Date of Award

Fall 2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Biomedical Engineering (ENAS)

First Advisor

Duncan, James

Abstract

Coronary artery disease remains the leading cause of death around the world. Acute myocardial infarction (MI) causes regional dysfunction which places remote areas of the heart at a mechanical disadvantage resulting in long term adverse left ventricular (LV) remodeling and complicated congestive heart failure (CHF). Stress echocardiography is currently the clinically established, cost-effective 2D imaging technique for detecting and characterizing myocardial injury by imaging the left ventricle at rest and after either exercise or pharmacologically-induced stress to reveal ischemia and/or infarct. However, the inherent limitations of a 2D echocardiography make it difficult to characterize the whole 3D volume of ischemic/infarct zone, and the qualitative assessment of wall-motion abnormality to characterize myocardial deformation leads to variability among experts. Although 3D echocardiography (3DE) has potential to address the limitations of 2D imaging, it is not widely accepted in standard clinical use due to the low signal-to-noise ratio (SNR). With the recent advancements in deep learning algorithms, many segmentation and registration tasks have achieved near expert level accuracy. Previous works have shown the utility of strain analysis as a way to quantify the degree of wall-motion abnormality in cardiac imaging modalities. Still, many of the current deep learning frameworks focus largely on intensity-based features which are difficult to train on 3D echocardiography datasets due to the low SNR, which in turn leads to poor strain analysis. In this work, we explore the use of attention mechanisms in deep learning frameworks to improve both the segmentation and motion tracking of left ventricle in 3D echocardiography. We first develop a multi-frame attention architecture for segmentation of LV in 3DE. We show that incorporating multiple inter-frame spatiotemporal features allows improved segmentation performance. The model extracts highly correlated features between target volume and reference volume to guide the location of the myocardial boundaries even when the signal intensity was low. Our proposed framework is also easily scalable to incorporate as many frames as reference volumes. Second, we show the feasibility of end-to-end unsupervised learning framework for LV motion tracking in 3DE by using a Spatial Transformer Network. Third, we refine the unsupervised 3D LV motion tracking further by developing a novel co-attention spatial transformer network (Co-Attention STN) . Co-Attention STN extracts inter-frame dependent features between end-diastole and end-systole frames to improve the motion tracking in an otherwise noisy 3D echocardiographic image sets. The addition of our novel temporal constraint further regularizes the motion field to produce smooth and realistic cardiac motion. Experimental results demonstrate that our Co-Attention STN provide a superior performance compared to existing methods. Finally, we conducted both acute and chronic post myocardial infarction studies in pigs to assess the regional mechanical changes in the left ventricle by assessing three dimensional strain from our proposed Co-Attention STN. We evaluated our echocardiography-derived strain maps by comparing to \textsuperscript{99m}Tc-tetrofosmin SPECT perfusion maps and showed that infarct localization from echo-derived strain maps correspond well with the regions of decreased radiotracer uptake in SPECT. Our results show that strain analysis from Co-Attention STN correspond well with the matched SPECT perfusion/viability maps, demonstrating the clinical utility for using 3D echocardiography for infarct localization. In summary, our work aims to provide an objective, quantitative tools for characterizing wall-motion abnormality with strain analysis in 3D echocardiography.

Download

Share

COinS