How does AI detect diabetic retinopathy from retinal photos? A heatmap analysis of 54 deep learning models

The research examined 54 artificial intelligence systems based on 27 different architectures developed to detect diabetic retinopathy from retinal photographs. Models were trained on data from DDR, BRSET, and Kaggle databases and subsequently analyzed 749 images with diabetic lesion annotations. Most models achieved acceptable performance with AUROC greater than 0.8. Heatmap analysis revealed that models focused primarily on the central retinal area (macula) while neglecting the optic disc. Models most frequently identified exudates and cotton wool spots, while venous beading and neovascularization at the disc showed poor coverage. Models classifying according to the international grading scale demonstrated better detection of all features. The research showed that individual AI models do not uniformly use all diabetic signs when detecting retinopathy, which may limit their performance in unusual cases.