI obtained through 3 orthogonally oriented DWIs or diffusion trace images, S 0 is the signal intensity of the ROI acquired through reference T2-weighted images, and b is the gradient b factor with a value of 1000 s/mm 2 . ADC maps were calculated on a pixel-by-pixel basis.
DCE-MR imaging of the brain was acquired after DWI scans as part of a standard clinical protocol with an axial 3D T1WI echo-spoiled gradient echo sequence (TR ϭ 4 -5 ms; TE ϭ 1-2 ms; section thickness ϭ 5 mm; flip angle ϭ 25°; FOV ϭ 24 cm; matrix ϭ 256 ϫ 128; temporal resolution ϭ 5-6 seconds; number of sections ϭ 10 -15; and total time ϭ 3.3-4 minutes). Ten phases were acquired preinjection followed by a 30-phase dynamic injection imaging and a 40-mL saline flush. Matching contrastenhanced T1-weighted (TR/TE ϭ 600/8 ms; thickness ϭ 5 mm) images were also obtained, along with standard sequences including T2-weighted images, FLAIR images, and susceptibility-weighted images. T1 mapping provides a method to calculate the T1 value of each voxel during the noncontrast phase and has been shown not to significantly alter DCE quantification. 18,19 Hence, we do not perform T1 mapping for DCE correction at our institution, and it was not available for image processing in this study.

Image Postprocessing and Analysis
ADC maps, DCE-MR imaging perfusion raw data, contrast-enhanced T1WI, and FLAIR images were transferred to an off-line workstation.
Postprocessing DCE-MR imaging perfusion data were processed with FDA-approved commercial software (nordicICE; NordicNeuroLab, Bergen, Norway). The signal-to-noise ratio and arterial input function were optimized individually for each patient. For the arterial input function, an appropriate artery was semiautomatically selected to characterize the input function curve and concentrationtime curve. 20 The linear assumption between change in signal intensity and gadolinium concentration was made to convert the signal intensity curve to a concentration-time curve. Curves showing an optimal relationship between arterial input function and the concentration-time curve were selected. We used the perfusion analysis method based on the 2-compartment pharmacokinetic model proposed by Tofts et al 20 to calculate pharmacokinetic parameters, including Vp and K trans and to display the results as parametric maps.

Image Analysis
Two experienced operators (1 radiology fellow with 3 years and 1 medical student with 1 year of experience) manually outlined a VOI around the enhancing lesion on contrast-enhanced T1WI and the peritumoral nonenhancing lesion on FLAIR images; the 2 operated independently. The VOI was constructed by summing ROIs drawn around the lesion on all axial sections by the 2 operators, and the final VOI was approved by a board-certified neuroradiologist with 10 years of experience in MR imaging and functional imaging. VOIs were transferred to the ADC, Vp, and K trans parametric maps, and the corresponding measurements were re-corded for the all enhancing and nonenhancing lesions. Minimum values of zero pixels were removed. To reduce variability related to scanner heterogeneity, contrast, and patient physiology (eg, cardiac output), we normalized all parameters to normal brain by placing ROIs (standardized area of 40 -60 mm 2 ) in the normal-appearing white matter of the contralateral hemisphere at the midlevel of the tumor. 21-23 The ROI was placed on the contrast-enhanced T1weighted images and then transferred to the VP and K trans maps and adjusted, if necessary, to avoid potential outlier areas that may harbor subtle microvascular leakage. The ADC, Vp, and K trans measurements were binned, and histogram analysis was performed to determine the mean and 90th percentile normalized values for Vp (rVp mean ; rVp 90%tile ) and K trans , and the mean and fifth percentile normalized values for ADC (rADC mean , rADC 5%tile ). The 90th percentile for Vp and K trans characterizes the portion of the tumor with the highest perfusion, while the fifth percentile for ADC determines the portion of the tumor with the greatest degree of restricted diffusion. To facilitate comparisons with prior studies, we also recorded absolute (non-normalized) ADC mean and ADC 5%tile values.

Statistical Analysis
Wilcoxon rank sum tests were conducted to assess differences between GBM and PCNSL groups for normalized ADC, K trans , and Vp parameters. Measurements from enhancing and nonenhancing lesions were analyzed separately. The significance of P values was adjusted by using the false discovery rate approach. Receiver operating characteristic analysis was performed for the parameters with statistically significant differences, and the area under the curve (AUC) was computed. 24 Optimal thresholds were estimated with consideration for sensitivity and specificity. Overall survival was estimated by using the Kaplan-Meier method, starting from the date of tumor diagnosis until death. Patients who did not die during the study period were censored at the date of last available follow-up. Statistical analysis was performed by using R statistical and computing software (http:// www.r-project.org/), including the "ROCR" and "survival" packages. We retrospectively reviewed all patients with pretreated primary CNS lymphoma who underwent evaluation with DCE-MR imaging at our center and selected a GBM cohort matched for age and sex; hence, a sample size calculation was not performed for this study.

Patient Population
We identified 18 patients with PCNSL (11 men; mean age, 68.7 years) and 36 matched patients with GBM (22 men; mean age, 68.6 years) who met all the inclusion criteria. The patient-selection process and clinical data are summarized in On-line Fig 1 and  Table 1, respectively. There are no statistically significantly differences among the 3 AUCs.  In the PCNSL and GBM cohorts, the median Karnofsky performance status at tumor diagnosis was 70 and 90, respectively. At last follow-up, all except 1 (94.4%) patient with PCNSL were alive, with a median follow-up of 22.1 months. One patient died 21.3 months after diagnosis, and the 12month overall survival in the PCNSL cohort was 100%. The 12-month overall survival in the GBM cohort was 48.5% (95% CI, 29.6%-64.9%).

Imaging Findings
The imaging results are summarized in Table 2

ADC
For the enhancing lesions, the median rADC mean was lower for PCNSL than for GBM (PCNSL versus GBM, 1.5 versus 1.9; P Ͻ .001). On the basis of receiver operating characteristic analysis, a rADC mean threshold of Ͻ1.7 indicated PCNSL with a specificity of 78% and sensitivity of 75%. For the nonenhancing lesions, the rADC mean was not significantly different (P ϭ .21). When tested without normalization to a ratio, the median ADC mean of the enhancing lesions was also lower in PCNSL than in GBM (1.1 ϫ 10 Ϫ3 mm 2 /s versus 1.4 ϫ 10 Ϫ3 mm 2 /s; P Ͻ .001); an ADC mean threshold of Ͻ1.3 ϫ 10 Ϫ3 mm 2 /s indicated PCNSL with a specificity of 89% and sensitivity of 69%. There was no significant difference in the AUCs for ADC mean and rADC mean (P ϭ .88).

DCE-MR Imaging
For the enhancing lesions, the median rVp 90%tile was lower for PCNSL than for GBM (3.7 versus 5.0; P ϭ .006), while the median K trans 90%tile was not significantly different (P ϭ .83) between the two. For the nonenhancing lesions, neither rVp 90%tile nor the rK trans 90%tile was significantly different (P Ն .16). If one optimized specificity and sensitivity, a rVp 90%tile threshold of Ͻ4.6 indicated PCNSL with a specificity of 72% and sensitivity of 58%. The AUC for rVp 90%tile was not significantly different from that for ADC mean (P ϭ .87) or rADC mean (P ϭ .66).

Combining ADC and DCE-MR Imaging
If one combined ADC mean and rVp 90%tile at the enhancing lesion, a binary threshold of ADC mean Ͻ 1.3 ϫ 10 Ϫ3 mm 2 /s and rVp 90%tile Ͻ 4.6 predicted PCNSL with a specificity of 61% and a sensitivity of 83%. Conversely, a binary threshold of ADC mean Ն 1.3 ϫ 10 Ϫ3 mm 2 /s and rVp 90%tile Ն 4.6 predicted GBM with a specificity of 94% and sensitivity of 47%.

DISCUSSION
We retrospectively examined the use of ADC and DCE in differentiating pretreated PCNSL from GBM. We found that ADC mean , rADC mean , and rVp 90%tile , but not rK trans in the enhancing regions, distinguished PCNSL and GBM (Table 2). An ADC mean threshold of 1.3 ϫ 10 Ϫ3 mm 2 /s discriminated PCNSL and GBM with the best specificity and sensitivity; a binary threshold that combined ADC mean and rVp 90%tile values was helpful for predicting GBM but less so for PCNSL. Our study findings are consistent with previous studies demonstrating significantly lower ADC in PCNSL than in GBM. 6,7,25,26 In a subset of these studies, DWI results were correlated with histologic information and showed a clear inverse relationship between ADC and tumor cellularity, 6,7,26 suggesting that untreated PCNSL has higher tumor density than untreated GBM. In contrast to these studies, we also studied relative normalized ADC values but did not find any improved discriminative performance with rADC mean over absolute ADC mean. We therefore advocate the use of ADC mean , which is a simpler and more direct measurement than rADC mean , to distinguish PCNSL and GBM.
Our study also examined the use of the DCE-MR imaging technique and showed that rVp was able to distinguish PCNSL from GBM. Vp was not part of the original Tofts model and was introduced later in the modified Tofts model to account for intravascular tracer. 15 In gliomas, Vp has been shown to differentiate tumor grade. 27 However, Vp has not be shown to be useful in differentiating PCNSL from GBM. 14 Our study, to the best of our knowledge, is the first to demonstrate the discriminating value of rVp 90%tile ; this positive finding may be related to our choice of a normalized parameter to reduce interscanner, interrater, and interpatient variability.
K trans is the other commonly measured DCE-MR imaging parameter. It measures the degree of increase in T1 due to contrast accumulation in tissue and can be affected by multiple factors such as blood flow and capillary wall surface area. In general, it is used to represent leakiness due to capillary permeability and blood-brain barrier disruption. 16 A study by Kickingereder et al 13 showed that PCNSL had significantly higher K trans than GBM and further correlated these radiologic findings with histologic demonstration of destroyed vessel architecture in the 11 PCNSLs and intact vascular integrity in the 60 GBMs. With CT perfusion imaging, K trans was also shown to be significantly higher in PCNSL than in GBM in one study 28 but not in a subsequent study 29 ; this latter negative study was attributed to the use of the Patlak model, which fails to account for backflow of contrast agent from the extravascular extracellular space to the blood plasma, compromising K trans measurement. Our study also showed a trend for higher K trans in PCNSL than in GBM, but the difference was not statistically significant.
Our results indicated that rVp is not superior to ADC for identifying PCNSL, and a combined model including rVp 90%tile and ADC mean is not superior to either technique alone. Perfusion MR imaging has been suggested as a valuable part of the imaging strategy for the differential diagnosis of undiagnosed brain masses. 30,31 In addition, perfusion MR imaging has demonstrated additional value in gliomas with correlations to glioma grade, EGFR gene amplification and vIII status, MGMT methylation status, time-to-progression, and overall survival. 21,32-37 Therefore, while not superior for differential diagnosis purposes, DCE-MR imaging may be helpful to support the presumptive diagnosis of PCNSL when marked hyperperfusion is absent, and it can be considered to provide important prognostic data about tumor vascularity and leakiness independent of data about tumor cellularity.
Unlike previous studies that examined imaging parameters solely at contrast-enhancing regions, we evaluated imaging parameters at both enhancing and nonenhancing regions but failed to show any added discriminative value. This outcome suggests that increases in tumor cellularity, microvascular permeability, and vascular proliferation are less marked in nonenhancing areas of different brain tumors compared with contrast-enhancing regions. Although nonenhancing regions in glioblastomas are known to include a combination of tumor and peritumoral edema, 38 quantification of tumor-related imaging characteristics is dependent on the relative abundance of tumor cells and may be diluted by the amount of edema. 26 Our study has several potential limitations. First, we did not examine all DCE-MR imaging parameters, including extravascular extracellular volume (Ve), which was shown in 1 study to differentiate PCNSL from other brain tumors. 14 This study by Abe et al 14 used a DCE-MR imaging sequence with a shorterthan-standard acquisition time, however, which potentially overestimates K trans and underestimates Vp and Ve, 39 thereby limiting the generalizability of the results. In addition, the physiologic meaning of Ve remains elusive, with conflicting studies demonstrating its correlation with tumor cellularity. 7,26,40 The use of Ve is further refuted by an earlier study that failed to show a significant difference in Ve between PCNSL and GBM. 13 We did not evaluate Ve in our study and, instead, chose ADC as a more reliable and widely applied marker of tumor cellularity. 7,26 Second, we did not evaluate the performance of other conventional MR imaging parameters such as the presence of necrosis and rim enhancement. Previous studies have shown that PCNSL has variable degrees of T1 and T2 intensities, extent of necrosis, and pattern of contrast enhancement-suggesting that these features are unreliable in distinguishing lymphoma from other CNS lesions. 41,42 Consequently, in 1 prior study, the use of rim enhancement and central necrosis resulted in misclassification of 3 (17.6%) of 17 PCNSLs and malignant gliomas. 30 Given the statistical limitations of our relatively small sample size of patients with PCNSL who underwent DCE-MR imaging, we chose to focus our study on examining ADC, rather than all the other conventional MR imaging parameters. Our small sample size also limited our ability to use cross-validation to assess the operating characteristics and validate the sensitivities and specificities of the cutoff values determined in our study. Third, we did not account for the presence of microbleed, which may confound ADC and DCE-MR imaging results. However, the additional step of eliminating microbleed by using an additional SWI sequence may pose impractical constraints and limit the applicability of our study. Fourth, due to the lack of the routine use of MR spectroscopy in this retrospective study, we were unable to evaluate the utility of MR spectroscopy, which has been shown to distinguish PCNSL from GBM in 1 study with a specificity exceeding 90%. 43 The added value of MR spectroscopy in discriminating the diagnoses should be incorporated in future prospective studies. Fifth, the manual drawing and transfer of VOIs to parametric maps could have introduced variability. We sought to reduce the variability by having all VOIs drawn by operators who had at least 1 year of experience and reviewed by an experienced board-certified neuroradiologist and by exploiting the histogram function to interrogate only the most abnormal parts of the tumor.

CONCLUSIONS
Pretreatment differentiation of PCNSL and GBM is challenging. Our study suggests that DCE-MR imaging is helpful in identifying PCNSL, though rVp did not outperform ADC and a combined model did not outperform either metric alone. Further prospective studies are needed to confirm these findings.