A Study on Brain Tumor Segmentation in Noisy Magnetic Resonance Images

Abstract

AbstractPerformance of brain tumor segmentation from Magnetic Resonance (MR) images has been significantly improved in the last decade by various advanced machine learning techniques. However, it remains challenging when MR images suffer from bias field and various types of noise, while acquisition process. Noise typically gets introduced in MR images because of uneven measurements and selection of several statistical factors of radio frequency pulse sequences such as T1 and T2 relaxation times, repetition time, time to echo, etc. In this work, we present a study of brain tumor segmentation on noisy MR images by adding Gaussian noise on different brain MR modalities. Random Forest (RF) and Support Vector Machine (SVM) classifiers are utilized to segment various labels of brain tumor, viz., complete brain tumor region, brain tumor core region, and enhancing tumor region. Classifiers are trained after applying appropriate noise removal filtering algorithms such as Median Filter and Wiener Filter, followed by extracting relevant statistical and texture image features. The proposed work is evaluated on the real glioma MR images of benchmark BRATS dataset. Experimental results demonstrate favorable segmentation performance, i.e., 95% for segmenting complete brain tumor region in terms of Dice Similarity Coefficient (DSC).KeywordsBrain tumor segmentationMagnetic resonance imagingNoise removal