Texture feature analysis based on gray level co-occurrence matrix for differential diagnosis of glioblastoma and primary central nervous system lymphoma
10.13929/j.issn.1672-8475.2020.04.009
- Author:
Yunfeng SHI
1
Author Information
1. Department of Medical Imaging, Shanxi Medical University
- Publication Type:Journal Article
- Keywords:
Glioblastoma;
Lymphoma;
Magnetic resonance imaging;
Texture analysis
- From:
Chinese Journal of Interventional Imaging and Therapy
2020;17(4):228-232
- CountryChina
- Language:Chinese
-
Abstract:
Objective: To observe the value of texture analysis based on gray level co-occurrence matrix in differential diagnosis of glioblastoma multiform (GBM) and primary central nervous system lymphoma (PCNSL). Methods: Image data of 46 cases of GBM (GBM group) and 36 cases of PCNSL (PCNSL group) confirmed by pathology were retrospectively analyzed. MaZda software was used to manually draw ROI on the maximum level of tumor on enhanced-T1WI and ADC images, and then texture parameters including angular second moment energy (AngScmom), Entropy, Contrast, correlation (Correlat) and inverse difference moment (InvDfMom) were extracted respectively. Multivariate Logistic regression model was constructed for texture feature parameters with statistically significant differences between 2 groups, and ROC curve was used to analyze differential diagnostic efficiency of GBM and PCNSL based on texture parameters and Logistic regression model. Results: There were significant differences of AngScMom, Contrast, Correlat and Entropy on enhanced-T1WI images, also of AngScMom, Correlat and Entropy on ADC images between GBM group and PCNSL group (all P<0.01). Parameters with statistical significances between 2 groups were brought into the binary Logistic regression analysis, and the Logistic regression model was obtained. ROC curve showed that the efficiency of Entropy for identifying GBM and PCNSL was the highest both on enhanced-T1WI and ADC images, AUC was 0.81 and 0.72, the sensitivity was 78.26% and 56.52%, and specificity was 77.78% and 80.56%, respectively. AUC of Logistic regression model for identifying GBM and PCNSL was 0.92, the sensitivity and specificity was 91.30% and 83.33%, respectively. Conclusion: Texture feature based on gray level co-occurrence matrix was helpful for differential diagnosis of GBM and PCNSL.
