Table 3.
Generalization to an unseen dataset.
| Baseline | Wu et al. (18) | Ensembles | MC-drop | Ours (MTL) | |
|---|---|---|---|---|---|
| Overall | 72.45 ± 0.26 (0.02) | 74.04 (<0.01) | 73.34 ± 0.26 (0.02) | 67.28 ± 0.26 (<0.01) | 78.13 ± 0.63 |
| Density | |||||
| A | 74.30 ± 0.26 (0.28) | 64.27 (<0.01) | 72.15 ± 0.78 (0.03) | 71.01 ± 0.22 (0.02) | 78.91 ± 0.23 |
| B | 79.91 ± 0.35 (0.09) | 80.23 (0.11) | 77.43 ± 0.89 (0.07) | 73.34 ± 0.13 (<0.01) | 85.57 ± 0.32 |
| C | 55.05 ± 1.21 (0.11) | 78.83 (<0.01) | 56.64 ± 0.62 (0.13) | 52.00 ± 0.11 (0.01) | 65.61 ± 0.22 |
| D | 67.70 ± 1.09 (0.82) | 62.50 (0.47) | 64.39 ± 1.23 (0.62) | 62.38 ± 0.17 (0.46) | 70.83 ± 0.32 |
| View angle | |||||
| CC | 74.61 ± 0.73 (0.02) | 72.43 (<0.01) | 68.12 ± 1.46 (0.01) | 63.79 ± 0.26 (<0.01) | 83.06 ± 0.43 |
| MLO | 72.69 ± 0.47 (0.76) | 79.83 (<0.01) | 70.98 ± 1.83 (0.07) | 70.20 ± 0.26 (0.05) | 73.62 ± 0.37 |
Binary cancer classification performance [“AUC score (p-value to ours)”] of the compared methods on the entire DINB dataset and for the subsets filtered per density class and per view angle.