Characterization of Brain Glioma in Magnetic Resonance Images using Texture Analysis Techniques

Abstract

This study aimed to characterize brain glioma in magnetic resonance images using image texture analysis techniques in order to recognize the tumor and surrounding tissues by means of textural features. This an analytical case control study was conducted in radiation oncology department at radiation and isotopes center of Khartoum (RICK), which included 100 patients underwent MRI for brain (50 with brain glioma and the rest with normal MRI (case control) scan), FLAIR, T2, T1, and T1 with contrast sequence was performed then the image extracted as DICOM images and then converted to TIFF format which used as input data for an algorithm generated using IDL (interactive data language) for textural features extraction. Three basic textural features types was used to classify the brain images using five different window sizes (3x3, 5x5, 10x10, 15x15, and 20x20 pixels) which are first order statistics (FOS), second order statistics (SGLD), and diagonal features (dSGLD), to recognizes 4 different classes (brain gray and white matter, tumor, background and CSF); further analysis and image segmentations was performed to remove background from the images for purpose of image enhancement. The extracted feature classified using linear discriminant analysis. The result showed that the classification accuracy, sensitivity and specificity according to window sizes was (99.5%, 98.4% and 100%), (98.5%, 95.7% and 100%), (99.1%, 98.8% and 99.3%), (98.1%, 94.3% and 100%), and (96.1%, 90.0% and 98.8%) respectively for brain glioma. This study implies that 3x3 window gives a higher classification accuracy while the most significant features for classification includes; difference average of SGLD, mean and entropy of FOS.