Information set-based texture features to classify brain tumors

An early detection of patients with brain tumors from the Magnetic Resonance (MR) images is the problem that attempts to improve their survival rates. To this end, the proposed Gradient Local Pattern (GLP) method generates both gradient and local pattern images, representing the high- and low-frequency texture components, respectively, by applying Sobel and Local Binary Pattern (LBP) masks to the windows of an MRI. To account for the uncertainty in the texture components, the possibilistic Tsallis (PT) entropy function is derived from its probabilistic counterpart. This function paves the way for the generation of three types of Tsallis-based power form of information values, viz., Energy, Sigmoid, and log from texture components separately. In contrast, the products of pixel intensities and their membership function values give the product form of information values, and the Hanman filter converts these values into low and high-frequency information set texture components, whereas the Gabor filter directly derives these texture components from the pixel intensities using the complex Gabor function. We have two types of information set texture features, with the first being Gabor-based and Hanman-based magnitude and fused features, and the second being the Tsallis-based fused Energy, Sigmoid, and log features. It has been found by experiments that the Tsallis-based texture features outperform the Gabor-based and Hanman-based magnitude and fused features when used with the Hanman-Shannon Transform Classifier that gives the accuracies of 99.06%, 99.14%, and 98.92% in the classification of the 2-class, 3-class, and 4-class brain MRIs, respectively.