Fully automated segmentation of mitral valve in real-time three-dimensional ultrasound data and its applications

Abstract

Being the critical gateway that regulates the oxygenated blood flow from the left atrium to the left ventricle, the mitral valve has been extensively studied by clinical experts. In order to derive quantitative parameters that could lead to significant clinical decisions, the anatomy and the dynamics of the live mitral valve must first be imaged through the use of ultrasound devices. In recent years, the most commonly used non-invasive imaging modality is real-time three-dimensional transesophageal echocardiography (RT3DTEE). Although this latest imaging technology enables unprecedented in-vivo visualization of the mitral valve and its surrounding tissues, clinical experts are still required to spend hours to trace the mitral valve manually in three-dimensional (3D) and four-dimensional (4D) settings. This time-consuming and laborintensive manual work often requires a very demanding level of eye-hand coordination and mental concentration in order to have clinically-qualified delineations. Additionally, the inferior image quality of RT3DTEE causes many readily or commercially available solutions stumble. Hence, being able to fully automatically segment the mitral valve from RT3DTEE has always been a challenging problem.

This thesis first presents the background information on mitral valve and RT3DTEE technology. By exploiting the approximately radial-symmetric geometry of the mitral valve, a simple yet effective technique is proposed to determine whether the valve is in systole(closed-valve) or diastole(open-valve) from only what it is available in RT3DTEE images. This labeling exercise is often considered to be a sub-problem in the mitral valve segmentation problem. By doing so, clinical experts can then study the anatomy and dynamics with respect to the valve states, while algorithmic approaches can make use of such information to track the mitral valve in various time instances.

Next, this thesis focuses on a practical solution that fully automatically delineates the mitral valve by formulating the segmentation problem as a machine learning problem, of which the solution is further optimized by an energy minimization function. It is then demonstrated, when compared to other state-of-the-art approaches, the described approach can further reduce the initial size of the pre-collected training data from clinicians, can still perform well regardless of how the mitral valve is being imaged and, most importantly, is able to extract the mitral valve in a cardiac cycle while preserving its volumetric details. Finally, the applicability of the presented methods is demonstrated through the derivations of several important clinical morphological parameters of the mitral valve by comparing them against clinical experts’ measurements which is the gold standard in the experiments. Altogether, this work steers the mitral valve segmentation task to a even more systematic and automatic direction.