Synthetic data-driven diabetic retinopathy diagnosis with explainable AI: a clinically interpretable framework.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a leading cause of blindness, necessitating early and accurate diagnosis.
• Current diagnostic methods face challenges due to limited annotated datasets and the clinical adoption barriers of deep learning models.

Purpose of the Study:

• To develop a novel solution combining synthetic data generation and explainable machine learning for Diabetic Retinopathy diagnosis.
• To overcome limitations of data scarcity and improve clinical adoption of AI in DR detection.

Main Methods:

• Generated synthetic retinal fundus images to create a balanced dataset of 1200 instances for DR severity grading.
• Extracted handcrafted features using HSV color space, GLCM texture analysis, and lesion quantification.
• Implemented a Random Forest classifier with Out-of-Bag validation and integrated SHAP for explainability.

Main Results:

• Achieved 94% accuracy on synthetic data, surpassing established deep learning models like ResNet-50 and EfficientNet-B0.
• SHAP explanations identified clinically relevant features (e.g., exudates), confirming classifier reliability.
• Swarm plot analysis indicated consistent prediction confidence.

Conclusions:

• The proposed framework demonstrates the potential of synthetic data and explainable AI in DR diagnosis.
• Domain adaptation techniques integrating synthetic and real data can further improve clinical transferability and reduce performance gaps.