Research Spotlight: Deep Learning for Diabetic Retinopathy Detection Using EfficientNet and Swin Transformer

Diabetic Retinopathy (DR) is a leading cause of preventable blindness, especially in underserved communities. This research tackles the challenge of early and scalable DR detection using artificial intelligence. The study presents a rigorous comparison between two cutting-edge deep learning architectures—EfficientNet-B0 (a convolutional neural network) and the Swin Transformer (a vision transformer)—for automated DR grading using the APTOS 2019 retinal fundus image dataset.

Key Findings:

• Swin Transformer achieved higher overall diagnostic accuracy (QWK: 0.91), particularly excelling at detecting severe DR stages like proliferative DR.

• EfficientNet-B0 demonstrated remarkable precision (98%) for early-stage DR (mild NPDR), with high speed (55 FPS) and low computational demand—ideal for mobile and edge deployments.

• Grad-CAM visualizations revealed architectural differences: Swin-T analyzed global patterns (e.g., vessel structures), while EfficientNet focused on localized features (e.g., microaneurysms).

Clinical Implications:

• Use EfficientNet in high-volume, low-resource screening environments (e.g., rural health centers).

• Deploy Swin-T in tertiary care hospitals for accurate detection of advanced cases and referral decisions.

• Highlights the need for hybrid AI models that combine global and local feature reasoning.

  
  

“This work bridges the clinical and technical dimensions of AI, offering a path forward for equitable, scalable, and interpretable diabetic retinopathy screening systems.”

Technical Insights:

• Dataset: 3,662 annotated images from the APTOS 2019 competition.

• Techniques: Contrast enhancement, Frangi filtering, Grad-CAM explainability, 5-fold stratified cross-validation.

• Performance Metrics: QWK, F1-score, ROC-AUC, inference time, and VRAM usage.

Challenges & Future Scope:

• Tackled issues like class imbalance and lack of diversity in datasets.

• Advocates for multi-ethnic model validation, real-time explainability, and federated learning for ethical and inclusive AI deployment.

• Proposed hybrid CNN-Transformer architectures to optimize both accuracy and efficiency.