Table 3.
Paired augmentation ablation analysis on the CEUS domain-shift test.
| Metric | Combined augmentation | Photometric augmentation | Noise augmentation | Blur augmentation |
|---|---|---|---|---|
| Dice | Diff. = 0.008; p < 0.001 | Diff. = − 0.003; p = 0.353 | Diff. = − 0.039; p < 0.001 | Diff. = 0.009; p = 0.003 |
| IoU | Diff. = 0.008; p < 0.001 | Diff. = − 0.003; p = 0.336 | Diff. = − 0.039; p < 0.001 | Diff. = 0.013; p < 0.001 |
| Boundary F1 | Diff. = 0.004; p = 0.005 | Diff. = 0.006; p < 0.001 | Diff. = − 0.020; p < 0.001 | Diff. = 0.013; p < 0.001 |
| HD95 | Diff. = − 2.126; p = 0.158 | Diff. = 2.067; p = 0.081 | Diff. = 8.358; p < 0.001 | Diff. = 3.390; p < 0.001 |
| Pixel ECE | Diff. = − 0.017; p < 0.001 | Diff. = − 0.012; p < 0.001 | Diff. = 0.126; p < 0.001 | Diff. = − 0.038; p < 0.001 |
Paired differences are relative to Residual U-Net. p values were calculated using two-sided Wilcoxon signed-rank tests. For HD95 and Pixel ECE, lower values are favorable. Mean differences, 95% bootstrap confidence intervals, and the proportion of cases improved are reported in the supplement.