Cardiac magnetic resonance (CMR) images play a growing role in diagnostic imaging of cardiovascular diseases. Left Ventricular (LV) cardiac anatomy and function are widely used for diagnosis and monitoring disease progression in cardiology and to assess the patient's response to cardiac surgery and interventional procedures. For population imaging studies, CMR is arguably the most comprehensive imaging modality for non-invasive and non-ionising imaging of the heart and great vessels and, hence, most suited for population imaging cohorts. Due to insufficient radiographer's experience in planning a scan, natural cardiac muscle contraction, breathing motion, and imperfect triggering, CMR can display incomplete LV coverage, which hampers quantitative LV characterization and diagnostic accuracy.
To tackle this limitation and enhance the accuracy and robustness of the automated cardiac volume and functional assessment, this thesis focuses on the development and application of state-of-the-art deep learning (DL) techniques in cardiac imaging. Specifically, we propose new image feature representation types that are learnt with DL models and aimed at highlighting the CMR image quality cross-dataset. These representations are also intended to estimate the CMR image quality for better interpretation and analysis. Moreover, we investigate how quantitative analysis can benefit when these learnt image representations are used in image synthesis.
Specifically, a 3D fisher discriminative representation is introduced to identify CMR image quality in the UK Biobank cardiac data. Additionally, a novel adversarial learning (AL) framework is introduced for the cross-dataset CMR image quality assessment and we show that the common representations learnt by AL can be useful and informative for cross-dataset CMR image analysis. Moreover, we utilize the dataset invariance (DI) representations for CMR volumes interpolation by introducing a novel generative adversarial nets (GANs) based image synthesis framework, which enhance the CMR image quality cross-dataset.
