Novel Framework for Predictive Analytics of Brain Tumor Segmentation Using Recurrent Neural Network

Abstract

Background and Objective: Brain tumor is the fast growing illness in both emerging and developed countries. Brain tumor develops as the number of malignant cells in the brain increases. This continued development of cells will result in the infected person's death. The mortality can be reduced by the early prediction and diagnosis of brain tumor at the initial stage. Early detection of brain tumor is analyzed using the Magnetic Resonance Imaging (MRI), which would increase the survival rate of the infected person. This imaging technique will determine the brain tumor with the clear markings which results in detection of tumor at an early stage. Methods: In this proposed work, Weiner filter with the various bands are used for denoising and enhancing the input MRI image slices. MRI images include n number of pixels, which are represented as subsets are determined using potential field clustering method. The segmentation of tumor is carried out by using the global threshold and different morphological operations in the division of Fluid Attenuated Inversion Recovery (Flair) and T2 MRI. The segmented region consists of various features which are extracted by using the fusion of Local Binary Pattern (LBP) and Gabor Wavelet Transform (GWT). The extracted feature undergoes feature selection process which is carried out using genetic algorithm (GA). The selected features undergo the classification process using Recurrent Neural Network (RNN). Results: The implemented work yields the peak signal to noise ratio (PSNR), mean squared error (MSE) and structured similarity index (SSIM) which is approximately ranges 77.39,0.026 and 0.87 on T2 whereas in Flair 77.3,0.025,0.85. The segmentation process which is carried out using fusion technique includes fused features as well as individual features based on the pixel value. The pixel level comparison is carried out using the slices obtained from the input images which are represented as foreground (FG) pixels, background (BG) pixels, errorregion (ER) and pixel quality (Q). BRATS database is used in the proposed work which consists of collection of T2 stage and Flair as well as Normal brain images. The fused features are selected using genetic algorithm which is given as an input for the classification process. The classifiers which is implemented is the Recurrent Neural Network (RNN) consists of various hidden layers that can achieve the accuracy, sensitivity, specificity, area under curve (AUC) in the range of 0.97,0.93,0.99,0.96 on BRATS dataset.