Table 1.
Distribution of WSI patches across four different institutions
| Training sets | ||||
|---|---|---|---|---|
| Clients | ||||
| Label | C1 | C2 | C3 | C4 |
| Healthy | 1195 | 1363 | 1727 | 1517 |
| Tumor | 1143 | 1363 | 1210 | 1324 |
| Sum | 2338 | 2726 | 2937 | 2841 |
| Test set | |||||
|---|---|---|---|---|---|
| Clients | |||||
| Label | C1 | C2 | C3 | C4 | Sum |
| Healthy | 322 | 338 | 107 | 344 | 1111 |
| Tumor | 374 | 338 | 624 | 537 | 1873 |
| Sum | 696 | 676 | 731 | 881 | 2984 |
Class labels (healthy vs. tumor-containing) are defined at the WSI level, and are exactly balanced within each Client’s overall dataset. Patches in train and test sets are from non-overlapping patients, which forces some imbalance in class labels. The test sets are merged into a single multi-centric dataset for model evaluation.