Abnormality Classification on NIH CXR (External Baseline Reference)
82.2Mass AUROCFT ResNet-50 (2019)
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| FT ResNet-50 (2019)Backbone=ResNet50, Evaluation protocol=Fine-tuning2025.12 | 82.2 | 72.6 | 74.4 | 89.1 | 80 | 79.7 | 76.9 | — | |
| ResNet50 (M-50)Backbone=ResNet50, Masking strategy=M-50 (loose masking)2025.12 | 81.2 | 72.4 | 70.9 | 91.2 | 80.2 | 79.2 | 79.1 | 50 | |
| DenseNet121 (M-50)Backbone=DenseNet121, Masking strategy=M-50 (loose masking)2025.12 | 81.1 | 72.1 | 69.8 | 90 | 78.2 | 78.2 | 77.3 | 58 | |
| ResNet50Backbone=ResNet50, Masking strategy=None2025.12 | 81.1 | 72.4 | 71.1 | 91.6 | 80.9 | 79.4 | 77.2 | 54 | |
| ResNet50 (M-15)Backbone=ResNet50, Masking strategy=M-15 (tight masking)2025.12 | 80.9 | 72.4 | 69.4 | 90.6 | 79.8 | 78.6 | 76.9 | 52 | |
| DenseNet121Backbone=DenseNet121, Masking strategy=None2025.12 | 80.1 | 71.8 | 70.6 | 90.9 | 78.1 | 78.3 | 77.7 | 65.3 | |
| DenseNet121 (M-15)Backbone=DenseNet121, Masking strategy=M-15 (tight masking)2025.12 | 80.1 | 72 | 68.2 | 90.2 | 78.1 | 77.7 | 77.7 | 63 |