OrdPrune-KD: An Ordinal-Consistency-Based Model Compression Framework for Diabetic Retinopathy Grading

This study proposes OrdPrune-KD, an ordinal-consistency-driven model compression framework that integrates grade-aware structured pruning with Earth Mover's Distance (EMD)-based knowledge distillation for diabetic retinopathy (DR) grading. Unlike conventional approaches that only consider ordinal relationships at the loss level, the proposed method incorporates ordinal priors into both model compression and knowledge transfer stages. Extensive experiments on APTOS 2019, Messidor-2, and IDRiD demonstrate that the proposed framework achieves a favorable balance between model compactness and predictive performance. In particular, under a 77% parameter reduction, the student model achieves competitive performance relative to the teacher model in terms of QWK while maintaining strong high-risk sensitivity. Additional ablation studies and fairness-controlled comparisons confirm that the performance gains are primarily attributed to the proposed ordinal-aware design rather than output formulation differences. These results indicate that OrdPrune-KD provides an effective and deployable solution for lightweight DR grading systems.