TABLE 4.
Evaluation of WMH segmentation on different adversarial attacks
| Adversarial attacks | HyperMapper Bayesian | BIANCA | DeepMedic | |
|---|---|---|---|---|
| Dice similarity coefficient | Noise (sigma = 0.2) | 0.806 (±0.086) | 0.252 (±0.130) | 0.639 (±0.217) a |
| Downsampled (2 × 2 × 2) | 0.809 (±0.101) | 0.348 (±0.176) | 0.306 (±0.208) b | |
| Contrast (gamma = 0.5) | 0.822 (±0.102) | 0.389 (±0.184) | 0.677 (±0.198) | |
| Hausdorff distance (HD95) (mm) ↓ | Noise (sigma = 0.2) | 3.592 (±4.686) | 31.521 (±14.999) | 14.092 (±13.626) a |
| Downsampled (2 × 2 × 2) | 5.836 (±6.475) | 30.469 (±18.451) | 50.049 (±21.236) b | |
| Contrast (gamma = 0.5) | 5.047 (±6.172) | 29.488 (±17.077) | 10.742 (±10.5854) | |
| Absolute volume difference (%) ↓ | Noise (sigma = 0.2) | 15.037 (±13.169 | 159.602 (±278.002) | 42.627 (±24.601) a |
| Downsampled (2 × 2 × 2) | 19.745 (±18.260) | 156.712 (±325.622) | 77.265 (±18.012) b | |
| Contrast (gamma = 0.5) | 24.454 (±23.178) | 176.860 (±365.183) | 40.007 (±24.096) | |
| Sensitivity (recall) | Noise (sigma = 0.2) | 0.720 (±0.136) | 0.669 (±0.168) | 0.259 (±0.124) a |
| Downsampled (2 × 2 × 2) | 0.677 (±0.148) | 0.601 (±0.217) | 0.072 (±0.040) b | |
| Contrast (gamma = 0.5) | 0.601 (±0.200) | 0.759 (±0.186) | 0.348 (±0.191) | |
| F‐1 score | Noise (sigma = 0.2) | 0.505 (±0.160) | 0.139 (±0.142) | 0.376 (±0.134) a |
| Downsampled (2 × 2 × 2) | 0.718 (±0.010) | 0.316 (±0.169) | 0.131 (±0.066) b | |
| Contrast (gamma = 0.5) | 0.662 (±0.135) | 0.213 (±0.175) | 0.469(±0.166) |
Note: ↓ indicates that smaller values represent better performance.
Values in bold indicate best performance.
Abbreviation: WMH, white matter hyperintensity.
a
DeepMedic failed on one subject with increased noise.
b
DeepMedic failed on 62 subjects with lower resolution.