Diabetic Retinopathy Grading by Deep Graph Correlation Network on Retinal Images Without Manual Annotations

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy is a leading cause of vision loss requiring early diagnosis.
• Current deep learning models for diabetic retinopathy screening demand extensive, costly expert annotations.
• There is a need for automated, annotation-free screening tools to improve efficiency and accessibility.

Purpose of the Study:

• To develop an automated diabetic retinopathy grading system using a novel deep learning approach.
• To eliminate the need for professional annotations in training AI models for diabetic retinopathy.
• To enhance the efficiency and reduce the cost of diabetic retinopathy screening.

Main Methods:

• A deep graph correlation network (DGCN) was proposed, integrating graph convolutional networks with convolutional neural networks.
• The DGCN model leverages inherent correlations from retinal image features without manual labels.
• Three specialized loss functions (graph-center, pseudo-contrastive, transformation-invariant) were designed to optimize the DGCN.

Main Results:

• The DGCN model achieved high performance on the EyePACS-1 and Messidor-2 datasets.
• Key metrics included accuracy (89.9% EyePACS-1, 91.8% Messidor-2), sensitivity (88.2%, 90.2%), and specificity (91.3%, 93.0%).
• The model demonstrated strong effectiveness on ROC curves and t-SNE plots, with 95% confidence intervals.

Conclusions:

• The DGCN model provides an innovative, annotation-free method for automated diabetic retinopathy grading.
• Its grading performance approaches that of retina specialists and surpasses trained graders.
• This approach offers a promising pathway for advancing computer-aided diagnostic systems in ophthalmology.