Real-time segmentation of 3D echocardiograms using a state estimation approach with deformable models

Cardiac ultrasound, or echocardiography, is considered to be the quickest and most cost effective imaging modality for assessment of cardiac function. The modality is unique in that it allows for real-time imaging of the heart, using portable equipment. The latest generation of echocardiography scanners are capable of acquiring dense image volumes in either real time, or over a few heart beats. Methods for analyzing these images, however, are lagging behind, with all existing methods for segmentation of 3D ultrasound data considered too computationally intensive to operate at acquisition rate. Availability of methods for analysis of 3D ultrasound in real-time could open up possibilities for very quick and simple measurement of cardiac function, potentially conducted during image acquisition. Results from automatic methods would also lessen the amount of interor intra-examiner differences introduced during the analysis, which in turn could lead to more reproducible results. The left ventricle of the heart has traditionally been the chamber of most interest within the field cardiac image analysis. This thesis has therefore focused on the left ventricle, and investigated ways to measure different aspects of the chamber, such as chamber volumes, myocardial strain and myocardial mass, by means of state estimation techniques. The main goal of this thesis has been to explore the possibilities of using state estimation methods for segmentation and tracking of structures in volumetric data with deformable models. The ability of using non-iterative estimators, such as the Kalman filter, for fitting deformable models to image structures would radically reduce the computationally effort required for performing a 3D segmentation, and open up for real-time usage. An existing Kalman tracking framework have therefore been extended to operate in volumetric data. The framework has been successfully demonstrated to fit ellipsoids, spline surfaces, active-shape surfaces and subdivision surfaces to image data. Furthermore, theory for the simultaneous tracking of several models have been developed. Finally, methods for combining edge-detection measurements with speckle-tracking measurements have been shown, with the potential advantages of material tracking with the lack of drift in edge detection. The framework has been demonstrated to successfully conduct high quality segmentation and tracking of the left ventricle, while operating in real-time without user intervention. Processing times are in the range of milliseconds per frame on standard computer hardware, which is orders of magnitude faster than the state-of-the-art methods. Applications of the framework on 3D echocardiograms has been investigated to prove the feasibility of the framework for automatic analysis of various aspects of cardiac function, such as view alignment, measurement of chamber volumes, myocardial muscle volume and regional myocardial strain. The results for view alignment are considered state of the art, and the results for volume measurements are close to the results of similar studies with semi-automatic tools.