A fuzzy clustering framework with neighbourhood constraints and quantitative analysis for segmentation of MRI images

• Main Author: Yu, Lin
• Other Authors: Mohammed Yakoob Siyal
• Format: Theses and Dissertations
• Language: English
• Published: 2010
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Control and instrumentation::Medical electronics
• Online Access: https://hdl.handle.net/10356/42106
• Institution: Nanyang Technological University

The development of computer-aided medical image processing over the past several decades has been truly revolutionary. The advantages of magnetic resonance imaging (MRI) over other diagnostic imaging modalities are its higher spatial resolution and its better discrimination of soft tissue. Automated recognition and diagnosis using MRI data require image enhancement, segmentation and quantification tools. In the area of MR image segmentation, the well-known obstacles include noise, intensity imhomogeneity and partial volume effect. Noise can dramatically degrade the performance of image segmentation algorithms. In order to remove the noise from MR images, a fuzzy filter is used and a modified spatial fuzzy c-means algorithm is developed to overcome the intensity imhomogeneity in MR images. It is demonstrated through experiments that MR brain images can be segmented into three classes: gray matter, white matter, and cerebrospinal-fluid (CSF), while the intensity imhomogeneity is estimated and corrected. Comparisons with similar algorithms have been made. Experimental results show better achievement by the proposed method. In order to overcome the partial volume effect a concept of mixture descriptor is proposed. The MR image voxels are modeled as mixture of more than one tissue. The partial volume labeling process is used to find the most-likely mixtures, which oversegment them to different tissue types. The partial volume labeling method can be implemented with the above noise filter and bias field estimator. Experiments on simulated images and real MR images have been conducted, and results show that the proposed framework works accurately and efficiently. Also comparison results with other methods show our algorithm has better performance. Normally one complete MRI scans include PD, Tl-weighted and T2-weighted data. The multispectral MRI images can provide more information of human tissues. In order to take the advantage of additional information to improve the segmentation, three approaches are proposed and their performances are compared.