Proposed Model for the Detection of Diabetic Retinopathy Using Convolutional Neural Networks

Abstract

Diabetic retinopathy, an ocular complication associated with diabetes, is a major cause of vision loss if not treated early. This study aims to identify Diabetic Retinopathy using Convolutional Neural Networks. The proposed Methodology consists of four phases: Obtaining the dataset, Preprocessing, Model Training and Evaluation. In the proposed method, DR detection is performed using the IDRID labeled retinal image dataset, implementing and training the pre-trained models VGG-19, ResNet-50 and Inception-V3. The results highlight that the VGG-19 model achieves remarkable performance with accuracy, precision and recall of 95.64%, 92.98% and 99.64% for binary classification. Although the performance achieved by the other models ResNet-50 and Inception-V3 show intermediate performance, they show lower accuracy, indicating difficulties in classification. In summary, VGG-19 stands out as an effective option to identify DR, while Inception-V3 and ResNet-50 present different performances, pointing out areas of improvement for future research. These results underline the relevance of CNN architectures in the detection of Diabetic Retinopathy. Despite certain limitations, such as unbalanced data, quality and availability of retinal images, the findings demonstrate the great ability of CNN models to contribute to the understanding and improvement of diagnostic methods in this medical area. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.