The human cerebrovascular system forms a complex, hierarchical network that ensures the metabolic sustenance of the brain. Its structural integrity is a critical determinant of brain health, while its degradation frequently serves as a precursor to stroke and neurodegenerative diseases. To characterize its morphology, clinical research relies on 3D neurovascular images, such as Time-of-Flight Magnetic Resonance Angiography (TOF-MRA), which feeds vascular image analysis pipelines to extract quantitative biomarkers. While promising, the translation of raw volumetric data into actionable clinical insights is currently hindered by three structural barriers: the scarcity of expert annotations due to the time-consuming nature of manual delineation, the heterogeneity of annotation protocols across datasets, and the lack of automated tools for biomarker extraction and subsequent analyses. This thesis addresses these barriers through three methodological contributions spanning data selection, annotation infrastructure, and cerebrovascular quantitative analysis.
To mitigate annotation scarcity, this work introduces V-DiSNet (Vessel-Dictionary Selection Network), a one-shot active learning strategy exploiting the tree-like topology of cerebrovascular structures. By integrating dictionary learning with stratified sampling in a sparse latent space, V-DiSNet identifies a minimal set of samples for expert annotation. Evaluation on public 3D TOF-MRA datasets indicates that the method achieves accuracy comparable to fully supervised approaches while reducing the volume of labeled data.
Addressing the challenge of data heterogeneity, this dissertation presents VesselVerse, an annotation infrastructure comprising annotated images across multiple neurovascular imaging modalities. In contrast to static repositories, the framework integrates multi-expert annotation management, STAPLE-based consensus generation, and version control. Validation by a panel of experts demonstrates inter-rater agreement and confirms that the automated consensus synthesizes expert judgment, enabling community-driven refinement.
Finally, to enable population-scale quantification, this research develops CaravelMetrics, a graph-based analysis framework that extracts morphometric, topological, fractal, and geometric features across anatomically defined brain territories. Application to a cohort of adults spanning six decades quantifies age-related cerebrovascular remodeling, including a decline in total vessel length and an increase in tortuosity consistent with patterns reported in the literature regarding arterial stiffening. Demographic analysis reveals sexual dimorphism in vascular morphology and a positive correlation between educational attainment and vessel length, aligning with literature hypotheses suggesting that cognitive reserve may extend to cerebrovascular health.
Collectively, these contributions establish a pipeline ranging from data creation to population-scale discovery. By addressing the structural barriers of annotation, collaboration, and analysis, this work aims to accelerate vessel annotation and deep learning model training and enable the population-scale investigations necessary to understand the vascular determinants of cognitive aging and disease risk.
