A three-dimensional statistical model for image segmentation and its application to MR brain images

The objective of this thesis is to classify Magnetic Resonance brain images into component tissue types for formation into objects to be used in the accurate 3D reconstruction and visualization of the data. This research work deals specifically with the classification of the tissue types in MR brain images. A 3D statistical model is developed to segment the MR images prior to reconstruction. The algorithm developed makes use of prior knowledge and a probability based model designed to semi-automate the process of segmentation. In addition, a 3D filtering method is used to reduce the noise content of the images prior to segmentation. The eventual 3D visualization provides an aid to the physician in the identification of tissues and the visualization of the sequence. The algorithm was run on several MR 3D sequences. The 3D filtering was shown to be effective in reducing the noise content while minimizing blurring. It compared favorably with mean and median filters, and produced consistently better results. The 3D segmentation algorithm was shown to segment the sequences accurately. Comparisons with k-means and minimum distance classifiers showed the improvements over these algorithms. Finally the entire algorithm with knowledge-base was shown to produce accurate results for the segmentation of the sequences and the reconstruction in a 3D object.