Table 5.
Off-domain evaluation of performance of the vision transformer—standardized training data sizes.
| Train on: | Test on: | |||||
|---|---|---|---|---|---|---|
| Training strategy | Dataset [Size] | VinDr-CXR | ChestX-ray14 | CheXpert | MIMIC-CXR | PadChest |
| Local training | VinDr-CXR [n = 15000] (*) | OND | 66.4 ± 5.9 (0.001) | 69.3 ± 10.1 (0.001) | 73.4 ± 6.3 (0.001) | 79.6 ± 6.6 (0.001) |
| ChestX-ray14 [n = 60000] | 85.9 ± 6.7 (0.001) | OND | 75.0 ± 7.6 (0.001) | 76.5 ± 6.1 (0.001) | 82.9 ± 6.6 (0.001) | |
| CheXpert [n = 60000] | 85.3 ± 8.3 (0.001) | 75.3 ± 7.6 (0.039) | OND | 78.0 ± 6.6 (0.001) | 82.1 ± 7.9 (0.001) | |
| MIMIC-CXR [n = 60000] | 90.0 ± 5.4 (0.008) | 75.0 ± 4.6 (0.468) | 77.6 ± 7.1 (0.001) | OND | 85.1 ± 5.4 (0.747) | |
| PadChest [n = 60000] | 88.8 ± 5.0 (0.001) | 72.6 ± 5.6 (0.001) | 74.3 ± 7.9 (0.001) | 76.7 ± 6.2 (0.001) | OND | |
| Collaborative training | All datasets [n = 4 × 15000] | 91.1 ± 4.2 | 75.0 ± 6.0 | 76.5 ± 7.8 | 78.7 ± 5.8 | 85.2 ± 5.7 |
Table organization as in Table 4 above.