Fusion of Deformable and Biomechanical Models for Tracking Left Ventricular Endocardium by Echocardiography

Abstract

Biomedical image processing is a very important research area. Image analysis is one of the most important techniques in studies related to heart functions. The clinical assessment of LV function is very important to evaluate the heart function for patients or suspected heart disease sufferers. 2D echocardiography allows us to study the dynamic analysis of the heart which results in obtaining the quantitative and qualitative analysis of the LV. Cardiac function quantitative analysis depends on the heart’s shape characteristics like the enclosed area and heart wall thickness. The segmentation of medical images and obtaining the traces of the LV boundaries is an essential procedure to get the quantitative and qualitative analysis. Yet, in clinical procedure, this task depends on manual tracing which is slow, tedious and time consuming job. Hence, automating this clinical procedure during the cardiac cycle is of great importance. The aim of this thesis is to automate the manual process of detecting and tracking the LV boundaries of 2D echocardiographic image sequence. Instead of depending only on the imaging based techniques, the designed and implemented framework utilizes the LV mechanics beside the imaging based techniques. When it comes to information extraction from patterns which have been classified, it has been proved that the different contour detection methods complement each other. As a result, efficient combination of different contour detectors is expected to achieve better contour detection than if only one detector is used. This combination of contour detectors produces incremental gains in overall performance. In the first framework, the detection and tracking are accomplished by employing the extended Kalman filter framework to combine the contours estimated by the biomechanical model and the contours extracted using the deformable models. An alternative framework is used by employing averaging fusion followed by level set method. A gold standard is created from three manual outlines and utilized in the experimental results to evaluate the automated results. The tracking and segmentation of LV during the cardiac cycle was accomplished successfully in all cases. The results showed limits of agreement for an average perpendicular distance of 1.277 ±0.252 mm versus the created gold standard. This proved that this framework achieved better performance in tracking and segmenting the LV through the cardiac cycle.