NIMG-59. EVALUATION OF GLCM FEATURES BASED CLASSIFIERS IN DIFFERENTIATING TUMOR GRADES ON MULTI-MODAL MR IMAGES

Abstract

Abstract BACKGROUND Accurate classification of tumor grade is necessary for understanding tumor development critical in patient management. Radiomic features are gaining popularity in classifying the tumors with the application of various classifiers. We evaluate five classifiers using gray-level co-occurrence matrix (GLCM) features to classify low-grade gliomas (LGG) and high-grade gliomas (HGG). METHODS We included high resolution multi-modal MR images from pre-operative BraTS 2019 database. The database contains a total of 335 multi-modal MR images (259 HGG and 76 LGG) with manually-corrected segmentations of tumor compartments. Sixty four three-dimensional (3D) voxel-by-voxel GLCM feature images on T2-wighted (T2), fluid attenuated inversion recovery (FLAIR), and T1-weighted (T1) pre- and post-contrast images are computed. A total of 192 features within regions of enhanced tumor (ET), edema (ED), and non-enhanced tumor (NET) are obtained by taking averages of the GLCM features within each region. For classification, we evaluated k-nearest neighbor (kNN), learning vector quantization (LVQ), random forest classifier (RF), classification and regression trees (CART), and support vector machine (SVM) classifiers for differentiating LGG from HGG. The dataset is randomly split into ratio of 80 to 20 for training and validation. The models are trained using repeated 5-fold cross-validation. Best 10 models for each of the classifiers are selected based on accuracy by applying multiple random split. RESULTS The average accuracies of 10 best models selected for each of kNN, LVQ, RF, CART, and SVM classifiers are 0.88, 0.92, 1.00, 0.890, and 0.95 on training set, and 0.89, 0.88, 0.88, 0.89, and 0.90 on validation set. The performance of five classifiers on validation set is similar. The accuracy of SVM classifier is slightly higher on validation set even though the RF appears to be the best classifier on training set. CONCLUSION Voxel-by-voxel GLCM features help differentiate LGG and HGG with 0.89 of accuracy irrespective of the classifier.