Deep Learning-based Cerebrovasculature Segmentation in CT Perfusion Imaging

Accurate brain vasculature segmentation has various clinical applications, ranging from improving perfusion map generation, a key component of acute stroke triage, to surgical planning and prediction of patient treatment outcomes. Computed Tomography Perfusion (CTP) scans provide 4D spatiotemporal data by repeatedly imaging a single 3D volume over a defined time span. Unlike Computed Tomography Angiography (CTA), which captures a single static snapshot of vessel anatomy, CTP tracks contrast agent passage through the cerebral blood circulation, accumulating dynamic information on vessel topology and tissue perfusion. However, leveraging this rich temporal information is difficult, as CTP images are inherently noisy and suffer from poor contrast and partial volume effect, making manual annotation incredibly tedious and labor-intensive for clinicians. Additionally, the field currently lacks any consensus methodology for extracting the cerebrovascular tree from CTP data. While rapid development of Machine Learning and Deep Learning methods has benefited many medical fields, the application of such methods to brain vasculature segmentation in CTP scans remains limited, as existing methods are mainly developed for other higher resolution modalities and cannot be directly applied to the 4D low-resolution CTP imaging. This Thesis addresses this gap through three contributions. First, a protocol for annotating the cerebral vasculature in perfusion imaging, producing the labeled data required for training. Second, an adaptation of the nnU-Net framework to vessel segmentation in CTP. Third, an investigation of self-supervised, domain-specific pre-training, assessing whether representations learned directly from CTP improve segmentation performance.