Brain Tumor Classification by MRI Images Through Variants of Linear Discriminant Analysis

Abstract

This thesis is based on di erent methods of discriminant analysis applied to brain tumor data. Tumor detection is crucial to improving medical treatment. Magnetic Resonance Imaging (MRI) scans are crucial in several traits and therapeutic applications. For image-based classi cation problems, Linear Discriminant Analysis (LDA) is a potential candidate. In the current article, we have used the LDA variants including Flexible Discriminant Analysis(FDA), Mixture Discriminant Analysis(MDA), Sparse Discriminant Analysis(SDA), and Regularized Discriminant Analysis(RDA) for tumor classi cation based on MRI scans. For this MRI scans were rst compressed with Principal Component Analysis (PCA), moreover PCA helps to remove the outlier samples. It appears the outlier removal slightly increases the brain tumor classi cation ability. Further, the above-mentioned methods have several parameters to tune, which was done by Cross-Validation. The meta-analysis based on 100 Monte-Carlo simulation runs reveals that MDA-PCA and SDA-PCA have signi cantly (p − value ≤ 0.05) better able to classify the brain tumor on test data (82%), while RDA-PCA has worst ability to classify the brain tumor. The ndings indicate the LDA variants can be used not only for brain tumor classi cation but also for image-based other classi cation problems.