An Interpretable Ensemble Deep Learning Model for Diabetic Retinopathy Disease Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a significant complication of diabetes, potentially leading to irreversible blindness.
• Timely diagnosis of DR is critical for effective patient management and prevention of vision loss.
• Current diagnostic methods may face challenges in early and accurate detection.

Purpose of the Study:

• To develop an automated system for accurate, image-level detection of diabetic retinopathy.
• To enhance the performance and robustness of DR detection models through ensemble methods.
• To provide visual explanations for DR detection results using interpretable AI techniques.

Main Methods:

• Utilized multiple pre-trained deep learning models for image-level diabetic retinopathy detection.
• Integrated diverse deep learning models using the Adaboost algorithm to minimize individual model bias.
• Employed weighted Class Activation Maps (CAMs) to visualize and localize suspected retinal lesions.
• Implemented data augmentation with eight image transformation techniques to increase fundus image diversity.

Main Results:

• The proposed ensemble deep learning system demonstrated superior performance compared to individual models.
• The system exhibited enhanced robustness in detecting diabetic retinopathy.
• Weighted CAMs effectively highlighted potential lesion areas, aiding in result interpretation.

Conclusions:

• The integrated deep learning approach offers a robust and accurate solution for automated diabetic retinopathy detection.
• Ensemble methods and data augmentation significantly improve diagnostic performance.
• Interpretable AI techniques like CAMs can enhance clinical trust and understanding of automated diagnostic systems.