Learning anatomical image representations for cardiac imaging

Advances in cardiac imaging techniques play a vital role in fighting against cardiovascular diseases by helping achieve early and accurate diagnoses as well as guidance in interventional cardiac procedures. Along with these imaging systems, computer-aided diagnosis (CAD) is essential in modern healthcare to yield useful and comprehensive clinical information in a short time. However, CAD systems have not been widely adopted in clinical practice due to technical limitations imposed by the imaging systems. In particular, cardiac cine Magnetic Resonance (MR) imaging has a low through-plane resolution and may contain motion artefacts due to the acquisition process. Similarly, the quality of cardiac ultrasound images is operator-dependent and often has poor signal-to-noise-ratio, which presents difficulties for automated image analysis. To tackle these limitations and enhance the accuracy and robustness of the automated image analysis, this thesis focuses on the development and application of state-of-the-art machine learning (ML) techniques in cardiac multi-modal imaging. Specifically, we propose new image feature representation types that are learnt with ML models and aimed at highlighting the correspondences between multi-modal data. These representations are also intended to visualise the cardiac anatomy in more detail for better interpretation and analysis. Moreover, we investigate how quantitative analysis can benefit when these learnt image representations are used in segmentation, motion-tracking and multi-modal image registration. Specifically, a probabilistic edge-map representation is introduced to identify anatomical correspondences in multi-modal cardiac images and demonstrate its use in spatial image alignment and anatomical landmark localisation problems. Additionally, a novel image super-resolution framework is introduced for the enhancement of cardiac cine MR images and we show that high resolution image representation can be useful and informative for various types of subsequent analysis including volumetric measurements and strain analysis.