Brain Tumor Classification with Fisher Vector and Linear Classifier for T1-Weighted Contrast-Enhanced MRI Images

Abstract

This paper presents the development of a computational method for classifying three types of brain tumors - i.e. meningioma, glioma and pituitary - from T1-weighted contrast-enhanced MRI images. The proposed method performs feature extraction on a specified set of tumor pixel intensity and uses the extracted information to determine the corresponding type of brain tumor. In feature extraction, the specified tumor area was first augmented to incorporate the sample of the surrounding tissue, prior to intensity extraction with dense local patches. Afterwards, the extracted intensity from each patch was fitted to a Gaussian Mixture Model (GMM) and processed into Fisher Vector representation. Furthermore, we applied four linear classifiers to the Fisher Vector representation and evaluated their classification performance. Our experiments showed that the logistic regression gave the best performance with average accuracy, sensitivity and specificity of 89.9%, 95.2%, and 89.0% respectively.