<![CDATA[Background of study: Diabetic Retinopathy (DR) is a severe microvascular complication of diabetes mellitus that can lead to vision loss if not detected early. With over 93 million individuals affected globally, the need for accurate and efficient diagnostic systems has become urgent. Traditional screening methods depend on manual interpretation of fundus images by ophthalmologists, which is time-consuming and prone to subjectivity.Aims: This research seeks to create and assess a deep learning diagnostic model designed to reliably identify the severity levels of diabetic retinopathy using retinal fundus images. The research also explores model interpretability using Shapley Additive Explanations (SHAP) to increase transparency in AI-assisted medical diagnosis.Methods: Convolutional Neural Networks (CNNs) were implemented using transfer learning with pretrained architectures such as ResNet50 and InceptionV3. The EyePACS dataset, containing images categorized into five DR severity levels, was used for model training. Preprocessing techniques, including contrast enhancement, histogram equalization, and data augmentation, improved image quality and model generalization. The models were optimized with the Adam and assessed through accuracy, precision, recall, F1-score, and AUC. Additionally, SHAP analysis was employed to interpret and illustrate the model’s predictions.Results: The proposed CNN-based model achieved 98.5% accuracy, with a sensitivity and specificity of 0.99, demonstrating strong performance across multiple DR stages. Comparison with existing studies revealed a notable improvement in diagnostic accuracy. SHAP visualizations confirmed that critical retinal features such as microaneurysms, hemorrhages, and cotton-wool spots were key predictors influencing model decisions.Conclusion: The findings validate the efficacy of deep learning, particularly CNNs, in enhancing early detection and classification of diabetic retinopathy. The integration of SHAP interpretability bridges the gap between AI predictions and clinical trust, making this approach a promising tool for large-scale automated DR screening and supporting ophthalmologists in timely diagnosis and treatment.]]>
