Hybrid Deep Learning Model for Diabetic Retinopathy Severity Detection and Classification

Background: A leading cause of vision loss and common complication in patients with diabetes is Diabetic retinopathy. The detection of these complications can be done in the early stages of the disease and can be managed through non-invasive retinal image analysis. However, traditional manual detection methods have low accuracy and high error rates, making it difficult to reliably identify key indicators like microaneurysms, hemorrhages, and exudates. This underscores the need for improved diagnostic methods.Objectives: The objective of this research study is to propose a diagnostic system using deep learning models to improve the accuracy and reliability of diabetic retinopathy detection.Materials and Methods: This study proposes an automated system for diabetic retinopathy detection using deep learning techniques. Two deep learning techniques, EfficientNet-B3 and ResNet18 CNN, were trained on retinal and non-retinal image datasets to identify indicators like microaneurysms, hemorrhages, and exudates.Results: Our experiments demonstrated high accuracy in detecting and classifying diabetic retinopathy. The deep learning model achieved an accuracy of 98.18%, while the verification model reached 99%. These results highlight the system's ability to reliably identify critical features such as microaneurysms, hemorrhages, and exudates.Conclusion: The automated system developed for diabetic retinopathy detection, powered by deep learnng models, significantly enhances diagnostic accuracy and efficiency. Its high performance in identifying key retinal indicators offers a practical solution for early detection and monitoring, especially in resource-limited areas. This approach shows potential for enhancing patient outcomes and alleviating the global impact of diabetic retinopathy by providing accessible and dependable diagnostic support.Keywords: Diabetic retinopathy (DR), Deep learning models, EfficientNet-B, ResNet18, Medical image classification.