Table 2.
Comparison of accuracy, sensitivity, and specificity of various deep networks trained and tested on the same test set
| Training model | Accuracy (%) | Sensitivity (%) | Specificity (%) |
|---|---|---|---|
| ResNet34 | 86.8 (305 of 351) | 83.81 (151 of 180) | 90.0 (154 of 171) |
| ResNet50 | 90.0 (316 of 351) | 90.5 (163 of 180) | 89.4 (153 of 171) |
| ResNet50 - No preprocessing | 85.1 (298 of 350) | 82.1 (147 of 179) | 88.3 (151 of 171) |
| ResNet152 | 87.1 (306 of 351) | 83.3 (150 of 180) | 91.2 (156 of 171) |
| CheXpert | 80.6 (283 of 351) | 81.1 (146 of 180) | 80.6 (137 of 171) |
| VGG16 | 85.2(299 of 351) | 81.6 (147of 180) | 88.8 (152 of 171) |
| Ensemble* | 90.3 (317 of 351) | 90.5 (163 of 180) | 90.0 (154 of 171) |
*Bold: model with best accuracy and sensitivity is the ensemble shown in bold