From pixels to attribution reports: Diabetic retinopathy grading with CNN–transformer ensembles and vision–language models

Diabetic retinopathy (DR) grading from retinal fundus images remains challenging due to class imbalance, ordinal disease severity, and visual similarities between adjacent stages. This study presents an integrated artificial intelligence framework combining convolutional neural networks (CNNs), Vision Transformers, ensemble learning, explainable AI, and vision–language models (VLMs) for five-class diabetic retinopathy grading and non-diagnostic attribution reporting. Six deep learning architectures were evaluated using the APTOS 2019 dataset under stratified five-fold cross-validation. Multiple ensemble strategies, including hard voting, weighted soft voting, stacking, and hybrid class-level fusion, were investigated to improve classification robustness. ResNet-50 and ConvNeXt-Tiny demonstrated strong individual performance, while weighted soft voting achieved the most consistent ensemble results. External validation on Messidor-2 highlighted generalization challenges associated with cross-dataset distribution shifts. To improve model transparency, Grad-CAM++ visual attribution maps were combined with safety-constrained textual summaries generated by vision–language models. The study also examined computational efficiency and deployment considerations. The findings demonstrate the potential of ensemble learning and multimodal explainability for developing more transparent AI-assisted diabetic retinopathy grading systems while emphasizing the importance of external validation, uncertainty assessment, and clinical evaluation before real-world deployment.