Evaluation benchmark for natural robustness evaluation of retinal vessel segmentation models

A dataset contains benchmark images for natural robustness evaluation of deep learning models for retinal vessel segmentation. The dataset consists of three mainstream retinal vessel segmentation datasets: DRIVE, STARE, and CHASE_DB1.

For each dataset are provided:

• images - directory containing fundus images augmented using AugOOD tool for fast image augmentation for OOD robustness evaluation.
• labels - directory with labels that correspond to the images.
• masks - directory with FoV masks that correspond to the images.

The benchmark is used in the paper Robustness of deep learning methods for ocular fundus segmentation: Evaluation of blur sensitivity to evaluate natural robustness of a portfolio of deep learning models for retinal vessel segmentation from fundus images.