Hybrid Approach for Retinal Vessel Segmentation: Bridging Retinal and Cerebral Microcirculation

Cerebrovascular diseases, including stroke and transient ischemic attacks, are major causes of mortality and disability. Retinal vasculature offers a unique window into cerebral microcirculation, enabling non-invasive detection of vascular anomalies. This paper presents a hybrid segmentation framework combining U-Net and Generative Adversarial Networks (GANs) to take out retinal vessels from fundus images. The U-Net architecture generates an initial vessel map by preserving both low-level and top-level features, while the GAN-based module refines the segmentation to ensure globally coherent vessel structures. Through adversarial training, the discriminator drives the generator to produce more realistic outputs, mitigating issues such as false positives and broken vessel segments. The proposed method demonstrates high accuracy on public datasets, achieving a Dice coefficient of up to 0.92, alongside strong sensitivity and specificity. We also analyse vessel density, branching angles, and tortuosity, which are key indicators for cerebrovascular risk assessment. Additionally, we explore the potential of longitudinal monitoring to capture disease progression and propose extending our framework to artery–vein classification for finer vascular biomarkers. Our proposed work can effectively support clinicians and researchers in detecting microvascular changes, potentially improving the early diagnosis and management of cerebrovascular diseases. Future work includes multi-modal data fusion integration.