5-class grading on APTOS 2019 (Stratified 5-fold CV)
0.967QWKChilukoti et al.
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Chilukoti et al.Protocol=Transfer learning across datasets (as reported), Model / Fusion=EfficientNet-based ensemble; DR-specific pretraining, Explainability / Text=Focus on grading; no VLM text evaluation reported2026.04 | 0.967 | — | — | — | |
| Sun et al.Protocol=Per-class 6:2:2 train/val/test split (as reported), Model / Fusion=MAFNet (multi-scale + attention modules), Explainability / Text=No VLM text; focuses on architectural attention modules2026.04 | 0.936 | — | — | — | |
| Backbone benchmarking + weighted soft voting (best)Protocol=Stratified 5-fold CV (mean,SD), Model / Fusion=Backbone benchmarking + weighted soft voting (best), Explainability / Text=Grad-CAM(++) + VLM rationales; quantitative text evaluation2026.04 | 0.934 | — | — | — | |
| Mohsen et al.Protocol=80/20 train/test; 10% of train for validation (as reported), Model / Fusion=RadFuse (fundus + RadEx representation fusion), Explainability / Text=Model comparison and error analysis; no VLM text evaluation reported2026.04 | 0.932 | — | — | — | |
| Dharrao et al.Protocol=Cross-validation + held-out test (as reported), Model / Fusion=EfficientNetB0 with attention gates, Explainability / Text=Grad-CAM-style visual interpretability; no text rationale evaluation2026.04 | 0.923 | — | — | — |