Medical Image Classification on ISIC 2019 (test)
58.4MEL AccuracyFL
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FLTraining Loss=Focal Loss, gamma=2, epochs=152026.05 | 58.4 | 88.7 | 72 | 35.4 | 46.7 | 27.8 | 26.3 | 38.3 | 72.5 | 49.2 | 20.8 | |
| FL+SSMode=Surgery on the FL checkpoint, spikes=7, iterations=202026.05 | 58.1 | 82.8 | 65.2 | 41.5 | 54.3 | 38.9 | 26.3 | 44.7 | 69.8 | 51.5 | 16.4 | |
| CBTraining Loss=Class-Balanced2026.05 | 55.9 | 78.3 | 72.8 | 30.4 | 50.5 | 57.1 | 70.6 | 46 | 67.9 | 57.7 | 14.8 | |
| CB+SSMode=Surgery on the CB checkpoint, spikes=7, iterations=5, Note=early stop2026.05 | 54 | 75.1 | 69.6 | 47.7 | 59.9 | 44.4 | 68.4 | 44.7 | 67 | 58 | 11.3 | |
| SSMode=Spectral Surgery only (starting from baseline)2026.05 | 51.3 | 87.8 | 74 | 38.5 | 42.6 | 50 | 26.3 | 31.9 | 70.7 | 50.3 | 19.6 | |
| BaselineBackbone=ResNet-50, Training Loss=Cross-Entropy2026.05 | 47.2 | 85.9 | 74.4 | 26.2 | 36.5 | 33.3 | 10.5 | 31.9 | 67.7 | 43.2 | 23.6 | |
| tau-normAdjustment=classifier column-norm rescaling, tau=0.252026.05 | 46.9 | 86 | 74.4 | 24.6 | 35.5 | 33.3 | 10.5 | 31.9 | 67.6 | 42.9 | 23.8 | |
| Logit Adj.Adjustment=logit shift, tau=0.252026.05 | 23 | 1.7 | 40.4 | 32.3 | 12.7 | 83.3 | 57.9 | 34 | 14.9 | 35.7 | 24.1 |