Ultrasound cardiac modeling, segmentation and tracking
MetadataVis full innførsel
Echocardiography plays a key role in assessing cardiac diseases, especially in evaluating left ventricular (LV) function. It enables real-time, non-invasive, and relatively low cost acquisition of cardiac images. Furthermore, it can be performed bedside using portable equipment. The latest generation of echocardiography scanners allows the acquisition of volumetric images of the heart in real time, which could further improve the accuracy of functional analysis. However, accurate and automated analysis of 3D+T recordings is a challenging task. This is due to echocardiography imaging artifacts (e.g. speckle noise, signal dropouts), and the necessity of computationally ecient algorithms to exploit the real-time nature of the modality. A Kalman filter based tracking framework was previously proposed for automatic and real-time analysis of LV structures in 3D echocardiography recordings. The approach was validated for detection and tracking of the endo- and epicardial borders of the LV, and noteworthy results were reported. The main goal of this thesis has been to extend the existing framework with more advanced algorithms for improving endocardial border tracking accuracy. In this work:Advanced edge detection methods including graph-cut based, maximum likelihood, empirical Bayes, and generalized step criterion endocardial edge detectors have been introduced. In addition, a polynomial regression based method has been proposed to lter endocardial edge measurements.Biomechanically constrained tracking of multi-resolution Doo-Sabin surface models has been investigated; an isoparametric nite element analysis (FEA) approach for Doo- Sabin surface models, and modi cation of the tracking framework to use isoparametric FEA have been introduced.The proposed endocardial edge measurement and biomechanically constrained tracking approaches were evaluated using manually segmented 3D echocardiography recordings provided by medical experts. The comparative analyses showed that:The graph-cut based edge detector improves endocardial detection accuracy of the tracking framework at end-diastole.The maximum likelihood and empirical Bayes edge detectors improve detection accuracy for the whole cardiac cycle, while introducing additional computational complexity.The generalized step criterion edge detectors enable real-time maximum likelihood detectors.The polynomial regression based edge ltering provides an intuitive controller for the tradeo between edge detection bias and variance.Biomechanical constraints can signi cantly improve endocardium tracking accuracy of subdivision surface models with high control node resolutions.