Answers are grounded in this paper and cite their sources.
Introduction Automated diabetic retinopathy (DR) grading systems must recognize when their predictions are unreliable. Most deep learning approaches optimize solely for classification accuracy and produce overconfident predictions, limiting their safety in screening and referral workflows. Existing prototype-based methods do not model ordinal disease progression or calibrated confidence. Methods We propose EviProto-DR, a framework combining grade-supervised ordinal prototype learning with evidential deep learning for DR grading and predictive uncertainty estimation. The framework comprises three components: Grade-Supervised Ordinal Prototypes (GSOP), which produce class-anchored embeddings under a monotonicity constraint reflecting DR severity ordering an Evidential Prototype Classifier (EPC), which derives Dirichlet-based grade predictions and epistemic uncertainty in a single forward pass and a Prototype-Uncertainty Consistency Loss (PUCL), which aligns prototype geometry with predictive uncertainty. Results On APTOS 2019, EviProto-DR achieves a Quadratic Weighted Kappa (QWK) of 0.918 and an Expected Calibration Error (ECE) of 0.061, while achieving a QWK of 0.897 on DDR. External evaluation on MESSIDOR-2 provides an initial assessment of cross-dataset performance without retraining. In a simulated referral setting, the model routes high-uncertainty cases to a specialist while retaining 91.3% accuracy on non-deferred cases. Discussion These results indicate the potential of EviProto-DR for uncertainty-aware DR grading, though prospective clinical evaluation remains necessary before deployment.
No takes yet. Share an insight, caveat, or question.
Kirubakaran et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: