Machine Learning For Simulation-Assisted X-ray Angiography
Cardiovascular disease is the leading cause of death worldwide, and within that, neurovascular disease is especially critical: vessel occlusions or ruptures can kill brain tissue outright. These events are often rooted in pre-existing vessel malformations that weaken vessel walls and disrupt normal blood flow. X-ray-based angiography, typically performed by injecting a contrast agent into the vessel of interest, is essential for diagnosing and treating these malformations, offering high-resolution views of intracranial blood vessels and flow. But the physical processes behind an angiographic image, how X-rays interact with tissue and how blood and contrast medium actually flow, are only partially measurable, even though understanding them matters a great deal clinically.
Computer simulation can model these processes by combining measured data with mathematical models, but the underlying numerical algorithms for X-ray-matter interaction and blood flow are computationally expensive, often taking hours to run even on high-performance computers. This thesis uses machine learning to cut that runtime down dramatically, training models as fast surrogates that approximate simulation results directly, while also using neural networks to improve data quality, discover and tune model parameters, and combine with conventional solvers in hybrid approaches.
The thesis develops several such methods, including surrogate models for predicting 3D X-ray scatter radiation distributions and time-resolved 3D blood flow velocity fields, using network architectures specifically designed around the underlying physics rather than generic designs. These surrogates predict scatter distributions and flow fields within seconds, with only a small increase in error compared to full simulation. The thesis also presents a reconstruction method for time-resolved angiographic contrast flow: by training on simulated angiographic images, a neural network learns the physical relationship between local and temporal contrast distribution, enabling efficient time-resolved reconstruction with acceptable accuracy. Together, these results highlight the potential of combining deep learning with physics-based medical simulation.