Segmentation of Spin-Echo MRI brain images: a comparison study of Crisp and Fuzzy algorithms

Abstract

This thesis presents a scheme for segmenting Spin-Echo MRI brain images based on Fuzzy C-Mean (FCM) clustering techniques. This scheme consists of feature extraction, feature conditioning or evaluation, and thresholded FCM clustering. Feature extraction involves the inference of a meaningful set of features from MRI images whereas feature conditioning evaluates the selected features in terms of tissue separability. It is shown that the combination of feature conditioning and fuzzy clustering improves the accuracy of the segmentation as compared to the Crisp C-Mean clustering, and K-Nearest Neighbor, and Bayesian Classification techniques. The newly proposed Contextual Thresholded FCM clustering algorithm is shown to provide the best results among the six algorithms compared in this study. The algorithms are evaluated clinically and via phantom volumetric measurement.