Robustness of Retinal Vessel Segmentation to Image Quality Degradations

Abstract

Retinal vessel segmentation is essential for automated fundus image analysis, but its reliability depends strongly on image quality. Because benchmark datasets contain mostly high-quality images, the robustness of vessel segmentation methods under realistic degradations remains insufficiently characterized. We investigate this issue using controlled synthetic degradations and two U-Net-based models, one trained on mixedquality images and the other exclusively on high-quality images. The effects of blur, contrast reduction, and non-uniform illumination are evaluated across multiple severity levels, together with no-reference image quality assessment (NR-IQA) metrics as candidate indicators of segmentation reliability. The results reveal degradation-specific failure modes. Blur and severe compound degradations substantially reduce performance, mainly through recall loss, whereas contrast reduction causes a sharp performance drop in both models. Although several NR-IQA metrics are associated with performance under specific degradations, no metric shows consistent alignment with segmentation reliability across conditions. These results motivate task-aware quality control for reliable clinical deployment.