Join
ODELIA breast MRI
Lesion classification on multi-centre breast MRI, over seven cross-validation folds (placeholder)
| # | AUROC ↑ | Step | What changed |
|---|---|---|---|
| 7 | 0.902 | ↑ 0.003 | A later commit on this route |
| 6 | 0.899 | ↑ 0.006 | Average the seven fold models, with test-time flips |
| 5 | 0.893 | ↑ 0.004 | Focal loss with centre-balanced sampling |
| 4 | 0.889 | ↑ 0.005 | Cross-attention across the T1, T2 and subtraction sequences |
| 3 | 0.884 | ↑ 0.001 | A 3D state-space backbone in place of the ResNet |
| 2 | 0.883 | ↑ 0.004 | Cross-attention across the T1, T2 and subtraction sequences |
| 1 | 0.879 | ↑ 0.003 | A 3D state-space backbone in place of the ResNet |
| 0 | 0.876 | The baseline, 3D ResNet-18 |
