Open Collections

Abstract

Echocardiography (echo) plays an important role in cardiac imaging and provides a non-invasive, low-cost, and widely available diagnostic tool for the comprehensive evaluation of cardiac structure and function. However, ultrasound image interpretation remains a challenge, particularly for the novices. In this thesis, I propose several machine learning methods with the aim of helping in cardiac assessment. In particular, the proposed methods take advantage of novel advancements in spatio-temporal analysis of videos to find a better representation of the cardiac cine series. First, I propose a customized supervised learning method in order to find two specific phases in a cardiac cycle. Specifically, identification of the end-systolic (ES) and end-diastolic (ED) phases from the echo cine series is a critical step in the quantification of cardiac chamber size and function. Later, I develop a self-supervised learning method for frame-rate up-conversion to augment conventional imaging without the need of specialized beamforming and imaging hardware. In another work, I propose a self-supervised learning framework to synchronize various cross-sectional 2D echo videos without any human supervision or external inputs. I show that such rich, yet free semantic representation can be used not only for synchronization of multiple cine series, but also for fine-grained cardiac phase detection. Finally, I propose a semi-supervised learning method to detect the cardiac rhythm based solely on echo without the need for an electrocardiogram (ECG), which is commonly used for cardiac rhythm detection.