INTRODUCTION

Primary central nervous system lymphomas (PCNSLs) are sometimes difficult to distinguish from glioblastomas (GBMs) based on routine magnetic resonance (MR) examination. To date, the apparent diffusion coefficient (ADC) and different patterns of contrast enhancement have been valuable for the differential diagnosis of PCNSLs and GBMs. ADC is based on a quantitative value, whereas contrast-enhanced T1-weighted images (CET1WI) are based on appearance instead of a quantitative value. Therefore, we aimed to determine whether statistics (kurtosis, skewness, entropy, and angular second moment) can be derived from histograms on signal intensity distribution in CET1WIs of brain tumors using a commercial software. In addition, we aimed to evaluate the utility of histogram analysis of intratumoral contrast-enhanced regions using CET1WIs of PCNSLs and GBMs to determine whether the abovementioned statistics differed between these tumors.

METHODS

CET1WI of 20 patients with PCNSLs (n = 10) or GBMs (n = 10) were obtained using a 3D-spoiled gradient recalled acquisition in the steady state (SPGR) sequence. All tumors were pathologically confirmed. Further, MR imaging was performed using a 3.0-T whole body scanner (Discovery MR750, GE Healthcare, Waukesha, WI, USA) and a 32-channel head coil. Imaging parameters were as follows: FOV = 240 mm, matrix size = 420 × 192, slice thickness = 1.4 mm, TR/TE/TI = 7.2/2.2/700 ms, and number of slices = 192–270. A standard dose of contrast agent (0.1 mmol/kg) was injected. The region surrounding the outermost layer of the tumors was removed using a workstation (ZIOSTATION; Amin Co., Ltd., Tokyo, Japan) to reveal the contrast-enhanced areas. Moreover, signal intensity distribution was measured within the tumors. Commercial software (Microsoft Excel; Microsoft Corporation, Redmond, WA, USA) was used to calculate statistics, and Wilcoxon rank sum test was used to estimate differences in terms of statistics (skewness, kurtosis, entropy, and angular second moment) between PCNSLs and GBMs. To differentiate between these tumors using statistics, area under the curve (AUC) and cut-off values were determined using receiver operating characteristic (ROC) curve analysis. Sensitivity, specificity, and accuracy were calculated. The AUC value was calculated using logistic regression analysis of the judgments of the two radiologists. DeLong’s test was used to examine significant differences between the AUC values of the statistics individually and those of the radiologists’ judgments. Using all combinations of different statistics, we calculated the AUC values using multivariate logistic regression analysis. We calculated the differences between the AUC values of combined statistics and those of the radiologists’ judgments. Further, we used the Akaike information criteria (AIC) to determine the optimal statistics combination for the differential diagnosis of the two tumors. Because kurtosis, entropy, and angular second moment were positive for the four variables, we performed logarithmic transformation of these statistics and calculated the AUC values for all combinations using multivariate logistic regression analysis. Similarly, we estimated the optimal statistics combination with the lowest AIC values. We performed DeLong’s test using a one-sided test between the AUC values of the optimal statistics combination and that of the judgment of each radiologist. The difference was found to be significant (p < 0.05).

RESULTS

Univariate analysis
There were significant differences in terms of skewness, entropy, and angular second moment between PCNSLs and GBMs (Figure 1; p < 0.01) but not in terms of kurtosis (p = 0.579). From the results of the ROC curve analysis, we could differentiate between the two tumors based on skewness, entropy, and angular second moment with cut-off values of −0.034, 4.474, and 0.018, respectively, as well as accuracies of 90%, 85%, and 95%, respectively. The AUC values of skewness, entropy, and angular second moment were 0.900 (95% confidence interval [CI]: 0.754–1.00), 0.890 (95% CI: 0.736–1.00), and 0.930 (95% CI: 0.790–1.00), respectively (Table 1). Additionally, the AUC values of the radiologists’ judgments were 0.750 (95% CI: 0.562–0.938) and 0.800 (95% CI: 0.615–0.985) (Table 2). Notably, there were no significant differences between the AUC values of each statistics and that of the radiologists’ judgments.

Multivariate analysis
Angular second moment was used to determine the optimal statistics combination. However, despite combining the statistics, a significant difference was not observed between the AUC values of the optimal statistics combination and that of the radiologists’ judgments. Furthermore, following logarithmic transformation of the statistics, the skewness–angular second moment combination was found to be the optimal combination. There were significant differences between the AUC values of this combination (0.95 [95% CI: 0.847–1]) (Table 2) and that of each radiologist’s judgment (p = 0.0496 and 0.0388, respectively).

DISCUSSION AND CONCLUSION

In this study, the commercial software was used to derive statistics (skewness, kurtosis, entropy, and angular second moment) from histogram analysis. Skewness, entropy, and angular second moment determined from the histograms of CET1WI signal intensities are useful indices of measurement for the differential diagnosis of PCNSLs and GBMs. The optimal combination for the differential diagnosis was found to be of skewness and angular second moment.

Acknowledgements

No acknowledgement found.

References

1. Lubner MG, Smith AD, Sandrasegaran K, Sahani DV, Pickhardt PJ. CT texture analysis: definitions, applications, biologic correlates, and challenges. Radiographics 2017; 37:1483–503.