A reliable ensemble-based classification framework for glioma brain tumor segmentation

Abstract

Glioma is one of the most frequent primary brain tumors in adults that arise from glial cells. Automatic and accurate segmentation of glioma is critical for detecting all parts of tumor and its surrounding tissues in cancer detection and surgical planning. In this paper, we present a reliable classification framework for detection and segmentation of abnormal tissues including brain glioma tumor portions such as edemas and tumor core. This framework learns weighted features extracted from the 3D cubic neighborhoods regarding to gray-level differences that indicate the spatial relationships among voxels. In addition to intensity values in each slice, we consider sets of three consecutive slices to extract information from 3D neighbors. Classification procedure is defined based on combination of support vector machines regarding to an ensemble learning method. In order to regularize and improve the output of the classifier framework, we design a post-process procedure based on statistical concepts. The proposed framework is trained and tested with BRATS datasets, and comparative analysis is implemented. Experimental results indicate competitive performance compared to the state-of-the-art methods. The achieved accuracy is characterized by the overall mean Dice index of 88%.