Detection and validation of segmentation techniques for MR brain tumor of glioma patients

Abstract

Prediction of brain tumor becomes difficult with respect to the irregular shape, growth, location and volume of the tumor, thus segmentation is highly required for the proper detection of the tumor. Four sequences of MR images like T1, T2, T1 contrast and fluid attenuation inversion recovery (FLAIR) is collected from the BRATS 2015 dataset for the validation of segmentation techniques. In this paper two segmentation techniques like semi-automated active contour and fully-automated expectation maximization (EM) are discussed as both are widely used in the field of brain tumor analysis. The synthetic data obtained is skull stripped and noise free reducing the process time for detecting the tumor. The main objective is to extract the region of interest, validate and improve the accuracy, dice coefficient of the synthetic dataset with the ground truth available. Active contour is an iterative process with the initial contour selected manually and EM works on the probability of the intensity values. The result shows some of the images works better with active contour and some with EM. Time taken is less for EM compared to active contour. Accuracy, dice coefficient, sensitivity is better in EM compared to Active contour. Statistical features and textural features extracted from the above techniques plays vital role for the accurate diagnosis of the tumor. In this context segmentation is vital to further classify images into low grade and high grade glioma’s helping radiologists to accurately diagnose the abnormal tissue growth with effective planning of treatment.