Table 1.
Description of the datasets used for validation of QA along with the demonstrated speedup.
| Tissue scale | Histologic structure | Number of slides | Number of ROIs | Number of histologic structures | QA total time (min) | QA human time (QAt, min) | Manual time (M t, min) | Speed up (θ t) | f‐score |
|---|---|---|---|---|---|---|---|---|---|
| Small | Cell nuclei | 5 | 400 | 337,386 | 473 | 391 | 40,165 | 102× | 0.97 |
| Medium | Tubules | 10 | 100 | 5,692 | 121 | 101 | 923 | 9× | 0.95 |
| Large | Epithelium | 10 | 100 | 14,187 | 167 | 113 | 4,433 | 39× | 0.89 |
As mentioned above, the difference between QA total time and QA human time is that human time removes DL training time, as the human annotator was dismissed to perform other non‐related annotation tasks. On the other hand, QA total time includes model training time under the assumption that the user kept annotating during backend training. Manual time is derived by extrapolating the measured annotations per minute from a subset of the annotations.