A Hybrid CNN-LSTM Framework for Early Detection of Diabetic Retinopathy from Retinal Fundus Images

Diabetic Retinopathy is a severe eye disease that can be prevented if detected early enough. Although manual diagnosis is difficult for early-stage DR, it exhibits subtle retinal changes. In order to tackle this problem, A hybrid CNN – LSTM framework is introduced to perform DR detection in an early stage. Retinal fundus images are processed by CNNs for feature extraction, while LSTM networks are used to capture temporal behavior of the disease making a more precise classification possible. The novelty is Temporal Progression modeling since it is used to track changes in the retina over time that helps to find early signs of diabetic. This proposed model will be trained on benchmark datasets, such as Kaggle APTOS 2019 and Messidor, in order to robustly perform under different data sources. Data preprocessing continues to learn by performing more advanced data augmentation with the purpose of eventually counteracting overfitting for increased generalization. The accuracy shows 92. 50% in Hybrid CNN-LSTM. This hybrid CNN-LSTM is better than any single use CNN model because it records performance generally and particularly by generalizing to different levels of DR severity. Therefore, lessens the requirement for specialists in resource-poor settings, with possible incorporation into screening programs. Such a model provides a cheap and scalable mechanism to detect DR and result in better patient outcomes and timely intervention leading to prevention of vision loss.