Calibration and acceleration of thrombosis modelling in intracranial aneurysms

Approximately 5-8% of the general population harbours an intracranial aneurysm (IA), which is a localised dilation or ballooning of the cerebral blood vessel caused by the weakness of the wall of a cerebral artery or vein. Untreated IAs may eventually rupture and lead to death. Nowadays, the non-invasiveness of endovascular approaches is often used as a first-choice treatment due to its low morbidity and mortality risk. The efficacy of endovascular treatment for IA is influenced by both haemodynamics and thrombosis. Currently, in vivo or image-based analysis of thrombosis haemodynamics in realistic anatomies and physiologies is very difficult, if not impossible. Computational modelling has proven to be a powerful tool in predicting thrombosis haemodynamics in IAs before and after endovascular treatment and thus in patient-specific treatment planning or in silico trials. However, before being applied in clinical practice, there is a need to demonstrate the credibility of computational thrombosis modelling, defined as trust in the predictive capability of a computational model. According to the American Society of Mechanical Engineers (ASME) V&V 40, the credibility of computational modelling can be assessed using clinical studies, robust model calibration studies, or population-level validation studies. Given the complexity of thrombus formation, involving blood flow and the net results of a series of biochemical reactions, it is important to improve the efficiency of computational thrombosis modelling for using in population-level model credibility assessment studies.
This thesis aims to improve both the credibility and efficiency of patient-specific computational thrombosis modelling. This work contributes to the following aspects: (1) We create a fully automatic multi-scale modelling workflow that enables population-based in silico studies to calibrate haemodynamic thresholds of thrombus formation against real population-specific data. (2) We identify the most influential factors of our thrombosis model through a comprehensive global sensitivity analysis and further validate the thrombosis model based on a real patient case using patient-specific parameters for those identified as influential ones and the calibrated trigger thresholds of thrombosis initiation. (3) We investigate the use of a physics-informed deep learning model to accelerate thrombosis modelling by leveraging the power of neural networks and GPUs.