Segmentation of cardiac structures in 3-dimensional echocardiography

Cardiac ultrasound is a key component in modern cardiology, and is widely used to assess and quantify the heart’s anatomy and function. In the past two decades, real-time 3-dimensional ultrasound has emerged as a promising modality, allowing a more accurate appreciation of the heart’s complex geometry, compared to the more conventional 2-dimensional ultrasound. However, the huge increases in information within volumetric images demand accurate and efficient methods for automatically analyzing and quantifying the heart. One of the fundamental problems in the processing of cardiac images is segmentation; the process of extracting a geometric model from the image, describing the shape and motion of an anatomical structure. A framework has previously been proposed for solving this problem in 3D ultrasound images, in which the segmentation is expressed as a state-estimation problem and solved with a Kalman filter. This framework is generic, and allows for incorporating different surface representations, image measurements and prior information on shape and motion, and it has been applied to the left ventricle. The main goal of this thesis has been to apply the existing Kalman filter segmentation framework to different anatomical structures, in particular the aortic root and right ventricle, and to develop a method for biventricular segmentation. Four main contributions have been made. Firstly, a method providing fully automatic quantification of the aortic valve size from 3D ultrasound images has been developed and its feasibility demonstrated. This is a key measurement to be made prior to procedures on diseased valves, and performing manual measurements is time-consuming and subject to inter-observer variation. Secondly, two methods have been developed for automated segmentation of the Right Ventricle (RV). Historically, the importance of the RV has been underestimated. Although the RV’s role in cardiovascular diseases has been more widely recognized in recent years, few methods for segmentation of the RV have been proposed. Because the RV has a more complex and asymmetrical shape compared to the left ventricle, the existing segmentation framework is not immediately applicable to the RV. Therefore, the method was extended with a geometric model learned from statistical analysis on a set of manual segmentations, as well as prior information on the appearance of the ultrasound images. Thirdly, one of the two RV methods uses a novel geometric surface representati