Classification of low-grade and high-grade glioma using multi-modal image radiomics features.

Abstract

Gliomas are primary brain tumors arising from glial cells. Gliomas can be classified into different histopathologic grades according to World Health Oraganization (WHO) grading system which represents malignancy. In this paper, we present a method to predict the grades of Gliomas using Radiomics imaging features. MICCAI Brain Tumor Segmentation Challenge (BRATs 2015) training data, its segmentation ground truth and the ground truth labels were used for this work. 45 radiomics features based on histogram, shape and gray-level co-occurrence matrix (GLCM) were extracted from each FLAIR, T1, T1-Contrast, T2 image to quantify the property of Gliomas. Significant features among 180 features were selected through L1-norm regularization (LASSO). Based on LASSO coefficient and selected feature values, we computed a LASSO score and gliomas were classified into low-grade glimoa (LGG) or high-grade glimoa (HGG) through logistic regression. Classification result was validated by a 10-fold cross validation. Our method achieved accuracy of 0.8981, sensitivity of 0.8889, specificity of 0.9074, and area under the curve (AUC) = 0.8870.