Table 7.
TB-consistent ROI segmentation performance achieved using the U-Net models with and without AT using the Montgomery TB CXR test set.
| Model | Dice | IOU | AP@[0.5:0.95] | % Improvement in AP@[0.5:0.95] |
|---|---|---|---|---|
| Standard U-Net | 0.301 | 0.1844 | 0.3171 (0.2526,0.3816) | 1.26 |
| Standard U-Net (AT) | 0.4531 | 0.3021 | 0.3297 (0.2645,0.3949) | |
| V-Net | 0.4121 | 0.2785 | 0.4114 (0.3432,0.4796) | 5.68 |
| V-Net (AT) | 0.4365 | 0.3021 | 0.4682 (0.399,0.5374) | |
| Improved attention U-Net | 0.3903 | 0.2573 | 0.4218 (0.3533,0.4903) | 12.0 |
| Improved attention U-Net (AT) | 0.4806 | 0.3273 | 0.5418 (0.4727,0.6109) | |
| VGG16-U-Net | 0.4263 | 0.2949 | 0.5275 (0.4583,0.5967) | 1.15 |
| VGG16-U-Net (AT) | 0.4771 | 0.3211 | 0.539 (0.4699,0.6081) | |
| VGG19-U-Net | 0.4238 | 0.2881 | 0.5029 (0.4336,0.5722) | 10.14 |
| VGG19-U-Net (AT) | 0.4789 | 0.3148 | 0.6043 (0.5365,0.6721) | |
| VGG16-CXR-U-Net | 0.4667 | 0.3200 | 0.4191 (0.3507,0.4875) | 16.32 |
| VGG16-CXR-U-Net (AT) | 0.5261 | 0.3743 | 0.5823 (0.5139,0.6507) | |
| VGG19-CXR-U-Net | 0.4809 | 0.3306 | 0.4935 (0.4242,0.5628) | 3.37 |
| VGG19-CXR-U-Net (AT) | 0.4694 | 0.3226 | 0.5272 (0.458,0.5964) |
Data in parenthesis are 95% CI for the AP@[0.5:0.95] values measured as the binomial (Clopper–Pearson’s) “exact” method corresponding to separate 2-sided CI with individual coverage probabilities of √0.95. The best performances are denoted by bold numerical values in the corresponding columns.