An unsupervised learning with feature approach for brain tumor segmentation using magnetic resonance imaging

Abstract

Segmentation methods are so much efficient to segment complex tumor from challenging datasets. MACCAI BRATS 2013-2017 brain tumor dataset (FLAIR, T2) had been taken for high grade glioma (HGG). This data set is challenging to segment tumor due to homogenous intensity and difficult to separate tumor boundary from other normal tissues, so our goal is to segment tumor from mixed intensities. It can be accomplished step by step. Therefore image maximum and minimum intensities has been adjusted because need to highlight the tumor portion then thresholding perform to localize the tumor region, has applied statistical features(kurtosis, skewness, mean and variance) so tumor portion become more visualize but cann't separate tumor from boundary and then apply unsupervised clusters like kmean but it gives hard crisp membership and many tumor membership missed so texture features(Correlation, energy, homogeneity and contrast) with combination of Gabor filter has been applied but dimension of data increase and intensities became disturb due high dimension operation over MRI. Tumor boundary become more visualize if combine FLAIR over T2 sequence image then we apply FCM and result is: tumor boundaries become more visualized then applied one statistical feature (Kurtosis) and one texture feature(Energy) so tumor portion separate from other tissue and better segmentation accuracy have been checked with comparison parameters like dice overlap and Jaccard index.