A robust ensemble-based deep learning framework for automated retinal disease detection.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Automated retinal disease detection models often lack generalizability and accuracy.
• Clinical decision-making requires reliable diagnostic tools for various eye conditions.

Purpose of the Study:

• To develop a robust deep learning framework for automated multi-class retinal disease detection.
• To enhance generalizability and accuracy beyond existing models for clinical application.

Main Methods:

• A novel ensemble model, ResEfficientNetB3, was developed by integrating EfficientNetB3 and ResNet50 architectures.
• Two Kaggle datasets (4217 and 8230 images across 4 and 8 classes) were utilized with data augmentation.
• Models were trained using the Adam optimizer with early stopping and dropout, assessed via cross-validation and cross-dataset validation.

Main Results:

• ResEfficientNetB3 achieved 99.0% accuracy on Dataset 1 and 98.2% on Dataset 2, surpassing individual models.
• Five-fold cross-validation confirmed model robustness (99.0% ± 0.2 and 98.2% ± 0.3).
• Cross-dataset validation demonstrated strong transferability, achieving 94.5-95.8% accuracy.

Conclusions:

• ResEfficientNetB3 effectively combines EfficientNetB3 and ResNet50, yielding superior performance.
• The model demonstrates high accuracy, robustness, and generalization capabilities for retinal disease detection.
• This framework provides a reliable, clinically applicable tool for real-world automated diagnostics.