From discarding to leveraging: quality-aware collaborative learning for robust diabetic retinopathy grading

Real-world diabetic retinopathy (DR) screening faces a paradox: the most diagnostically critical images are often the lowest in quality, because advanced disease itself produces vitreous hemorrhage, proliferative tissue, and media opacities that degrade fundus imagery. We characterize this quality-severity coupling quantitatively (Spearman ρ = 0.420, odds ratio 4.17 for referable DR in Reject vs. Good strata on DDR, p < 0.001) and show that conventional pipelines work against it: filtering low-quality images discards the most severe cases, while uniform processing leads to misclassification. Both behaviors stem from treating image quality assessment (IQA) as a binary preprocessing decision. We argue that quality should serve as a continuous guidance signal that conditions the diagnostic process, and propose QGDR, a quality-guided dynamic routing framework realizing this paradigm through three coordinated mechanisms: (i) a multi-level IQA module that extracts hierarchical quality features across backbone stages; (ii) a quality-conditioned context gating mechanism that modulates spatial attention according to the predicted quality state; and (iii) an adaptive gated fusion mechanism that routes inputs to scale-specialized experts, with high-quality images preferentially activating fine-scale experts for subtle lesions and degraded images relying on coarse-scale experts for robust global pattern recognition. On EyeQ and DDR, QGDR attains 78.32% accuracy with 0.6863 QWK and 80.85% accuracy with 0.8231 QWK respectively, outperforming representative CNN, transformer, and foundation-model baselines while remaining within the compute envelope of standard single-stream backbones. Counter-intuitively, performance is preserved or improved on the lowest-quality stratum (77.61% on EyeQ-Reject; 82.21% on DDR-Reject, exceeding the Good-quality accuracy on the same dataset), and a Shuffled-IQA counterfactual confirms that QGDR exploits the semantic content of quality information rather than a generic auxiliary signal. Test-only evaluation on the external IDRiD and DeepDRiD cohorts confirms cross-dataset generalization. By treating image quality as guidance rather than as a filter, QGDR preserves screening coverage without sacrificing diagnostic reliability.