Grade-Skewed Domain Adaptation via Asymmetric Bi-Classifier Discrepancy Minimization for Diabetic Retinopathy Grading

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a major cause of preventable vision loss globally.
• Deep learning shows potential for DR grading but faces challenges in generalization across different datasets.
• Domain shift and imbalanced grade distribution in DR datasets hinder model performance.

Purpose of the Study:

• To address the challenges of domain shift and grade imbalance in diabetic retinopathy grading using deep learning.
• To propose a novel image-level supervised DR grading method for improved generalization.

Main Methods:

• Developed Asymmetric Bi-Classifier Discrepancy Minimization (ABiD) for grade-skewed domain adaptation.
• Optimized feature extractor by minimizing classifier prediction discrepancies to enhance adjacent grade feature exploration.
• Maximized classifier differences using distribution compensation to mitigate pseudo-label bias.

Main Results:

• The proposed ABiD method significantly outperforms state-of-the-art methods on public and private DR datasets.
• Demonstrated improved generalization performance in cross-center, cross-vendor, and cross-user scenarios.
• Effectively handled imbalanced grade distributions and small lesion detection.

Conclusions:

• ABiD offers a robust solution for domain-adaptive DR grading, overcoming key limitations of existing methods.
• The approach shows promise for more reliable and widespread clinical application of AI in ophthalmology.
• Further research can explore advanced techniques for even greater accuracy and adaptability in DR detection.