What question did this study set out to answer?

This research aims to enhance diabetic retinopathy detection through a hybrid deep learning architecture that leverages various models.

June 11, 2026Scientific Reports0 citationsOpen Access

Hybrid deep ensemble architecture for robust diabetic retinopathy classification: leveraging transfer learning and CNN-transformer synergy

YMYoussef MaaodGalala University TATamer AbuhmedSungkyunkwan University EAEslam AmerUniversity of Portsmouth

Key Points

This research aims to enhance diabetic retinopathy detection through a hybrid deep learning architecture that leverages various models.
Utilized APTOS 2019 dataset to evaluate 23 pre-trained models for DR classification.
Fine-tuned top models and created hybrid architectures combining CNNs and vision transformers.
Implemented soft voting architecture for final performance assessment.
Achieved 93.10% accuracy with model performance metrics indicating strong predictive capabilities.
ROC AUC reached 99.22%, demonstrating excellent discrimination ability.
F1-score of 93.07% indicated high precision and recall in DR classification.

Abstract

Diabetic retinopathy (DR) is still one of the main reasons for vision loss worldwide, especially in places where people do not have easy access to regular eye checkups. Early and accurate disease detection is important to avoid permanent damage, but traditional methods are slow and sometimes inconsistent. This study proposes a deep learning framework that combines convolutional neural networks (CNNs), vision transformers, transfer learning, and ensemble techniques to improve DR detection. We used the APTOS 2019 dataset and tested the capabilities of 23 different pre-trained models. Then, we fine-tuned the top models and designed hybrid architectures by combining the best-performing CNNs and transformers in parallel and sequential ways to capture both image spatial features in short and long contexts. The best performance came from combining the top sequential hybrid models using the soft voting architecture, where we got an accuracy of 93.10%, ROC AUC of 99.22%, and F1-score of 93.07%. The optimized model showed that mixing different models and using ensemble methods can lead to better and more stable DR detection decisions. Our approach is a step toward building a reliable and automated system that could help doctors in real-world settings.

Ask AI

Helpful

Bookmark

View Full Paper