Binary glioma grading framework employing locality preserving projections and Gaussian radial basis function support vector machine

Gliomas are one of the most dangerous brain tumours, which are life-threatening in their highest grade. Manual diagnosis and detection of brain tumours are time-consuming. Thus, early and accurate detection of such tumours becomes essential for its prognosis. In this work, we aim to develop a fully automated glioma grade classification framework. The proposed method aims to learn relevant and non-redundant features from magnetic resonance brain tumour images. Gabor filter banks extract discriminant features from the brain images, and locality preserving projections are utilized for feature reduction to achieve non-redundant features. We train Gaussian radial basis function-support vector machine classifiers using these features for classifying gliomas into low grade glioma (LGG) or high-grade glioma. We apply Synthetic Minority Over-Sampling Technique on the minority class (LGG) to mitigate the class imbalance problem. The performance of the proposed method is validated by conducting fivefold cross-validation. We evaluate the performance of our proposed framework on different BRATS datasets: BRATS 2013, BRATS 2015, and BRATS 2017. We perform extensive experiments on each dataset separately, and experimental findings demonstrate that our proposed approach provides significantly superior performance compared to state-of-the-art techniques.