Tracking and quantification of left ventricle borders in echocardiographic images with improved segmentation techniques

Abstract

Echocardiography is the examination of heart using ultrasound which allows direct visualization of heart in motion. Segmentation of echocardiographic images has a vital role in understanding, analysis and diagnosis of heart diseases. Specifically, tracking the Left Ventricle (LV) borders is of greater clinical importance. The major challenge of this work is to devise robust techniques to address the above said problem. Hence, this thesis proposes new automatic segmentation techniques to track the inner wall of left ventricle (endocardium), which can assist the physician in diagnosing the heart function. The need to track the LV contour is to measure the wall thickness, the ventricular mass and the ejection fraction. Hence, six segmentation techniques have been proposed and implemented in this research work to track the left ventricle borders under three broad categories of image segmentation: texture based, probability based and region based methods. The first method, Transtexture based segmentation technique uses fuzzy logic to resolve the ambiguity in texture variations present between endocardium (inner wall of LV) and blood pool, and between endocardium and myocardium (Thick muscular layer covering the heart) for differentiating the edge and non-edge pixels and produced 89% of acceptance level. In the second work, the novel concept of Fuzzyshed has eliminated the setbacks of conventional morphological watershed by incorporating fuzzy rules to perform segmentation. In the third method, a new Full Causal Hidden Markov Model based segmentation scheme has been proposed by modifying the traditional HMM. It resolves the ambiguity in identifying the edge pixels present in the image under analysis. These proposed techniques were tested and were found to be more robust in producing accurate results when compared with existing algorithms. Amongst the proposed techniques, octagonal tree combined with BFS based fast watershed algorithm produced faster and accurate results. newline newline newline