Ventricular Segmentation and Coronary Artery Disease Classification in Echocardiography with Limited Data

Document Type



Coronary artery disease (CAD) is the third most prominent cause of mortality in the world and is linked to 17.8 million deaths per year. Blockage of the coronary arteries leads to reduced blood supply to the heart muscles, causing complications such as myocardial infarction (MI), otherwise known as a heart attack. If left untreated, this can result in long-term damage. MRI or CT scans are typically performed to check for MI. Although it is challenging, diagnosis of CAD using echocardiography is possible while being relatively inexpensive and quicker than other imaging modalities. Furthermore, automation of this process can help reduce the burden on cardiologists, and even make diagnosis accessible where clinical experts are not present. The diagnostic process for CAD and other heart diseases may involve segmentation of the heart chambers such as the left ventricle in order to assess cardiac function. In the first part of this project, a self-supervised contrastive learning method is proposed to segment the left ventricle from echocardiography when limited annotated images exist. Furthermore, the effect of contrastive pretraining is studied on two well-known segmentation networks, UNet and DeepLabV3. The results show that contrastive pretraining helps improve segmentation performance, particularly when annotated data is scarce. We achieve results comparable to state-of-the-art fully supervised algorithms with self-supervised pretraining followed by fine-tuning on just 5% of the data. We also show that our solution outperforms what is currently published on a large public dataset (EchoNet-Dynamic), achieving a Dice score of 0.9252. We also compare the performance of our solution on another smaller dataset (CAMUS) to demonstrate the generalizability of our proposed solution. The second part of this project presents an end-to-end deep learning approach for classification of MI in echocardiography videos. We show how our fully automatic method outperforms existing published work for automatic MI classification (+1.8% F1-score) and shows slightly lower results compared to the best performing semi-automatic method (F1-scores of 82.90 vs. 85.71), without the need for intermediate segmentation. Our work has been developed using multiple publicly available datasets and was evaluated on an MI classification dataset with an F1-score of 87.09%.

First Page


Last Page


Publication Date



Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfillment of the requirements for the M.Sc degree in Computer Vision

Advisors: Dr. Mohammad Yaqub, Dr. Karthik Nandakumar

Online access for MBZUAI patrons