Table 5.
Performance of Resnet-101 trained with imbalanced data, oversampling, and undersampling (without and with Data Augmentation (DA)) in the multi-class classification of Luminal A vs. Luminal B1 vs. Luminal B2 vs. HER2 vs. TN, in terms of AUC. For each molecular subtype, the best AUC value obtained is shown in bold.
| AUC (%): Mean (stdDev) | ||||||
|---|---|---|---|---|---|---|
| Imbalanced Data | Oversampling | Undersampling | ||||
| without DA | with DA | without DA | with DA | without DA | with DA | |
| Luminal A | 61.01 (4.53) | 59.92 (2.65) | 57.53 (4.66) | 54.09 (1.80) | 53.11 (3.96) | 58.05 (6.23) |
| Luminal B1 | 55.52 (3.55) | 59.81 (2.09) | 51.00 (9.50) | 53.89 (3.14) | 58.16 (3.91) | 56.13 (2.87) |
| Luminal B2 | 56.43 (0.86) | 59.20 (4.22) | 52.09 (7.08) | 59.91 (0.28) | 57.65 (5.88) | 53.36 (2.57) |
| HER2 | 66.27 (2.40) | 64.40 (2.40) | 73.14 (4.38) | 70.96 (4.00) | 70.01 (4.34) | 66.78 (4.19) |
| TN | 58.06 (1.98) | 60.89 (5.51) | 53.39 (1.72) | 64.27 (2.26) | 57.70 (9.44) | 61.06 (2.60) |
| Average | 59.46 (2.66) | 60.84 (3.37) | 57.43 (5.47) | 60.62 (2.30) | 59.33(5.51) | 59.08 (3.69) |