Toward Reliable Diabetic Retinopathy Screening

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) grading is crucial for timely treatment but faces challenges due to data variability and distribution shifts.
• Existing methods struggle with reliable five-grade severity assessment across diverse datasets.

Purpose of the Study:

• To introduce PRISM-DR, a multi-objective framework for robust five-grade diabetic retinopathy grading.
• To enhance automated DR screening by improving accuracy, robustness, and capacity.

Main Methods:

• A feedforward pipeline incorporating a ConvNeXtV2-Base backbone, Recurrent BiFPN neck, and a multi-scale reasoning transformer.
• Dual classification heads (categorical and ordinal) with gradient-isolated pathways and a prototype memory module.
• Multi-crop training with a phased loss curriculum on imbalanced DR datasets.

Main Results:

• Achieved high Quadratic Weighted Kappa (QWK) scores across multiple datasets (e.g., 0.835 on IDRiD, 0.865 on APTOS2019).
• Outperformed existing models like RETFound and MedGemma-4B in AUC-PR across evaluated datasets.
• Demonstrated emergent, anatomically plausible spatial priors from grade-level supervision alone.

Conclusions:

• PRISM-DR offers a superior accuracy-robustness-capacity trade-off for automated DR screening.
• The framework effectively addresses acquisition variability and cross-dataset distribution shifts.
• PRISM-DR enables scalable and reliable automated diabetic retinopathy grading.