Deep generative model for synthesising and analysing cardiac magnetic resonance images

Cardiovascular disease (CVDs) is still the main disease causing many deaths around the world. According to the World Heart Federation's 2023 World Heart Report, approximately 20.5 million deaths in 2021 were attributed to CVDs, accounting for nearly one-third of global fatalities. Over the past few decades, deep learning algorithms have increasingly been applied in magnetic resonance imaging (MRI) in the medical field, and in particular, have become central to the diagnosis and prediction of CVDs. However, the dynamic motion of the heart and its complex and changeable anatomy pose many challenges to the interpretation of cardiac magnetic resonance (CMR) data. Traditional manual analysis methods are time-consuming and provide variable results. At the same time, generative models have advanced medical image analysis, especially for downstream cardiac image analysis tasks. The aim is to use these synthetic images as viable alternatives to real data in deep learning model training, providing cutting-edge solutions in data segmentation, registration, and strain analysis.

This thesis systematically investigated several probabilistic generative models applied specifically to cardiac image analysis, including multi-channel variational autoencoders (VAEs), generative adversarial networks (GANs), and latent diffusion models (LDMs), using cine CMR and tagging CMR images as primary subjects. Cine CMR provides high-resolution dynamic sequences to assess cardiac morphology and myocardial function throughout the cardiac cycle. Tagging CMR enables the quantification of myocardial deformation by encoding spatial modulation patterns into the myocardium. The efficacy of these models is validated through multiple metrics and downstream tasks such as cardiac segmentation and myocardial strain analysis. Initially, we comprehensively reviewed existing deep learning-based image generation techniques in medical image synthesis. Subsequently, we introduced a sparse multi-channel VAE to learn the joint latent representation of cine and tagging CMR images. The proposed model can generate tagging CMR from cine CMR alone, thereby
enabling myocardial strain estimation straight from cine CMR images. This represents a novel approach within cardiac imaging research and could potentially replace the conventional clinical use of tagging image sequences as a basis for myocardial motion and strain analysis. Furthermore, we introduced an innovative framework employing latent denoising diffusion implicit models (DDIM) to synthesise full-spatial cine CMR images. We investigated whether these synthetic images can serve as viable substitutes for real data in downstream cardiac image analysis tasks. Building upon this, we present a novel spatial-temporal generative model that leverages latent DDIM conditioned on demographic and clinical factors, capable of synthesising realistic 4D cardiac cine CMR image sequences.

Overall, the methodologies presented in this research demonstrate potential for innovation and practical applications. The method introduced here may potentially revolutionize traditional clinical diagnosis and intervention methods, and introduce new perspectives on applying deep learning models in medical imaging. These models show promising performance in the generative field, not only promising insights into cardiac conditions, but also advancing the development of personalized medical diagnosis and prediction solutions in the field of cardiology.