Cardiovascular diseases (CVDs) are the leading cause of death in the world, accounting for 17.9 million deaths each year, 31\% of all global deaths. According to the World Health Organisation (WHO), this number is expected to rise to 23 million by 2030. As a noninvasive technique, medical imaging with corresponding computer vision techniques is becoming more and more popular for detecting, understanding, and analysing CVDs. With the advent of deep learning, there are significant improvements in medical image analysis tasks (e.g. image registration, image segmentation, mesh reconstruction from image), achieving much faster and more accurate registration, segmentation, reconstruction, and disease diagnosis.
This thesis focuses on cardiac magnetic resonance images, systematically studying critical tasks in CVD analysis, including image registration, image segmentation, cardiac mesh reconstruction, and CVD prediction/diagnosis. We first present a thorough review of deep learning-based image registration approaches, and subsequently, propose a novel solution to the problem of discontinuity-preserving intra-subject cardiac image registration, which is generally ignored in previous deep learning-based registration methods. On the basis of this, a joint segmentation and registration framework is further proposed to learn the joint relationship between these two tasks, leading to better registration and segmentation performance. In order to characterise the shape and motion of the heart in 3D, we present a deep learning-based 3D mesh reconstruction network that is able to recover accurate 3D cardiac shapes from 2D slice-wise segmentation masks/contours in a fast and robust manner. Finally, for CVD prediction/diagnosis, we design a multichannel variational autoencoder to learn the joint latent representation of the original cardiac image and mesh, resulting in a shape-aware image representation (SAIR) that serves as an explainable biomarker. SAIR has been shown to outperform traditional biomarkers in the prediction of acute myocardial infarction and the diagnosis of several other CVDs, and can supplement existing biomarkers to improve overall predictive performance.
