Automated segmentation of the heart in high-dimensional computed tomography

Computed tomography (CT) of the heart is a low-cost, non-invasive medical imaging technique used for various clinical indications. It visualizes the coronary arteries, myocardium, heart chambers including valves, and surrounding tissue with high spatial resolution. While three-dimensional (3D) CT is commonly used in clinical practice, valuable information can be retrieved from additional dimensions, namely temporal and spectral information. For various applications, the delineation of specific objects or structures in images is necessary (i.e. segmentation). Since manual segmentation of 3D medical images is extremely time-consuming, automated methods have been developed. In recent years, deep learning with convolutional neural networks (CNNs) has become the state-of-the-art for image segmentation. Despite the high accuracy achieved by these methods, they often do not generalize well to unseen images with different contrast enhancement protocols, different CT acquisition protocols, different patient populations with different pathologies, and different cardiac phases. Furthermore, they sometimes produce anatomically implausible shapes of heart structures. This thesis presents approaches to tackle these challenges with a focus on the automated segmentation of the heart in high-dimensional CT.