RetinoDeep: Leveraging Deep Learning Models for Advanced Retinopathy Diagnostics.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss globally.
• Manual DR screening is labor-intensive, subjective, and faces specialist shortages.
• There is a need for scalable, objective, and interpretable automated diagnostic tools for DR.

Purpose of the Study:

• To develop and evaluate deep learning frameworks (RetinoDeep) for automated DR detection and classification across seven severity levels.
• To enhance model transparency and clinical trustworthiness using explainable AI (XAI).
• To improve the accuracy, stability, and generalization of DR screening systems.

Main Methods:

• Proposed four novel deep learning models: EfficientNetB0 with SPCL transformer, ResNet50 with Bi-LSTM, GA-optimized Bi-LSTM, and SHAP-explained Bi-LSTM.
• Trained and evaluated models on 757 augmented retinal fundus images.
• Benchmarked against state-of-the-art models using accuracy, F1-score, and precision.

Main Results:

• The Bi-LSTM model optimized with Particle Swarm Optimization (PSO) achieved superior stability and generalization.
• SHAP visualizations confirmed that learned features align with key retinal biomarkers, enhancing interpretability.
• The proposed models demonstrated improved diagnostic performance compared to baselines.

Conclusions:

• RetinoDeep frameworks, particularly Bi-LSTM with PSO and SHAP, offer a promising approach for automated DR screening.
• The integration of advanced optimization and explainable AI enhances diagnostic accuracy and clinical trustworthiness.
• These systems have the potential for integration into real-world clinical workflows for improved DR management.