LF-UNet: An Attention-Based U-Net for Retinal Vessel Segmentation

Automated retinal vessel segmentation serves as a significant aid to the clinical practice of ophthalmologists. Through segmenting the vascular structures in fundus images, physicians can observe the morphology and distribution of blood vessels more easily to detect and diagnose ocular diseases. However, due to the complex structure of the retinal vascular system, the conventional U-network cannot extract tiny vascular features. Besides, the direct connection of low-level features with high-level features results in the underutilization of information. To address these challenges, we propose a new U-shaped network (LF-UNet) for retinal vessel segmentation. The application of large kernel attention enables our network to learn the difference between local vascular features and global features. The feature fusion module is designed to adjust the weights of the input feature maps adaptively, enabling the network to fully utilize features from different levels. We validate the LF-UNet on three public datasets, and the experimental results demonstrate the segmentation performance of this network.