Integrating explainable AI for diabetic retinopathy screening: A comparative and clinically grounded study

Diabetic retinopathy (DR) is a leading cause of vision loss worldwide. However, screening remains limited because of reliance on trained ophthalmologists and associated costs. Advances in artificial intelligence (AI) models, particularly convolutional neural networks (CNNs), have high performance in DR classification. However, when they function as opaque “black boxes” where predictions are generated without insight into the decision-making process, it limits clinical trust and adoption. Enhancing available AI models by integrating explainable AI (XAI) methods will help address these challenges and will allow clinicians to adopt these models in their practices. This study enhances interpretability by integrating XAI methods: Grad-CAM, Grad-CAM++, Score-CAM, and Guided Grad-CAM to state-of-the-art CNN models including DenseNet121, InceptionV3, and ResNet-50. We performed experiments on the APTOS 2019 dataset. The DenseNet121 model achieved the best performance (92.95% Quadratic Weighted Kappa (QWK), 85.29% accuracy). Model evaluation metrics include accuracy, F1-Score, and QWK. We quantitatively evaluated the XAI outputs’ accuracy by computing the lesion-overlap scores of 236 images in the lesion-annotated TJDR dataset. Additionally, we analyzed the computational cost of each method and how that may hinder clinical deployment. The variety of XAI outputs provides comprehensive information about the models’ inner thinking. This study shows how to combine high-performing CNN models with XAI techniques to create interpretable and clinically relevant solutions for DR screening while validating AI models, highlighting the importance of trustworthy outputs in real-world healthcare settings.