Robust and Lightweight Retinal Microvascular Segmentation via Zero-Leakage Multi-Expert Distillation

Deploying accurate deep learning models for retinal microvascular segmentation on resource-constrained devices requires a trade-off between diagnostic precision and computational efficiency. Furthermore, evaluations based on a single fixed data split can introduce optimistic bias and mask a compact model’s vulnerability to domain shift. To address these challenges, this study proposes a Zero-Leakage Multi-Expert K-Fold Distillation Framework. Using a Dual-Seed Control Strategy on the DRIVE and STARE datasets, we transfer the feature extraction capabilities of a high-capacity U-Net (138.34 M parameters) to a SegNet student with approximately 78.72% fewer parameters (29.44 M), while ensuring that evaluation images are excluded from training and model selection. All eight distilled students significantly outperform the SegNet model trained without distillation on the 28 evaluation images, according to the Wilcoxon signed-rank test with Holm correction. Among the K-fold configurations, the proposed Ensemble Intra-class Feature Variation Distillation (EnsIFVD) achieves the highest mean Dice score (0.6652) and also improves vessel continuity, as measured by clDice. Whereas a SegNet model trained on DRIVE alone exhibits severe performance degradation on STARE (Δshift = 0.2142), and joint training without distillation still results in performance degradation across both domains, EnsIFVD limits the cross-domain degradation to −0.0094 and −0.0075, respectively, achieving performance comparable to that of SegNet models trained on the respective target domains. Because distillation changes only the training objective, these gains incur no additional inference cost. These results indicate that multi-expert distillation is a promising approach for compact retinal vessel segmentation. However, validation on larger, independent, multicenter datasets is required before clinical deployment.