Table 2:
Linear evaluation protocol results for contrastive representation learning. Each experiment included three random initial seeds. Mean value and standard deviation (in parentheses) for each metric are reported here. The full table including false negative rates and slide-level metrics can be found in the Appendix E.
| Backbone | Methods | Patch | Patient | ||||
|---|---|---|---|---|---|---|---|
|
| |||||||
| Accuracy | MCA | MAP | Accuracy | MCA | MAP | ||
| ResNet50 | ImageNet | 68.3 (0.0) | 67.9 (0.0) | 72.9 (0.1) | 80.0 (0.0) | 82.9 (0.0) | 88.8 (0.1) |
| ResNet50 | SimCLR | 79.1 (0.4) | 78.9 (0.4) | 84.2 (0.6) | 83.9 (1.0) | 86.1 (0.9) | 92.4 (0.1) |
| ResNet50 | SupCon | 87.5 (0.3) | 86.8 (0.3) | 91.5 (0.5) | 90.0 (0.0) | 91.4 (0.1) | 94.6 (0.5) |
| ViT-S | ImageNet | 71.8 (0.1) | 71.1 (0.1) | 77.1 (0.1) | 88.3 (0.0) | 89.8 (0.0) | 93.9 (0.0) |
| ViT-S | SimCLR | 76.8 (0.5) | 76.3 (0.5) | 82.5 (0.3) | 80.0 (1.7) | 83.0 (1.3) | 92.3 (0.0) |
| ViT-S | SupCon | 81.4 (0.2) | 80.2 (0.3) | 85.6 (0.5) | 87.8 (1.0) | 89.4 (0.7) | 94.0 (0.4) |