Differentiation between Primary Cerebral Lymphoma and Glioblastoma Using the Apparent Diffusion Coefficient: Comparison of Three Different ROI Methods

Objective Apparent diffusion coefficients (ADC) can help differentiate between central nervous system (CNS) lymphoma and Glioblastoma (GBM). However, overlap between ADCs for GBM and lymphoma have been reported because of various region of interest (ROI) methods. Our aim is to explore ROI method to provide the most reproducible results for differentiation. Materials and Methods We studied 25 CNS lymphomas and 62 GBMs with three ROI methods: (1) ROI1, whole tumor volume; (2) ROI2, multiple ROIs; and (3) ROI3, a single ROI. Interobserver variability of two readers for each method was analyzed by intraclass correlation(ICC). ADCs were compared between GBM and lymphoma, using two-sample t-test. The discriminative ability was determined by ROC analysis. Results ADCs from ROI1 showed most reproducible results (ICC >0.9). For ROI1, ADCmean for lymphoma showed significantly lower values than GBM (p = 0.03). The optimal cut-off value was 0.98×10−3 mm2/s with 85% sensitivity and 90% specificity. For ROI2, ADCmin for lymphoma was significantly lower than GBM (p = 0.02). The cut-off value was 0.69×10−3 mm2/s with 87% sensitivity and 88% specificity. Conclusion ADC values were significantly dependent on ROI method. ADCs from the whole tumor volume had the most reproducible results. ADCmean from the whole tumor volume may aid in differentiating between lymphoma and GBM. However, multi-modal imaging approaches are recommended than ADC alone for differentiation.


Introduction
Glioblastoma (GBM) is the most common malignant brain tumor in adults. GBM is marked by rapid growth [1]. Primary central nervous system(CNS) lymphoma is less common than GBM but its incidence is increasing [2]. For GBM, surgical resection is the primary treatment [3], while chemotherapy or radiation therapy is the treatment of choice for CNS lymphoma [4]. Therefore, an exact differential diagnosis is essential for making therapeutic decisions about GBM and CNS lymphoma. On conventional imaging, primary CNS lymphomas usually show homogenous and intense contrast enhancement. And primary CNS lymphomas are often hypointense to gray matter without large necrosis on T2-weighted image(T2WI) [5]. However, differentiation is often difficult because some of GBMs have considerable overlap in conventional magnetic resonance(MR) imaging findings [6].
Several studies have shown that apparent diffusion coefficient (ADC) values from diffusion-weighted imaging(DWI) can help differentiate between CNS lymphoma and GBM [7][8][9]. However, other studies have reported that ADC might not be helpful because of substantial overlap between values for CNS lymphoma and GBM [10,11].
These contradictory results are partly because ADC can be measured by a variety of methods to determine placement of the region of interest (ROI). Toh et al [9] drew the ROI in the center of the solid enhancing region Yamashita et al [12] and Doskaliyev et al [13] drew several small ROIs within the tumor. This might contribute to the wide variety in reported ADC results. Thus, it is necessary to evaluate the reliability of commonly used ROI methods in DWI. The purpose of this study was to compare whole tumor volume ROI, multiple ROIs and single ROI for ADC measurement for differentiating between primary CNS lymphoma and GBM.

Patients
Approval by Severance hospital institutional review board was obtained and informed consent was waived for this retrospective study. Patients' records and information were anonymized and deidentified prior to analysis. MR imaging of consecutive patients from Oct 2012 through Nov 2013 were retrospectively analyzed. We identified 30 immunocompetent patients with biopsy-proven primary CNS lymphoma. We excluded the 5 patients with primary CNS lymphoma because they received the steroid therapy before they performed MR imaging. Finally, 25 patients with primary CNS lymphoma (15 women, 10 men; mean age, 60 years; age range, 44-77 years) were included. We identified 62 patients (28 women and 34 men;mean age, 56.72 years; age range, 32-73 years) with histologically-confirmed, World Health Organization grade IV GBM in our medical record.

MR imaging
All images were obtained using a 3.0T MRI scanner (Achieva, Philips Medical system, Best, Netherlands) with a 16-channel sensitivity encoding (SENSE) head coil. Diffusion weighted image(DWI) was performed using a single-shot spin-echo (SE) echo planar sequence with following parameters: Echo time(TR)/ Repetition time(TE) = 8413/77 ms, 90u flip angle, 70 transverse sections, SENSE factor = 2, slice thickness = 2 mm, 1126112 matrix, field of view(FOV) = 220 mm. Diffusion-sensitizing gradients were applied sequentially in the x, y and z directions with b factors of 0 and 1000 s/mm 2 . ADCs were automatically calculated by the operating console of the MR scanner and displayed as corresponding ADC maps.

Image analysis
The size and location of tumor was recorded by the study coordinator. If there were multiple lesions, the largest one was measured. Three different ADC measurements for one lesion were obtained from the ADC map according to three distinct ROI protocols: (1) whole tumor volume; (2) multiple ROIs and (3) single ROI. For whole tumor volume, using the coregistration module integrated in the commercial software nordicICE (Nordic Imaging Lab, Bergen, Norway), ADC maps were coregistered to postcontrast T1-weighted 3D GRE image by the study coordinator. Two readers (a neuroradiologist with 5 years of experience and a neuroradiologist with 14 years of experience) independently drew freehand ROIs along tumor borders on coregistered images to cover tumors completely with consecutive slices. Minimum, maximum and mean value (min, max and mean) were calculated from ADC values from the whole tumor volume. For multiple ROIs, two readers independently drew circular 5 ROIs (area = 10 mm 2 ) on enhancing lesions in coregistered ADC map. For single ROI method, the readers reviewed the coregistered ADC maps and drew a single circular ROI (area = 20 mm 2 ) on any enhancing portion. Hemorrhage, cyst and necrosis were avoided when drawing all three ROI methods (Fig. 1). Min, max and mean were calculated as above.

Statistical analysis
Statistical analyses were performed using SPSS version 16.0 (SPSS Inc., Chicago, IL, USA). Interobserver variability of the readers for different ROI methods was calculated as intraclass correlation(ICC) coefficient (0.00-0.20 poor, 0.21-0.40 fair, 0.41-0.60 moderate, 0.61-0.80 good and 0.81-1.00 excellent correlation). ADCs were averaged between the two observers for further analysis. ADC min , ADC max and ADC mean were compared between GBM and lymphoma using a two-sample t-test for each individual ROI method. Sensitivity, specificity and accuracy for the discriminating between GBM and lymphoma were calculated for each parameter using an optimal cut-off value determined by receiver operating characteristic (ROC) analysis. Area-under-the-ROC curve (AUC) values for discrimination were calculated for the four parameters. P-values,0.05 were considered statistically significant.

Interobserver variability
Intraclass correlation coefficients between two readers for three ROI methods are in Table 1. ADC min , ADC max , ADC mean from

ROC analysis
ADC variables from three different ROI methods were evaluated for discriminative ability using ROC analysis (Table 3). ADC mean calculated from ROI 1 was a significant predictor for differentiating lymphoma from GBM (p = 0.03). The optimal cutoff value was 0.98610 23 mm 2 /s (sensitivity: 85%; specificity: 90%; AUC, 0.87). In ROI 2 , ADC min was a significant predictor for differentiating lymphoma from GBM (p = 0.02). The optimal cutoff value was 0.72610 23 mm 2 /s (sensitivity: 87%; specificity: 65%; accuracy: 0.84). Other variables from the three different ROI methods did not show significant discriminative ability (p. 0.05).

Discussion
Previous studies have used various ROI methods to measure ADC values for differentiating between lymphoma and GBM [7][8][9][10]. Toh et al [9] drew a single ROI in the center of solid enhancing region and Yamashita et al [12] and Doskaliyev et al [13] drew several small ROIs. Kang et al [14] used the whole tumor volume ROI. These various ROI methods may account for previous inconsistent results. However, there has been no study comparing the reproducibility of various ROI selections. According to our results, interobserver reproducibility of ADC calculations was dependent on the selected ROI method. ADC measurements from the whole tumor volume (ROI 1 ) were most reproducible followed by multiple ROIs, then by the single ROI method. Several studies reported that quantitative measurement from the whole tumor volume is the most reproducible, although their subjects was not the brain [15][16][17]. Our results suggested that the whole tumor volume ROI method is favored, and single ROI method should be avoided when measuring ADC values. A single ROI method can be subjective and prone to a sampling bias [18]. We found that the ADC mean from the whole tumor volume was significantly lower for lymphoma than for GBM. Meanwhile, ADC mean from multiple ROIs or a single ROI was not significantly different between lymphoma and GBM. It is well known that GBM may have heterogeneous histologic features. Although we draw ROIs avoiding large necrosis, GBM may have microscopic necrosis with surrounding clustered nuclei, so called ''pseudopalisading'' features, which may increase the overall ADC mean [19,20]. These features make it easier to differentiate between lymphoma and GBM. On the contrary, an ADC from multiple ROIs or a single ROI may not reflect heterogeneity of GBM [21].
Also of note was that ADC min from the whole tumor volume was not a significant predictor but ADC min from multiple ROIs was a significant predictor for differentiating between lymphoma and GBM. ADC min has been suggested to reflect the highest tumor cell density or the most proliferative portion of a tumor within heterogeneous tumors. ADC min from whole tumor volumes might be influenced by the susceptibility of MR to generate artifacts from blood products and might not represent true ADC min of the tumor parenchyma.
However, our results should be carefully interpreted, because the ranges of ADCs between lymphoma and GBM still substantially overlapped (Fig. 2) and ADC alone might not be sufficient to differentiate lymphoma from GBM. Other advanced imaging techniques such as dynamic contrast-enhanced MRI (DCE), dynamic susceptibility-weighted imaging (DSC), susceptibility-weighted imaging (SWI) and FDG-PET have been reported to improve differential diagnosis of lymphoma and GBM [22][23][24][25]. Kickingereder et al [26] reported multimodal imaging integrating these advanced sequences allowed reliable differentiation of lymphoma and GBM. Therefore, Multiple advanced imaging techniques in conjunction with ADC should be preferred than ADC alone when differentiating lymphoma from GBM.
Our study has limitations. First, selection bias was ineluctable in this study because only the patients who had pathologically proven lymphoma and GBM were enrolled. Second, it was difficult to spatially co-localize pathology with MR images. Therefore,  interpreting the pathological meaning of ADC from each ROI was difficult. Third, the number of cases was not enough to draw a solid conclusion. Fourth, we did not perform ADC histogram analysis and the distribution of ADCs were not assessed.
In conclusion, ADC values were significantly dependent on ROI method. ADCs from the whole tumor volume had the most reproducible results. ADC mean from the whole tumor volume may aid in differentiating between lymphoma and GBM. However, multi-modal imaging approaches are recommended than ADC alone for the differentiation.

Author Contributions
Conceived and designed the experiments: SKL. Performed the experiments: SJA HJS. Analyzed the data: SJA HJS. Contributed reagents/ materials/analysis tools: SJA HJS JHC. Wrote the paper: SJA.