Generalization of retinal vessel segmentation models to real patient populations

Determine how deep learning–based retinal blood vessel segmentation models generalize to real patient populations across differences in retinal diseases (e.g., diabetic retinopathy, age-related macular degeneration, glaucoma), image quality (illumination and color, blur, and contrast), and dataset source domains.

Background

Public retinal vessel segmentation datasets such as DRIVE and CHASEDB1 are small, vary in image quality and labeling, and contain few diseased cases, raising concerns about overfitting and limited external validity. As a result, the extent to which models trained on these datasets generalize to diverse real-world clinical populations—spanning different diseases, acquisition setups, and quality conditions—has been a key uncertainty.

The FIVES dataset provides larger scale and higher quality annotations, enabling a more rigorous assessment of cross-dataset and cross-disease robustness. Nevertheless, the underlying question of generalization behavior across disease-induced and acquisition-induced domain shifts is explicitly identified as unclear.