New Weighted Voting Ensemble Strategies for Retinal Vessel Segmentation

Retinal vessel segmentation is essential for the diagnosis and monitoring of ocular and systemic diseases, yet manual delineation remains time-consuming, labor-intensive, and prone to inter-expert variability. We propose a weighted ensemble learning framework that integrates five UNet variants (Unet++, Attention U- Net, SA-UNet, Res2UNet, and Residual Attention U-Net) trained on the DRIVE dataset. Ensembles of two to five models were constructed using four selection strategies: (1) ranking by Dice coefficient, (2) by Sensitivity, (3) by Specificity, and (4) a multi-criteria decision-making method (TOPSIS). Performance was evaluated with standard metrics, qualitative comparisons, and validated using the Scott–Knott test and Borda Count ranking. Results show that ensembles consistently outperform single models. The best configuration, combining three learners under the MCDM strategy, achieved a Dice score of 0.8871, significantly surpassing all base learners. These findings demonstrate that weighted ensembles guided by balanced selection criteria provide robust and accurate vessel segmentation, with strong applicability to other imaging modalities and clinical decision support systems.