Robust segmentation of vascular network using deeply cascaded AReN-UNet

Abstract Retinal vessel segmentation is an essential step for non-invasive diagnosis and analysis of ocular pathologies such as diabetic retinopathy, glaucoma, etc. Although several deep learning networks have been implemented for segmenting vascular maps, still further modification can be carried out on the existing deep learning networks for precise segmentation of vascular maps. This paper presents a novel cascaded AReN-UNet (Attention Residual U Network), driven by the integration of attention and residual modules. The proposed network is implemented by cascading two deep learning networks of depth 4. In the second network, each encoder receives the feature maps from the previous convolutional block. In addition to this, the feature maps of a respective convolutional block of the preceding network are also fed as input to the convolutional block of the second network. Furthermore, aggregated residual and attention modules in the cascaded AReN-UNet are used to improve convergence and stability of the network which eventually reduces the vessel breakdowns in the vascular map. The proposed model is trained and evaluated on different datasets such as DRIVE, CHASE_DB1, and one locally collected dataset. The proposed network illustrates the state-of-the-art performance by achieving an accuracy, F1 score, sensitivity, specificity, and Area Under the Curve (AUC) of 96.96%, 82.63%, 83.68%, 98.35%, and 98.67% respectively on the DRIVE dataset and 97.70%, 82.01%, 85.60%, 98.35%, and 99.01% respectively on the CHASE_DB1 dataset.