Diabetic Retinopathy Grading

The field studies how to assign ordered severity levels to diabetic retinopathy, usually from color fundus photographs and sometimes from OCT or OCT angiography. Common approaches include convolutional networks, vision transformers and foundation models, with ordinal-learning methods that reflect the progression between grades; lesion detection, anatomical graphs and multimodal fusion aim to make predictions more informative. Models are compared on datasets such as APTOS, EyePACS, Messidor and DDR, using measures such as quadratic weighted kappa, sensitivity and performance across datasets. A central challenge is that strong results on one dataset may not transfer to different cameras or populations, while sparse lesion labels and uncertain predictions complicate safe screening and referral.