Ensemble Transfer Learning Model for the Detection of Diabetic Retinopathy Disease

Ishaq Aiche; Youcef Brik; Bilal Attallah; Oussama Bouguerra

doi:10.5935/jetia.v12i59.3423

Authors

Ishaq Aiche LASS, Laboratory of Signal and Systems Analysis, Department of Electronics, Faculty of Technology, University of M’sila, Algeria. https://orcid.org/0009-0004-5133-4745 (unauthenticated)
Youcef Brik Electronic Department, Faculty of Technology, Laboratory of Signal and Systems Analysis, Mohamed Boudiaf University, PO Box 166 Ichebilia, 28000 M'sila, Algeria https://orcid.org/0000-0003-1565-0594 (unauthenticated)
Bilal Attallah Electronic Department, Faculty of Technology, Laboratory of Signal and Systems Analysis, Mohamed Boudiaf University, PO Box 166 Ichebilia, 28000 M'sila, Algeria https://orcid.org/0000-0003-3661-957X (unauthenticated)
Oussama Bouguerra Electronic Department, Faculty of Technology, Laboratory of Signal and Systems Analysis, Mohamed Boudiaf University, PO Box 166 Ichebilia, 28000 M'sila, Algeria https://orcid.org/0009-0001-8564-5291 (unauthenticated)

DOI:

https://doi.org/10.5935/jetia.v12i59.3423

Abstract

Diabetic retinopathy (DR) is a leading cause of vision loss, and early detection is essential for preventing blindness. This study investigates using transfer learning, deep learning and model fusion techniques to improve DR detection. Five pre-trained Convolutional Neural Networks (CNNs) ResNet50, VGG16, VGG19, DenseNet121 and EfficientNet-B7 were evaluated individually and in combination using the APTOS 2019 RGB retinal image dataset. Transfer learning was applied to fine-tune the models for the binary classification of DR and non-DR cases. Four fusion strategies were implemented to exploit the complementary strengths of the individual models. The experimental results demonstrate that all fusion approaches outperformed the standalone CNNs in terms of accuracy, sensitivity, specificity and Area Under the Curve (AUC). Of the individual models, DenseNet121 and EfficientNet-B7 performed best. Fuzzy-based fusion strategies further improved diagnostic accuracy. The best fusion model achieved an accuracy of 98.64% and a recall of 99.19%, demonstrating the robustness of the proposed approach. These findings highlight the potential of integrating transfer learning with advanced model fusion to create automated, reliable and scalable systems for the early screening of Diabetic Retinopathy (DR). Such systems could facilitate timely clinical intervention, enhance patient outcomes, and contribute to more efficient healthcare delivery.