AI-based methods for diagnosing and grading diabetic retinopathy: A comprehensive review

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a major global cause of blindness.
• Early detection and accurate grading of DR are crucial for effective treatment.
• Artificial intelligence (AI), computer vision, machine learning, and deep learning (DL) offer automated solutions for DR diagnosis.

Purpose of the Study:

• To comprehensively review and evaluate AI-based methods for DR detection and classification.
• To analyze studies utilizing fundus photography, OCT, OCT-angiography, and fluorescein angiography.
• To compare characteristics of 23 public DR datasets.

Main Methods:

• Systematic review of 91 studies on AI for DR detection and classification.
• Analysis of AI methods across fundus photography, OCT, OCT-angiography, and multimodal imaging.
• Comparison of dataset characteristics for 23 public DR datasets.

Main Results:

• Deep learning (DL) approaches generally outperform traditional AI methods across imaging modalities.
• Fundus images were the most utilized (81%), followed by OCT (9%), OCT-angiography (6%), and multimodal (2%).
• AI was predominantly used for multi-way classification (62%) compared to binary classification (28%).

Conclusions:

• AI, particularly DL, demonstrates significant potential for automated DR detection and grading.
• Future research should focus on explainable AI, multimodal data integration, and clinical workflow integration.
• Standardized protocols are needed to integrate AI tools into healthcare systems for improved DR management.