Date of Award

Spring 2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Engineering and Applied Science

First Advisor

Duncan, James

Abstract

Cardiovascular diseases are the leading cause of death globally and impose great economic and societal burden. Of these diseases, coronary artery disease is the most prevalent. Occlusions in the coronary artery restricts the flow of blood to the myocardium, leading to tissue damage and subsequent adverse cardiac remodeling and congestive heart failure. Early detection of myocardial dysfunction is imperative in reducing the overall mortality rate through timely intervention. To that end, there is significant motivation to develop automated techniques to characterize cardiac function using cost-effective and non-invasive tools. Motion and morphology analysis using echocardiography give insights into overall cardiac health by providing estimates of important indicators such as myocardial strain, ejection fraction, and ventricular shape indices. Achieving this requires accurate and reliable tracking and segmentation, which are very challenging tasks in echocardiography due to the relatively low image quality. However, both of these tasks are known to be highly related and one task can be used to guide the other. Combining this idea with the inherent spatiotemporal nature of echocardiography yields useful constraints that can improve both motion tracking and segmentation predictions. This dissertation focuses on developing a unified multi-task learning framework that can simultaneously segment the left ventricular myocardium and track its motion through a 3D echocardiographic sequence by exploiting overlapping latent features between the two tasks. We first explore how the incorporation of shape information can be used to guide and improve motion field estimations by developing an unsupervised motion tracking network that learns to optimally deform a source frame to match a target frame. Corresponding source and target segmentation masks serve as constraints to penalize deviations caused by inaccurate motion estimates. Next, we then develop and explore multiple architectural ways to bridge the tasks of motion tracking and segmentation. At the foundation of each of these networks are separate task-specific motion tracking and segmentation branches each with their own set of loss functions and regularizations. The joint decoder network combines these branches during the upsampling, feature reconstruction stage. The iterative network trains the branches separately and uses predictions from one branch to influence and refine the other in a back-and-forth manner. The joint encoder network combines the two branches in a Siamese-style manner during the downsampling, feature deconstruction stage. All of these networks were shown to improve the performances of each task relative to their single-task counterparts. Expanding on these ideas, we then arrive at the cross-stitch network. Cross-stitch units placed at every layer of each branch interweave features pertinent to each task by learning optimal linear combinations of said features. A novel shape consistency unit is introduced which encourages similarity between motion propagated masks and directly predicted segmentations. In addition, a temporal constraint is added in order to incorporate additional time frames for added motion information. Experiment results demonstrate excellent tracking and segmentation performance relative to alternative methods. Finally, we demonstrate the clinical utility and feasibility of our work using both canine and porcine studies with induced myocardial infarctions. The canine studies consists of expert manual tracings of the left ventricular myocardium as well as sonomicrometer crystal readings which provide a reasonably reliable benchmark for strain estimation. The porcine dataset consists of expert manual tracings in the end-diastolic and end-systolic frames as well as corresponding 99mTc-tetrofosmin SPECT perfusion and CT images which provide clinically reliable benchmarks to assess strain and volumetric/morphological indices. Our echocardiography-derived strain curves exhibit strong correlations with the sonomicrometer crystal-derived strains and our estimated infarction localization maps correspond well with 99mTc-tetrofosmin SPECT perfusion maps. Additionally, calculations of left ventricular ejection fraction and sphericity using our method were in general agreement with the CT-derived and manually-derived metrics. Overall, the objective of our work is to provide a framework for a more reliable and objective tool to better characterize myocardial function and detect abnormalities using 3D echocardiography.

Download

Share

COinS