Geometric Self-Supervised Learning: A Novel AI Approach Towards Quantitative and Explainable Diabetic Retinopathy...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness in working-age adults.
• Early detection of DR is critical for preventing vision loss but faces challenges with manual methods.
• Automated detection of DR lesions is necessary to overcome limitations of manual analysis.

Purpose of the Study:

• To develop and validate an annotation-free deep learning strategy for automatic detection of exudates and bleeding spots.
• To evaluate the strategy on both color fundus photography (CFP) and ultrawide field (UWF) retinal images.
• To assess the model's performance without requiring manual annotations for training.

Main Methods:

• Utilized three cohorts: two CFP (Kaggle-CFP, E-Ophtha) and one UWF.
• Developed an algorithm using a U-Net model trained on contrast fields derived from retinal images.
• Evaluated model performance using sensitivity and false positive rates on independent test sets with manual annotations.

Main Results:

• Achieved high sensitivities for DR lesion detection on CFP images (e.g., microaneurysms 91.5%, hemorrhages 92.6%).
• Demonstrated variable but high sensitivity for bleeding and exudate detection on UWF images, with performance varying by lesion size.
• Reported varying false positive rates, generally higher for smaller lesions across both image types.

Conclusions:

• The study successfully demonstrates the feasibility of an annotation-free deep learning approach for DR lesion detection.
• This method can effectively train neural networks to identify and segment DR-related lesions.
• The developed strategy offers a promising avenue for automated DR screening and diagnosis.