Deep Learning for Diabetic Retinopathy Detection: A Survey on Model Architectures, Datasets, and Evaluation Metrics

Abstract

Diabetic retinopathy, also known as DR, is a serious complication of diabetes that, if not recognized and treated in a timely manner, can result in vision impairment and even blindness. As a result of the dramatic progress that has been made in deep learning, automated DR detection has emerged as a potentially fruitful topic of research. The purpose of this study is to provide a comprehensive assessment on the most recent deep learning models, datasets, and evaluation metrics that are utilized in DR detection. A number of distinct model architectures, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Generative Adversarial Networks (GANs), and hybrid models, are discussed in this article. We highlight the advantages and disadvantages of each model architecture, as well as the applications that may be found in various stages of DR severity detection and grading. A comprehensive analysis of publicly accessible datasets, such as Kaggle's EyePACS, Messidor, and DDR, is also included in the survey. The analysis focuses on the distinctive characteristics, dimensions, and quality of fundus images that are present in these datasets. In order to provide a comprehensive perspective, we will cover the performance evaluation criteria, which include accuracy, sensitivity, specificity, and the Area Under the Receiver Operating Characteristics (AUROC) curve. These criteria are essential for determining the clinical relevance of these models. In conclusion, the study identifies major issues, such as class imbalance and interpretability, and offers future research areas to enhance the effectiveness and reliability of deep learning-based DR detection systems. These challenges are discussed in the paper.