Assessing anatomy and function of the heart using 4D cardiac MRI and deep learning

Cardiovascular magnetic resonance (CMR) imaging is the reference modality for morphological and functional assessment of the heart. Typically, obtaining a high-resolution image and shape of the cardiac anatomy is crucial for such assessment. Short-axis CMR imaging, covering the entire left and right ventricles is routinely used to determine quantitative parameters of both ventricles' function. For this, manual or (semi-)automatic segmentation of the left and right endo- and epicardial structures in short-axis CMR images (CMRI) for at least end-diastole and end-systole is a key task. Manual segmentation of CMRIs is laborious and prone to intra- and inter-observer variability. Moreover, to quantify parameters of cardiac motion requires segmentation across a complete cardiac cycle, comprising 20 to 40 phases per patient. Over the last few years many state-of-the-art deep learning segmentation approaches for short-axis CMRI have been developed. For automatic left ventricle segmentation such methods can achieve performance level of human experts. However, even the best performing methods generate anatomically implausible segmentations. Therefore, existing semi-automated or automated segmentation methods for CMRIs regularly require (substantial) manual intervention. Furthermore, short-axis CMR scans with high temporal resolution are often highly anisotropic and suffer from respiratory motion induced inter-slice misalignment. Moreover, caused by the low through-plane resolution, short-axis CMR volumes often lack whole-heart coverage predominately at the apex and base of the heart. These shortcomings may hamper correct assessment of cardiac anatomy and subsequently hinder accurate analysis of cardiac function. This thesis presents approaches to tackle these challenges.