Automated volume measurements in echocardiography by utilizing expert knowledge

Abstract

Left ventricular (LV) volumes and ejection fraction (EF) are important parameters for diagnosis, prognosis, and treatment planning in patients with heart disease. These parameters are commonly measured by manual tracing in echocardiographic images, a procedure that is time consuming, prone to inter- and intra-observer variability, and require highly trained operators. This is particularly the case in three-dimensional (3D) echocardiography, where the increased amount of data makes manual tracing impractical. Automated methods for measuring LV volumes and EF can therefore improve efficiency and accuracy of echocardiographic examinations, giving better diagnosis at a lower cost.

The main goal of this thesis was to improve the efficiency and quality of cardiac measurements. More specifically, the goal was to develop rapid and accurate methods that utilize expert knowledge for automated evaluation of cardiac function in echocardiography.

The thesis presents several methods for automated volume and EF measurements in echocardiographic data. For two-dimensional (2D) echocardiography, an atlas based segmentation algorithm is presented in paper A. This method utilizes manually traced endocardial contours in a validated case database to control a snake optimized by dynamic programming. The challenge with this approach is to find the most optimal case in the database. More promising results are achieved in triplane echocardiography using a multiview and multi-frame extension to the active appearance model (AAM) framework, as demonstrated in paper B. The AAM generalizes better to new patient data and is based on more robust optimization schemes than the atlas-based method. In triplane images, the results of the AAM algorithm may be improved further by integrating a snake algorithm into the AAM framework and by constraining the AAM to manually defined landmarks, and this is shown in paper C. For 3D echocardiograms, a clinical semi-automated volume measurement tool with expert selected points is validated in paper D. This tool compares favorably to a reference measurement tool, with good agreement in measured volumes, and with a significantly lower analysis time. Finally, in paper E, fully automated real-time segmentation in 3D echocardiography is demonstrated using a 3D active shape model (ASM) of the left ventricle in a Kalman filter framework. The main advantage of this approach is its processing performance, allowing for real-time volume and EF estimates.

Statistical models such as AAMs and ASMs provide elegant frameworks for incorporating expert knowledge into segmentation algorithms. Expert knowledge can also be utilized directly through manual input to semi-automated methods, allowing for manual initialization and correction of automatically determined volumes. The latter technique is particularly suitable for clinical routine examinations, while the fully automated 3D ASM method can extend the use of echocardiography to new clinical areas such as automated patient monitoring.

In this thesis, different methods for utilizing expert knowledge in automated segmentation algorithms for echocardiography have been developed and evaluated. Particularly in 3D echocardiography, these contributions are expected to improve efficiency and quality of cardiac measurements.