Cardiovascular diseases are the leading cause of death worldwide, with neurovascular diseases playing a particularly crucial role, as vessel occlusions or ruptures can lead to brain tissue death. The causes of ruptures and occlusions are often pre-existing malformations of blood vessels, which can result in weakened vessel walls and patho logical blood flow. Medical imaging is essential for the treatment and diagnosis of these malformations. In particular, X-ray-based angiography enables high-resolution visualization of intracranial blood vessels and blood flow. This usually involves in jecting a radiopaque contrast medium into the vessel of interest, enhancing contrast against surrounding structures. Various physical processes contribute to the formation of an angiographic image. These underlying processes are often not fully measurable, yet understanding them is of essential importance. These include complex interactions of X-rays with matter and the flow behavior of blood and contrast medium. Computer-assisted methods enable the simulation of these physical processes. To achieve this, measurement data and extracted parameters can be coupled with a mathematical model, which enables the approximation of these processes on a computer. Subsequently, potentially clinically relevant information can be extracted from these simulations. Computer-assisted simulations for X-ray-matter interactions and blood flow are based on numerical algorithms that require extensive computational effort. This sig nificant computational demand leads to long run times of several hours even on high performance computers. Machine learning methods offer the potential to drastically reduce runtime. To achieve this, models are usually trained as surrogate algorithms that can efficiently predict simulation results during inference. Additionally, neural networks offer the possibility of improving data quality, discovering and optimizing model parameters, and integrating with conventional solution algorithms for hybrid methods. Within the scope of this thesis, multiple machine learning and simulation-based methodologies are presented. These include surrogate models for the efficient pre diction of three-dimensional X-ray scatter radiation distributions and time-resolved three-dimensional blood flow velocity fields. Network architectures specifically tai lored to the underlying physical processes are developed, offering advantages over conventional networks. The outcomes of the experiments conducted demonstrate that scatter radiation energy fluence distributions and blood flow velocity fields can be predicted within seconds with a small additional error. Furthermore, a recon struction method for time-resolved angiographic contrast agent flow is presented. By training on simulated angiographic images, the underlying physical laws between lo cal and temporal contrast agent distributions can be exploited by a neural network, resulting in efficient time-resolved reconstruction with acceptable error. The pre sented methods and results underscore the potential of the synergy between deep neural networks and medical computer-assisted physics
Noah Maul (Thu,) studied this question.