Automatic Classification of Brain Tumor by in Vivo MRS Data Based on LDA and SVM

摘要

Recently MRS has been an effective tool for aiding the radiological diagnosis of brain tumor. In this study, our purpose is to evaluate whether we could get a good predictive accuracy by applying different pattern recognition techniques. The classification target is the following four categories: normal tissue, low-grade glioma, high-grade glioma and metastasis. LCModel is used to quantify the in vivo spectra data. The classifiers select different metabolite concentration as input features based on the classification target and statistical analysis result. In general, this study achieves quite good performance for each category. The accurate rate exceeds 95% except for low grade glioma versus high grade glioma, which is hard to distinguish in clinical. The classifier of LS-SVM with an RBF kernel obtains 87.7% accuracy by lipids and lactate as features. Combination MRS with MRI could maybe improve the accuracy.