Supervised Contrastive Learning with Angular Margin for the Detection and Grading of Diabetic Retinopathy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) diagnosis from fundus images is crucial for preventing vision loss.
• Deep learning, particularly supervised contrastive learning (SupCon), shows promise for intelligent DR diagnosis.
• Existing SupCon methods do not differentiate between augmented and same-label positives, limiting performance.

Purpose of the Study:

• To introduce and evaluate a novel Angular Margin concept integrated into SupCon for improved DR detection and grading.
• To address the limitations of current SupCon methods in distinguishing positive sample types.

Main Methods:

• Proposed the Angular Margin concept and incorporated it into the supervised contrastive learning framework.
• Tested the enhanced SupCon method on two DR datasets for detection and grading tasks.
• Utilized standard metrics (Accuracy, Precision, Recall, F1, AUC), alignment, uniformity, and UMAP for evaluation and visualization.

Main Results:

• DR detection achieved state-of-the-art performance with Accuracy = 98.91%, Precision = 98.93%, Recall = 98.90%, F1 = 98.91%, and AUC = 99.80%.
• DR grading also reached state-of-the-art results, with Accuracy = 85.61% and AUC = 93.97%.
• Angular Margin demonstrated effective representation learning and improved diagnostic capabilities.

Conclusions:

• The proposed Angular Margin strategy significantly enhances deep learning-based intelligent medical diagnosis for diabetic retinopathy.
• This method offers a superior approach for both DR detection and grading compared to existing techniques.
• Angular Margin represents a valuable advancement in AI-driven ophthalmological diagnostics.