Adaptive Gating and Focal Debiasing for robust few-shot retinal disease classification

Hybrid attention few-shot models are promising for automated retinal disease classification in data-scarce settings, yet two issues persist: static fusion of global and local features, and weak performance on minority classes. We propose Adaptive Gating and Focal Debiasing (AGFD), which adds a Dynamic Attention Gating (DAG) module that learns input-specific weights for global and local attention branches, and replaces cross-entropy with focal loss to shift learning toward hard, underrepresented cases. On ODIR-5K under episodic evaluation, AGFD improves over a strong hybrid-attention baseline across all shot settings, reaching 78.7% accuracy and 76.9% macro-F1 in the 5-shot setting, and 83.2% accuracy at 10-shot. Minority classes benefit most, with +11–15 percentage-point F1 gains for Glaucoma, Cataract, Hypertension, and Other. Gate analysis shows higher global weighting for widespread conditions and higher local weighting for lesion-driven diseases. Coupling adaptive fusion with a debiased objective improves overall accuracy and reliability on underrepresented classes, a step toward more clinically useful screening.