Abstract: In the clinical diagnosis of retinal diseases, the use of deep learning techniques for retinal image segmentation is crucial. However, existing methods have not sufficiently considered the importance of receptive fields in high-resolution image feature extraction. Additionally, retinal images are influenced by various factors such as lighting variations, noise, and individual patient differences. To address this, this paper proposes a deep learning model (FPN-Unet) that combines Feature Pyramid Networks (FPN), U-Net, and Channel attention mechanisms. At the same time, Boundary Label Relaxation (BLR) is introduced to better capture the edges of the target. FPN effectively handles multi-scale feature information, while the symmetric structure of U-Net enhances the utilization of contextual information. By introducing attention mechanisms, the model can focus more on important vascular regions and suppress background noise, thereby improving segmentation accuracy. We validated the model on several public datasets, demonstrating its effectiveness and comparing it with existing segmentation models. The experimental results show that FPN-Unet significantly improves retinal image segmentation accuracy, achieving an accuracy of 0.9877 and 0.9732, and sensitivity of 0.9085 and 0.8324 on the high-resolution Fives and HRF datasets, respectively, outperforming traditional retinal image segmentation methods. Moreover, it excels in handling small targets and complex backgrounds. These results suggest that the combination of FPN、U-Net and Channel Attention Mechanism (CAM) for retinal image segmentation holds great promise for providing more accurate and reliable support in clinical ophthalmology applications.