Detection and Classification of Diabetic Retinopathy in Fundus Images Using Deep Learning Models

Abstract

This paper explores how deep learning architectures, namely Convolutional Neural Networks (CNN), ResNet and DenseNet can be used to detect and classify diabetic retinopathy (DR) in fundus images. A set of more than 35,000 manual hand-labeled retinal images were used to evaluate the models based on the severity of DR. DenseNet was the most effective one with its accuracy of 92 percent, precision of 90 percent, recall of 88 percent, F1-score 89 percent, and the AUC-ROC 0.94 surpassing ResNet (accuracy 90 percent, precision 88 percent, recall 85 percent, F1-score 86 percent, and the AUC-ROC 0.93) and CNN (accuracy 87 percent, precision 85 percent, recall 83 percent, F1-score 84 percent, and the AUC-ROC 0. These findings reveal the high capacity of DenseNet to correctly identify DR, especially the initial stages and its use in large-scale screening. The paper highlights the uniqueness of deep learning in the automation of DR detection, which is highly effective compared to the old methods that are not only time consuming but also error prone and rely on human expertise. DenseNet performed well because the layers in the network were densely linked, and feature reuse was possible, which increased the accuracy of the classification. This study presents an encouraging method of improving the diagnosis of diabetic retinopathy especially in the environments where there is scarce access to healthcare providers. At the same time, the results highlight the significance of big annotated datasets, the potential of state-of-the-art deep learning models, and how real-time detection systems can bring screening of DR closer to patients, more precise, and scalable in clinical and remote settings.