Towards Non-invasive Stroke Diagnosis: A Neural Network Approach to CT Perfusion Imaging With Subsampling

Artificial Neural Networks are powerful tools that help doctors across various medical imaging applications. This thesis focuses on the generation of perfusion maps from computed tomography using neural networks. Doctors use these maps to identify tissues threatened by ischemia that are still potentially salvageable and may be a target for thrombolytic therapy. The results obtained with a U-Net-like network on the UnitoBrain dataset are already promising. However, epidemiologic studies highlight that the radiation dose from even two or three CT scans results in a detectable increase in cancer risk. What happens if we apply different subsampling methods on the volume data to reduce the number of input scans? This thesis shows that a low number of scans is enough to generate good perfusion maps to estimate the ischemic core, opening the path to less invasive exams with novel perfusion protocols with lower radiation doses.