MRI-based brain tumor detection using the fusion of histogram oriented gradients and neural features

Abstract

Computer-aided diagnoses are playing a remarkable role in analyzing the MRI images and thus assist the radiologist. The brain tumor is one of the most common and incursive diseases, leading to a very short life expectancy in their highest grade. Early detection of the brain tumor and its proper diagnosis may increase the survival rate of a patient. This paper has developed an intelligent system based on the fusion of histogram of oriented gradient and deep CNN-based neural features from MRI in the identification of tumors. This research has performed Principle Component Analysis as a feature optimization technique to attain more intuitive features from the fused feature vector. A machine learning-based ensemble classifier based on bagging and boosting known as Bootstrap Aggregation is then used on the optimized fusion vector to detect the tumor. An extensive experiment is performed using four types of MRI sequences T1, T2, T1c, and Flair from two publicly available datasets BRATS 2013 and BRATS 2015. This fusion-based method gives a satisfactory performance with a detection accuracy of 98.79% using a fivefold cross-validation technique.