Accurate segmentation of retinal vessels is essential for the diagnosis of various ocular diseases.Existing convolutional neural network-based retinal vessel segmentation networks often suffer from microvascular discontinuities and loss of detailed information due to the mismatch between convolutional kernels and vascular morphology

as well as rough feature fusion mechanisms.To solve the above problems

a structure-aware progressive refinement network(SPR-UNet)is proposed.In the encoder

the network incorporates a vessel residual block(VesselRes Block)which integrates asymmetric and multi-scale convolutions to enhance the ability to capture feature of vascular orientation and morphology.Meanwhile

a progressive multi-scale context-aware attention(PMCA)gate is constructed to replace the traditional skip connections

realizing the dynamic filtering and refinement of cross-layer features.Experimental results demonstrate that SPR-UNet achieves superior performance on the DRIVE and CHASE_DB1 datasets

outperforming comparative algorithms in the F1-score and area under the curve(AUC)(DRIVE:82.85%/98.57%;CHASE_DB1:82.39%/98.21%).This superiority stems from the model’s targeted optimization for the feature extraction and fusion process

which effectively enhance segmentation accuracy and continuity

and improve the balance between sensitivity and specificity.