A U-Net Framework with Dice Loss for High-Precision Retinal Vessel Segmentation

Retinal vessel segmentation is essential for diagnosing ocular diseases such as diabetic retinopathy and glaucoma. In this study, a U-Net-based segmentation framework is developed to address the challenges of capturing intricate vascular structures and dealing with class imbalance. The encoder-decoder network leverages skip connections to preserve spatial details, while Dice Loss directly optimizes the overlap with annotated masks, thus improving segmentation accuracy in sparse target regions. Experiments on real-world fundus images demonstrate that the method effectively reconstructs vessel morphology, with minor errors concentrated around vessel boundaries. These findings highlight the framework's clinical potential for aiding disease screening and pathology analysis. Future work may involve more diverse datasets and further enhancements to model robustness.