Transfer Learning-Based Stack Ensemble Deep Learning Approach to Predict the Severity of Diabetic Retinopathy

Diabetic retinopathy (DR) is the most commonly occurring eye disorder and the main reason underlying blindness in diabetics all around the world. Many technologies have emerged today for the accurate diagnosis of DR at an early stage. Of these, deep learning (DL) is one of the most effective methods. This research focuses on the prediction of DR severity into five classes, Normal, Mild, Moderate, Severe, and Proliferative DR (PDR), using pre-trained models. Transfer learning using models, such as EfficientNetB0, MobileNet, and Xception, were implemented with customization. Further, the Stack Ensemble model was applied to combine the predictions of all these pre-trained models using meta classifiers, such as Random Forest and Extra Trees Classifier to grade the DR severity. The performance the proposed model was evaluated on the KAGGLE and the Asia Pacific Tele-Ophthalmology Society (APTOS) retina datasets. The final outcome revealed that the proposed model outperformed state-of-the-art pre-trained models, with an accuracy of 0.96 and 0.97 on the KAGGLE and APTOS datasets, respectively.