Table 2.
percentage differences between each saliency map and the masked gradient saliency map. Masked gradient is used as a baseline for comparing saliency methods. The largest percentage improvement according to each measure is shown in bold
| Model D | NMI | Dice | NSS | CC | SIM |
|---|---|---|---|---|---|
| Gradient | -0.03% | -94.68% | -157.79% | -163.15% | -95.13% |
| Input X Gradient | 0.02% | -0.04% | 49.29% | 56.37% | 39.59% |
| Masked Gradient | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% |
| Guided Backprop | -0.03% | -94.28% | -157.19% | -163.05% | -94.99% |
| Guided GradCAM | -0.03% | -85.78% | -115.60% | -117.24% | -95.71% |
| Integrated Gradient | 0.04% | 0.40% | 87.73% | 87.08% | 78.00% |
| Gradient SHAP | -0.01% | -5.61% | -2.13% | -10.16% | 1.14% |