Plus disease detection on Kenyan cohort eye-level (five patient-grouped outer folds)
100SensitivityCalibrated (Eff-B2)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Calibrated (Eff-B2)Backbone=EfficientNet-B2, Variant=Temperature-calibrated2026.07 | 100 | 17.8 | — | — | |
| MIL (Eff-B2)Backbone=EfficientNet-B2, Variant=Multiple Instance Learning2026.07 | 95 | 10.2 | 30.5 | — | |
| ResNet-50Backbone=ResNet-502026.07 | 93.3 | 20.5 | 48.3 | 0 | |
| Multi-task (Eff-B2)Backbone=EfficientNet-B2, Task type=Multi-task2026.07 | 90 | 47.8 | 68.5 | — | |
| EfficientNet-B3Backbone=EfficientNet-B32026.07 | 90 | 22.4 | 56.7 | 0.204 | |
| EfficientNet-B2Backbone=EfficientNet-B22026.07 | 89.2 | 35.3 | 52.1 | 0.268 | |
| Ordinal (Eff-B2)Backbone=EfficientNet-B2, Variant=Ordinal2026.07 | 87.8 | 44.9 | 53.8 | — | |
| EfficientNet-B0Backbone=EfficientNet-B02026.07 | 85.6 | 38 | 51.1 | 0.212 | |
| DenseNet-121Backbone=DenseNet-1212026.07 | 78.3 | 28 | 44.4 | 0.063 | |
| Seg-then-classify (GBM)Classifier=Gradient Boosting Machine (GBM), Type=Two-stage pipeline2026.07 | 73.3 | 48.7 | 27.2 | — | |
| ConvNeXt-TinyBackbone=ConvNeXt-Tiny2026.07 | 71.7 | 68.4 | 63 | 0.327 |