Table 2.
Robustness of four models. Dice score, Hausdorff Distance (HD) and 95th percentile of HD (HD95) are measured between the ground truth and predicted segmentation by four models on synthetic data.
| Moderate - | Low - | No | Low + | Moderate + | ||
|---|---|---|---|---|---|---|
| TorchXRayVision | Dice | 48.67 ± 21.86 | 58.06 ± 21.53 | 62.26 ± 20.32 | 60.21 ± 22.25 | 53.27 ± 26.12 |
| HD | 2077.14 ± 115.53 | 2219.49 ± 32.33 | 2287.61 ± 32.06 | 2285.14 ± 41.69 | 2176.00 ± 82.62 | |
| HD95 | 1539.31 ± 100.88 | 1796.58 ± 90.51 | 1913.68 ± 49.05 | 1870.52 ± 84.12 | 1630.44 ± 45.73 | |
| U-Net(Real) | Dice | 79.12 ± 8.90 | 82.94 ± 10.46 | 83.79 ± 13.24 | 86.37 ± 6.61 | 76.71 ± 17.69 |
| HD | 6.91 ± 0.30 | 6.38 ± 0.11 | 6.31 ± 0.12 | 6.14 ± 0.17 | 6.69 ± 0.35 | |
| HD95 | 5.07 ± 0.23 | 4.38 ± 0.10 | 4.15 ± 0.19 | 4.01 ± 0.03 | 4.84 ± 0.39 | |
| U-Net() | Dice |
87.14 ± 4.58 (10.14%) |
88.84 ± 4.52 (7.11%) |
89.61 ± 2.78 (6.95%) |
89.14 ± 3.30 (3.21%) |
86.02 ± 7.57 (12.14%) |
| HD | 5.57 ± 0.27 | 5.43 ± 0.10 | 5.56 ± 0.08 | 5.64 ± 0.10 | 5.76 ± 0.13 | |
| HD95 | 4.08 ± 0.24 | 3.77 ± 0.03 | 3.63 ± 0.10 | 3.78 ± 0.06 | 4.07 ± 0.39 | |
| U-Net() | Dice |
88.50 ± 3.45 (11.86%) |
89.03 ± 3.47 (7.34%) |
89.21 ± 3.49 (6.47%) |
89.32 ± 2.19 (3.42%) |
88.95 ± 3.39 (15.96%) |
| HD | 5.36 ± 0.10 | 5.37 ± 0.09 | 5.36 ± 0.07 | 5.50 ± 0.06 | 5.40 ±0.04 | |
| HD95 | 3.71 ± 0.10 | 3.67 ± 0.04 | 3.63 ± 0.05 | 3.64 ± 0.06 | 3.62 ± 0.07 | |
Values are shown in groups of normal and adversarial data using the internal rotation feature . Values are shown as mean ± standard deviation. Values in bracket show the percentage increase of the real and synthetic data-trained models compared to real data-trained model U-Net(Real). Bold values indicate the best results for each internal rotation category.