Determination Diagnostic Characteristics of Glioma Grading using Radiomics Analysis of Brain Magnetic Resonance Images

Abstract

Purpose: Grading is one of the challenges in gliomas diagnosis and treatment. This study aims to evaluate the diagnostic performance of machine-learning based radiomics approach in grading of gliomas in an accurate and clinically reproducible way. Methods: This retrospective study was conducted on contrast-enhanced T1 weighted (CE T1W) magnetic resonance (MR) images of 120 glioma patients (60 with grade II and 60 with grade III). Their data were prepared from The Cancer Imaging Archive database. The patients were randomly divided into two groups of training (n=80) and test (n=40). The Gross tumor volume (GTV) regions of interest (ROIs) were manually delineated. Hundred radiomic features were first extracted from the MR images of each patient. A two-step feature selection method was used to select significant features for grading of gliomas. Three linear, quadratic, and medium Gaussian support vector machine (SVM) classifiers were generated. A 10-fold cross-validation was employed to avoid overfitting of classifiers. The classification performance of the models was assessed based on the area under the curve (AUC) of receiver operating characteristic. Results: Grades II and III gliomas were differentiated most appropriately by the linear SVM model with an AUC of 0.900 in the training group. This model could predict the grades of 34 out of 40 patients in the test group (Accuracy=85%). Conclusions: The use of linear SVM model along with the extraction of radiomic features from the GTV ROIs on MR images is an accurate and clinically reproducible method for the grading of gliomas. Keywords: Radiomics, Gliomas, Gross tumor volume, Support vector machine