Diabetic retinopathy feature utilisation in 54 convolutional neural networks: a heatmap analysis

Heatmap analysis across 54 convolutional neural networks for referable diabetic retinopathy shows most attention to the central/temporal macular region, with exudates and cotton wool spots highlighted most often.

• Why it matters: Clinicians need to know whether AI relies on clinically meaningful retinal findings, because uneven feature use can limit performance, especially in atypical presentations.
• What they did: Fifty-four models trained on DDR, BRSET, and Kaggle were evaluated on 749 annotated images using Grad-CAM, Integrated Gradients, LIME, and SHAP, then compared to pixel-level feature annotations.
• The result: Exudates and cotton wool spots had the highest heatmap coverage, while venous beading and neovascularisation at the disc were poorly covered for binary referable classifiers; ICDR grading models showed higher feature coverage and architecture-dependent variability.

AI summary of the abstract · may contain errors · not medical advice · read the paper.