MORPHOLOGIC ANALYSIS OF BRAIN TUMOR TO PERFORM SURVIVAL PREDICTION WITH A FOCUS ON ECCENTRICITY WITH SINGULAR VALUE DECOMPOSITION

Abstract

Advancments in imaging field have evolved enough to make the detection of tumor task more accurate through 3D MRI but at the same time more time consuming and complex for medical experts. Therefore, the need for a computational logic unit which never fails, process the information fast and never gets tired arises. This thesis will cover a whole mechanism of brain tumor severity determination starting from segmentation process till evaluation of eccentricity and volume. Segmentation step is performed with a 3D U-net whith some tweaked hyperparameters such as dropout values and learning rates to achieve better performance for the segmented parts. The accuracy of segmentation is reported to be 99%. Eccentricity and volume are measured over the segmented region. Estimation of eccentricity is calculated based on the energy values from the decomposition of segmented tumor in SVD where the sigma, or the energy matrix holds the values of which their ratio combined gives the eccentricity value. The dataset is part of the BRATS challenge 2020.