Machine learning based analysis of cardiovascular images

Abstract

Cardiovascular diseases (CVDs), including coronary artery disease (CAD) and congenital heart disease (CHD) are the global leading cause of death. Computed tomography (CT) and magnetic resonance imaging (MRI) allow non-invasive imaging of cardiovascular structures. This thesis presents machine learning methods for analysis of CT images of patients with coronary artery disease and MR images of patients with congenital heart disease. To identify CAD patients with increased cardiovascular risk, methods for automatic coronary artery calcium (CAC) scoring in cardiac CT are presented. A method is described that automatically identifies CAC lesions in non-contrast-enhanced CT. Candidate lesions can be selected for expert review based on classifier uncertainty, allowing both accurate and fast CAC scoring. This method and other (semi-)automatic methods for CAC scoring are evaluated in a standardized framework. This evaluation shows that automatic patient CVD risk categorization is feasible, but that CAC lesions at ambiguous locations such as the coronary ostia remain challenging. Furthermore, a deep learning-based method for automatic CAC scoring in contrast-enhanced cardiac CT is described. This method uses a pair of convolutional neural networks (CNNs). Automatically determined calcium scores showed a strong correlation with manual calcium scores in non-contrast-enhanced CT. This might obviate the need for a dedicated non-contrast CT scan for CAC scoring, and thus reduce the CT radiation dose received by patients. To further allow potential CT radiation dose reduction, a CNN is proposed that transforms low-dose CT images into routine-dose CT images. To encourage the CNN to generate realistic images, feedback from an adversarial CNN is used. The results show that the images produced are similar to routine-dose images and that noise reduction by the CNN allows CAC quantification in low-dose patient cardiac CT images in which calcium scoring was otherwise infeasible. In addition, methods are proposed for coronary centerline extraction and coronary artery segmentation in contrast-enhanced cardiac CT. In contrast to previously published methods, these methods do not use predefined vesselness filters. Instead, CNNs predict the direction and radius of coronary arteries or the probability that a voxel lies in a coronary artery, based only on CT values. Moreover, experiments show that a single CNN can be trained to not only perform coronary artery segmentation in cardiac CT, but also brain tissue segmentation in brain MRI and pectoral muscle segmentation in breast MRI. This may lead more general systems in which tasks in different imaging modalities are combined. Finally, to analyze cardiac MR images of patients with congenital heart disease, a CNN is presented for segmentation of the myocardium and blood pool. The network uses dilated convolution kernels, which provide high-resolution feature maps, and reduce the number of parameters in the network. The method achieved top-ranking results among evaluated automatic methods in a public challenge.