SAF-NET: Split Attention Fusion Network for retinal vessel segmentation from fundus images

Regular analysis of retinal vessels can alert the occurrence of many retinal pathological disorders like Diabetic Retinopathy (DR) and Glaucoma in their early stages. In view of this, a robust deep learning model, called the Split Attention-Fusion Network (SAF-Net) is implemented using two subnetworks to detect and generate two probabilistic maps for arteries and veins respectively. The first subnetwork is a modified version of U-Net and acts as a backbone for SAF-Net and extracts the most pertinent feature maps to detect the major blood vessels. Further, to detect the tiny vessel and branch junction pixels, an effective residual-like structured block named Multi-Scale Feature Fusion (MSFF) is proposed. The feature maps of both subnetworks are concatenated to produce the final segmented maps of retinal blood vessels with high precision. The efficacy of proposed SAF-Net model is evaluated using two standard publicly available datasets (Dualmodel2019 and AV datasets) and one local real-time dataset.