A Novel Attention-Based U-Net Approach for Retinal Image Segmentation Across Diverse Datasets

Retinal vessel segmentation is essential for the early detection and diagnosis of ophthalmic and systemic diseases. This study aims to enhance the segmentation performance of the widely used U-Net architecture through novel integrations of attention mechanisms, specifically Attention U-Net (Att-UNet) and Spatial Attention U-Net (SA-UNet). To systematically evaluate these methods, comparative experiments were conducted across three benchmark retinal datasets: DRIVE, STARE, and CHASE-DB1. Models were trained for 80 epochs using a hybrid loss combining Binary Cross-Entropy and Dice loss functions, and performance was rigorously assessed using metrics such as Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and Accuracy. The experimental results indicate that attention-enhanced variants consistently outperform standard U-Net in terms of segmentation quality. Specifically, Att-UNet achieved the highest accuracy (97.37%) on DRIVE and STARE datasets, providing optimal segmentation precision, whereas SA-UNet demonstrated superior computational efficiency with the lowest training times. These findings highlight the dataset-dependent benefits and computational trade-offs associated with attention mechanisms. Overall, this study confirms that integrating novel AI techniques, such as attention mechanisms, significantly improves retinal vessel segmentation, showcasing potential for real-world engineering and clinical applications.