Automated Eye Disease Detection of Diabetic Retinopathy Using Artificial Intelligence on Fundus Images

Diabetic retinopathy (DR) is a diabetes-related eye disease that is a major contributor to blindness in diabetic people. This paper introduces an AI-powered automated method for DR identification utilizing the publicly available MESSIDOR dataset. The dataset contains 1200 fundus pictures, 654 DR cases spanning different severity levels, and 546 normal photos. The approach included a number of preprocessing processes to prepare the images for disease feature extraction, including scaling, cropping, noise reduction, CLAHE-based contrast enhancement, and segmentation. To address class imbalance, SMOTE was applied, ensuring uniform class distribution. A ResNet50 deep learning model was proposed and fine-tuned using hyperparameter optimization, including the Adam optimizer and categorical cross-entropy loss. A model's performance was evaluated and compared with existing models, such as InceptionV3, KNN, AlexNet, DenseNet-201, and CNN, using accuracy, sensitivity, and specificity metrics. The proposed ResNet50 model outperformed others, achieving a remarkable accuracy of 99.02 %, sensitivity of 98.32 %, and specificity of 99.04 %, demonstrating its strong potential in accurately detecting and classifying DR stages for clinical applications.