Deep Learning Methods for Echocardiogram Analysis

Cardiovascular diseases are one of the leading causes of death globally. Non-invasive echocardiogram imaging is widely used by clinicians to assess and diagnose cardiac disease. Advances in deep learning have created tools to help clinicians analyze echocardiograms and therefore enhance diagnosis and monitoring of cardiac disease. This work focuses on the application of machine learning on echocardiogram videos for two different tasks; left ventricular ejection fraction (LVEF) estimation and patent ductus arteriosus (PDA) segmentation and identification. The focus of this work is to create deep learning based classification frameworks with interpretable intermediate results by incorporating segmentation architectures into both classification tasks. The proposed LVEF classification model uses LV segmentation masks to estimate left ventricular volume and forms a final ejection fraction prediction by identifying cardiac cycles within an echocardiogram clip. The proposed PDA classification model implements a simple neural network trained on features extracted from PDA shunt masks. Results for the ventricular estimation task were found to be comparable to state-of the-art frameworks with mean absolute error of 6.6%. For the PDA classification task the proposed framework achieved an accuracy of 75% on a novel test set of colour echocardiogram videos.