Deep Learning Applications for the Assessment of Valvular Heart Disease using Transthoracic Echocardiography

Abstract

Valvular heart disease (VHD) is a class of diseases that affect the heart valves, and are associated with a reduced cardiac output and a poor prognosis for the afflicted patients. Transthoracic echocardiography (TTE) is the cornerstone for the assessment of VHD. However, currently there does not exist any gold standard TTE method for the assessment of the VHD, hence cardiologists rely on an integrative approach of multiple measurements, which is both time-consuming and associated with high inter-observer variability. In light of recent advancements in ultrasound technology and within the field of deep learning (DL), some of the key challenges related to workload and inter-observer variability for the assessment of VHD could be addressed. The goal of this thesis was to explore the integration of DL-based approaches for both new and existing TTE methods for quantifying VHD severity. Three studies are presented. First, we present a DL approach for automatic monitoring of the mitral valve apparatus from TTE B-Mode images, which was demonstrated feasible for detecting and quantifying mitral valve prolapse and stenosis. Secondly, we present a DL framework for the automatic quantification of mitral regurgitation from segmentation and time integration of regurgitant flow convergence zones from 2-D Color Doppler sequences. This framework was able to distinguish between mild, moderate, and severe cases, and measurements had a good correlation with TTE and magnetic resonance imaging references. Finally, we present a simulationbased DL approach for sub-pixel segmentation of poor resolution images of regurgitant valves, in the context of a high frame rate 3-D Doppler ultrasound framework. Clinical feasibility was demonstrated, and central challenges were highlighted.