Accurate segmentation of retinal blood vessels is crucial for automated detection of diabetic retinopathy (DR), a leading cause of vision impairment worldwide. Although UNet and its variants are widely used for medical image segmentation, their comparative robustness under noisy image conditions remains underexplored. This study systematically evaluates five deep learning-based UNet architectures – UNet, Dense UNet, Attention UNet, Inception UNet, and Residual UNet – using three benchmark retinal image datasets. A uniform preprocessing and augmentation pipeline is applied, and all models are optimized using the Dice loss function with the Adam optimizer. Also, to simulate real-world imaging challenges and assess the noise resilience of these models, salt-and-pepper and Gaussian noise are artificially introduced. The architectures are benchmarked across multiple metrics, including accuracy, sensitivity, specificity, Dice coefficient, and Intersection over Union (IoU). Among the variants, Attention UNet consistently achieved superior performance, particularly in highlighting fine vascular structures and managing inter-image variability. These findings demonstrate that UNet-based deep learning models, especially those incorporating attention mechanisms, provide a robust and scalable approach for retinal blood vessel segmentation.