Assessment of imaging markers for intracranial aneurysms using artificial intelligence

Abstract

Intracranial aneurysms are bulges in brain blood vessels that, while often harmless, can rupture and cause a life-threatening subarachnoid hemorrhage. Since unruptured intracranial aneurysms are generally asymptomatic and therefore are often not diagnosed until rupture, early detection and treatment are crucial. Aneurysms can be diagnosed using time-of-flight Magnetic Resonance Angiography, a widely used non-invasive imaging method. Although the pathogenesis of intracranial aneurysms remains largely unknown, research suggests a strong familial deposition. This has prompted recommendations to screen first-degree family members of patients with (ruptured) aneurysms, usually at five-year intervals since aneurysms can develop at any time. However, current screening methods have limited effectiveness. The identification of risk factors capable of identifying individuals at the highest risk for aneurysms can improve the screening process. High-risk individuals could be screened more frequently based on an initial evaluation of the brain scan, whereas the screening frequency for low-risk individuals could be reduced or even stopped. This thesis explores computer-assisted and artificial intelligence-based approaches for (semi-)automated assessment of imaging markers associated with intracranial aneurysms, focusing on the circle of Willis. Two methods are presented to semi-automatically assess artery diameters and artery bifurcation angles. The analysis of large cohorts revealed notable anatomical differences, including sex-specific differences in the anatomy of the circle of Willis and associations between specific artery diameters or bifurcation angles and aneurysm presence, offering new insights into potential risk factors. In addition, the thesis investigates the use of advanced artificial intelligence techniques, such as graph neural networks, to analyse circle of Willis topology. By exploring methods for model interpretability and calibration, this research advances towards fully automated, explainable, and clinically implementable AI systems.