Non-invasive diagnosis of brain tumor grade

Recent technological developments in imaging tools that are used in diagnostic radiology provide a wide spectrum of information about tumor tissues. Imaging techniques of brain tumors aim to determine the localization, extend, type and tumor grade. Conventional Magnetic Resonance Imaging (MRI) techniques are the most preferred procedure to diagnose brain tumors. However, advanced MRI techniques also provide additional information about brain tumors. Particularly, researchers have shown that a strong relationship exists between the information acquired using advanced MRI techniques and the grade of the tumor. The grade of the tumor plays a central role in surgery and treatment planning. The conventional procedure for grading the tumor is histopathological biopsy. In addition to being an invasive technique, biopsy also suffers from the disadvantage that the histological samples obtained are subject to a certain sampling error. Thus, a non-invasive method for determining the tumor grade by using several imaging techniques may be beneficial in certain situations. In this thesis, we compared the performance of Logistic Regression(LR), Back Propagation (BP) and Self Organizing Maps (SOM) in the prediction of the tumors' grade preoperatively using the parameters of advanced MRI techniques, namely Diffusion Weighted Imaging (DWI), Magnetic Resonance Spectroscopy (MRS), Perfusion-weighted Magnetic Resonance Imaging (pMRI) and Diffusion Tensor Imaging (DTI). We concluded that the results have evidenced the complex nature of the data. Compared to the other two methods, BP training algorithm with Bayesian Regularization gives the best classification for fourteen out of fifteen data sets. We seperated models as validated models that provide better classification than the null model. BP algorithm classified better for 9 out of 13 models. The correct classification of brain tumor grade in validated models have a range between %67.50 and %100.00.