Identifying Diabetic Retinopathy in the Human Eye: A Hybrid Approach Based on a Computer-Aided Diagnosis System...

Area of Science:

• Biomedical Engineering
• Ophthalmology
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) diagnosis relies on identifying retinal blood vessel damage.
• Deep learning (DL) models show promise for automated DR detection using medical eye imagery.
• Challenges like imbalanced datasets and annotation errors hinder DL model performance.

Purpose of the Study:

• To compare state-of-the-art deep learning approaches for diabetic retinopathy detection.
• To evaluate methods for addressing class imbalance in DR datasets.
• To identify optimal hybrid modeling strategies for automated DR detection.

Main Methods:

• Utilized three benchmark datasets for diabetic retinopathy analysis.
• Compared various state-of-the-art deep learning techniques.
• Analyzed hybrid modeling including Convolutional Neural Network (CNN) and SHAP (SHapley Additive exPlanations) model derivations.

Main Results:

• Achieved high precision scores across different diabetic retinopathy severity levels: 93% (normal), 89% (mild), 81% (moderate), 76% (severe), and 96% (DR phases).
• Demonstrated the effectiveness of hybrid modeling strategies in improving DR detection.
• Identified specific hybrid approaches that mitigate the impact of class imbalance.

Conclusions:

• Hybrid deep learning models, combining CNN and SHAP, offer robust solutions for automated diabetic retinopathy detection.
• Addressing class imbalance is crucial for enhancing the performance of DL models in DR screening.
• The study provides insights into ideal hybrid modeling for clinical application in DR diagnosis.