Automated seven-stage diabetic retinopathy grading using optimized deep networks through systematic hyperparameter tuning

Diabetes is a systemic disease that can lead to various pathological changes in the eye. One of these changes, diabetic retinopathy, results from damage to the retinal tissue responsible for light perception and image transmission. High blood glucose levels can damage retinal capillaries, as in other vascular structures, thereby impairing visual function. Early detection of diabetic retinopathy is essential for effective treatment. Accurate classification of disease stages directly affects treatment timing and effectiveness, thereby improving prognosis. Early and accurate staging significantly contributes to preventing disease progression, reducing the risk of complications, and preserving patients’ quality of life. In this study, a comprehensive deep learning-based approach was developed to classify the seven stages of diabetic retinopathy using fundus images. The proposed methodology integrates CLAHE (Contrast Limited Adaptive Histogram Equalization) preprocessing to enhance image contrast and reveal subtle pathological features, followed by systematic data augmentation to improve dataset diversity and model generalization. The Optuna optimization framework was employed to systematically identify the best-performing pre-trained deep learning architecture, which was determined to be NASNetLarge. Hyperparameter tuning using Grid Search optimization further refined the model configuration, achieving a classification accuracy of 98.39% with 5-fold cross-validation, demonstrating robust performance (98.50 ± 0.21%). Furthermore, saliency map visualization confirmed that the model focuses on clinically relevant anatomical structures and pathological features, such as microaneurysms, hemorrhages, and vascular abnormalities, thereby enhancing interpretability and trustworthiness for potential clinical deployment. The results demonstrate that the proposed preprocessing, systematic model selection, and hyperparameter optimization techniques significantly improve classification performance and clinical reliability in diabetic retinopathy diagnosis.