Two Eyes Are Better Than One: Exploiting Binocular Correlation for Diabetic Retinopathy Severity Grading

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• Current deep learning models for DR grading often analyze only one eye, ignoring binocular correlations.
• The variability in DR symptoms complicates diagnosis and treatment.

Purpose of the Study:

• To develop a deep learning model that leverages the correlation between both eyes for more accurate diabetic retinopathy grading.
• To simulate the clinical practice of comparing both eyes simultaneously for diagnosis.

Main Methods:

• A two-stream binocular network was proposed, processing paired retinal images from both eyes.
• Identical subnetworks were used to analyze each eye's image during training.
• A contrastive grading loss function was designed to learn binocular correlations for five-class DR detection.

Main Results:

• The proposed binocular model demonstrated superior performance compared to monocular methods on the EyePACS dataset.
• The model effectively captured subtle correlations between the left and right eyes.
• Significant improvements in diabetic retinopathy grading accuracy were observed.

Conclusions:

• The binocular deep learning approach offers more accurate diabetic retinopathy predictions than monocular methods.
• This method can extract valuable graphical patterns from both eyes for clinical reference.
• The model enhances the diagnostic capabilities for diabetic retinopathy by considering binocular information.