Detection of diabetic retinopathy 5-stages by denseNet, a CNN model

Diabetes Mellitus, which is brought on by elevated blood sugar levels, causes Diabetic Retinopathy (DR), an incurable eye condition that damages the retina. The blood vessels that nourish the retina enlarge, deform, and get obstructed as a result of this condition. The International Classification of Diabetic Retinopathy (ICDR) Scale classifies Diabetic Retinopathy (DR) into five categories depending on the Severity of the DR. The five classes are No DR, Mild, Moderate, Severe, and Proliferative DR. Deep learning has shown promising diagnostic technique by utilizing various medical images since the 1990s, with low price and show high diagnostic accuracy via various artificial intelligence (AI) based techniques. In this experiment, we employed a deep Convolutional Neural Network (CNN), DenseNet, for the automatic categorization of eye disorders utilizing fundus images. The proposed model has been optimized with different hyperparameter tuning methods. Due to speed of execution, generalization, efficiency, and architectural size, DenseNet is employed here. For this experiment, in addition to the public data sets, 1491 color fundus images obtained from the Department of Ophthalmology, IGM Hospital, Tripura, India, are also utilized. The proposed model’s efficiency is measured by various performance factors such as accuracy, F1Score, precision, recall, specificity, Receiver Operating Characteristics (ROC) curve of each class, Confusion Matrix, micro-average ROC curve, and macro-average ROC curve.