GDFP-Net: A Parallel Feature Fusion-Based Network For Automatic Vessels Segmentation From Fundus Images

In clinical diagnosis, the curvature, shape and spatial distribution of retinal vessels are important criteria for diagnosing diabetes. Although deep convolutional neural networks (CNNs) have achieved good results in retinal vessel segmentation, the segmentation accuracy of thin vessels is still limited due to insufficient extraction of detailed features. To alleviate the problem of detail information loss, this paper proposes a global detail feature processing network (GDFP-Net), which consists of a global segmentation network (GSN), a detail enhancement network (DEN), and a feature processing network (FPN). Among them, GSN and DEN work in parallel to enhance the feature extraction capabilities of thick and thin vessels segmentation.