Cardiovascular disease (CVD) remains one of the leading causes of mortality worldwide. Early detection and prevention of CVD are crucial for improving patient outcomes and reducing healthcare costs. Since the early 20th century, researchers have identified retinal vascular abnormalities as potential biomarkers for CVD. The retinal vasculature can be non-invasively assessed using digital fundus images (DFIs), which are easily acquired with fundus cameras. This approach enables efficient analysis of retinal abnormalities, positioning the eye as a window into vascular health. Recent advancements in deep learning have provided powerful tools to automate and standardize retinal measurements, offering unprecedented opportunities to analyze large-scale DFI databases and study correlations between ophthalmic as well as systemic health diseases and retinal vasculature. This doctoral research focuses on four key computer vision-based pillars relevant to the analysis of retinal microvasculature in large DFI datasets:1) Development of an open-source vasculature biomarker computation toolbox; 2) Deep learning-based retinal vessel segmentation; 3) Enhancement of retinal vessel segmentation out-of-distribution generalization using generative AI; 4) Investigation of vasculature variation within a large glaucomatous cohort. As part of the first pillar, we reviewed the scientific literature and identified retinal microvascular biomarkers which were shown to be associated with CVD endpoints. This resulted in the development of an open-source, automated, and reproducible Python toolbox computing 16 vasculature biomarkers based on blood vessel segmentations. In particular, we introduce new algorithmic methods to estimate tortuosity and branching angles. The second pillar explored the development of deep learning algorithm for the automated segmentation of blood vessel from DFIs in order to provide automated and reproducible input for retinal microvascular biomarkers extraction. We developed and annotated the Leuven-Haifa dataset, comprising 240 DFIs with ground truth blood vessel segmentations. Our deep learning model, LUNet, achieved a DICE score of 83.3, significantly outperforming previously deep learning models and achieving performance comparable to a human medical graduate annotator. However, we observed limited generalization capability, with performance dropping to an average DICE score of 77.0 when evaluated on external datasets. The third pillar addressed the challenge of generalization in medical image segmentation. We proposed Retinal Layout Aware Diffusion (RLAD), a novel conditional diffusion model that generates new images while preserving similar vascular structures. RLAD was used to generate additional training data, leading to increased generalization performance. Additionally, we collected and published a novel composite dataset from diverse sources, totaling 586 DFIs with ground truth blood vessel segmentation. With our newly introduced model, the average DICE score on external datasets improved to 80.7. Furthermore, our model exhibited robust capability in segmenting out-of-distribution retinal fundus images such as those captured using portable devices, low-quality imaging systems, or ultra-wide fields of view—achieving a DICE score of 72.5 compared to 61.1 for the best benchmark. The fourth pillar leveraged the previously developed tools to conduct a large-scale blood vessel analysis study, examining blood vessel variations within a glaucomatous cohort versus a healthy control group. In total, 32,000 images were analyzed, providing robust insights into retinal vascular differences between the two groups. In particular, the analysis revealed significant independent similarities in retinal vascular geometry alterations associated with both advanced age and glaucoma. These findings suggest a potential mechanism of accelerated vascular aging in patients with glaucoma. This thesis presents important advancements in the analysis of retinal microvasculature by developing robust and automated computational tools and methodologies. Overall, this work highlights the transformative potential of deep learning in retinal microvasculature analysis, offering clinicians and researchers powerful tools for vascular health assessment, enhancing segmentation generalization, and enabling large-scale studies with direct implications for translational medical research. Supervised by Associate Prof. Joachim A. Behar and Assistant Prof. Moti Freiman Zoom link: https://technion.zoom.us/j/92105308309