Signal, Image and Video Processing Signal Processing
Automated segmentation of retinal vessels using the Vessel-Net + fully convolutional neural network
Abstract
Computer-aided diagnostic (CAD) systems for color fundus images are vital for early detection of retinal and systemic diseases. Despite advances in deep learning, segmentation remains challenged by limited annotations, large structural variability of vessels, and cross-dataset discrepancies leading to inconsistent performance, especially on thin vessels and complex bifurcations. We address these issues with a Vessel-Net + architecture: a U-shaped encoder–decoder with UNet ++ dense skip aggregation, augmented by Self-Adaptive Attention (SAA) blocks. Each SAA block jointly models pixel-wise (pooled attention) and channel-wise (correlation-based) dependencies, and applies lightweight channel and spatial gating; deep supervision with a BCE + Dice objective further stabilizes multi-scale learning. This design enhances long-range context, suppresses distracters, and preserves fine details critical for tiny vessels and branch regions. Our proposed model achieves 96.85% accuracy on DRIVE and 97.44% on CHASE-DB1 and Dice score 80.20% and 80.23% respectively surpassing other models. Qualitative analysis shows closely align with annotations on vessels structure emphasizing robustness and practical deployment in fundus analysis. The code will be available at: https://github.com/irshadkhalil23/vesselNet -.