Spatio-temporal artery analysis using geometric deep learning

Cardiovascular diseases are the leading cause of death worldwide. Vascular diseases may occur in different vessels in the human body, differing in size and tortuosity, and are often marked by abnormal vessel shapes, such as stenosis or aneurysms. These abnormalities may impair vascular function, worsen over time, and may eventually lead to life-threatening events. However, it remains unclear which patients are at risk of such outcomes. This thesis presents methods based on geometric deep learning to analyze diseased arteries in both space and time.

Medical imaging enables monitoring of diseased arteries over time. From these images, a digital 3D reconstruction of the vessel can be obtained through segmentation. While artificial intelligence (AI) has shown promising results on segmentation tasks, these approaches typically require large amounts of training data and often fail to generalize across different types of vessels. This thesis presents novel artery segmentation algorithms that exploit local symmetries in vascular structures. As a result, these algorithms require little training data and generalize to other arteries.

To track disease progression over time, patients occasionally return for follow-up scans, resulting in a time-series of vascular models obtained at discrete moments in time. However, vascular disease progression is in reality continuous and gradual. This thesis presents continuously deforming vascular models using neural fields, enabling interpolation of artery shapes between scans.

Building collections of vascular models over time can also reveal patterns of disease development on a larger scale. Such datasets enable training AI models to predict future disease progression in new patients. However, progression is influenced by many factors, including age, sex, and hemodynamic forces. This thesis presents an AI model that predicts the growth of an abdominal aortic aneurysm, based on vessel geometry and local features.

Together, the methods presented in this thesis provide a foundation for future research in cardiovascular disease. Integrating segmentation, continuous modeling, and progression prediction may lead to new insights into individual disease development and support the development of personalized patient care.