Analysis of Cardiovascular Disease using Cardiac Computed Tomography and Deep Learning
Cardiovascular disease is the leading cause of death in the United States. 30~50% of cardiovascular disease is caused by coronary artery disease (CAD). CAD is caused by the development of coronary stenosis (the narrowing of the coronary arteries), which restricts blood supply to the myocardium and causes myocardial ischemia and eventually heart attack. Accurate quantification of coronary stenosis is crucial to evaluate the severity of CAD and plan appropriate treatments. Further, the accurate detection and quantification of impaired myocardial function also have great prognostic value in patients with CAD. Computed Tomography (CT) was initially focused on evaluating CAD through CT coronary angiography; the recent developments in 4D CT allow the acquisition of full 3D volumes across the entire cardiac cycle and thus enable the assessment of myocardial function. This dissertation introduces novel analytical and deep learning-based techniques to analyze coronary stenosis and myocardial dysfunction from CT. For stenosis, we develop a novel quantification algorithm to overcome the demanding challenge of quantifying small stenosis below the image resolution and greatly enhance the accuracy of the estimates of stenosis severity. For myocardial dysfunction, we demonstrate that 3D myocardial regional shortening (RSCT) measured from CT is an outstanding quantitative classifier to detect regional myocardial wall motion abnormality (WMA). However, the clinical utility of regional myocardial function quantifications such as RSCT measurements is limited by the dependence on manual image analysis. To solve this unmet need, we develop a deep learning (DL) framework to automatically and simultaneously accomplish two essential image processing tasks: (1) segmenting heart chambers and (2) reformatting CT volumes into clinically standard planes. Furthermore, regional myocardial function analysis is computationally expensive to perform for each patient, whereas a trained DL model can be easily deployed and can quickly generate results. Thus, we present the first DL approach to detect regional WMA from high-resolution 4DCT empowered by unique features available from dynamic volume rendering. Overall, these novel techniques have outstanding promise to replace time-consuming manual work and lead to automatic, fast, and accurate diagnosis of cardiovascular disease.