Revisiting CT perfusion for acute ischemic stroke with deep learning

This thesis explores the intersection of medical imaging, artificial intelligence, and physics. Specifically, the thesis presents deep learning methods developed to improve CT perfusion imaging analysis for patients with acute ischemic stroke.
The research introduces a method that eliminates the need for traditional perfusion analysis by working directly with CT perfusion source data, instead of the so-called perfusion maps, to identify infarcted brain tissue. Moreover, the research addresses inherent CT perfusion noise problems through physics-informed neural networks that combine mathematical descriptions of tissue perfusion with deep learning, leading to more accurate perfusion analysis. Subsequently, the thesis introduces a method to accelerate such physics-informed perfusion analysis. Additionally, to tackle clinical limitations of radiation exposure and motion sensitivity, the thesis presents a method to reconstruct complete CT perfusion images from sparse measurements. This approach also enables perfusion analysis from alternative imaging modalities. Lastly, the work introduces an approach for collateral circulation assessment for distal vessel occlusions based on CT perfusion outcomes.
In conclusion, the research demonstrates potential to improve CT perfusion analysis and stroke diagnosis accuracy, reduce radiation exposure, mitigate motion artefacts, and enhance assessment of collateral circulation in distal vessel occlusions through innovative deep learning methods.