Table 2.

CNN performance overview.

Performance at WSI-level (private data)
Dataset	Micro-accuracy (SKET labels)	Micro-accuracy (GT labels)	Weighted F1-score (SKET labels)	Weighted F1-score (GT labels)
Catania	0.911 ± 0.004	0.918 ± 0.006	0.797 ± 0.011	0.807 ± 0.020
Radboudumc	0.906 ± 0.005	0.909 ± 0.008	0.744 ± 0.020	0.758 ± 0.025
Private data	0.908 ± 0.005	0.912 ± 0.006	0.769 ± 0.015	0.779 ± 0.019
Performance on publicly available images
Dataset	Accuracy (SKET labels)	Accuracy (GT labels)	Weighted F1-score (SKET labels)	Weighted F1-score (GT labels)
GlaS36	0.745 ± 0.059	0.745 ± 0.065	0.717 ± 0.050	0.750 ± 0.066
CRC37	0.876 ± 0.014	0.856 ± 0.024	0.878 ± 0.019	0.855 ± 0.024
UNITOPATHO31,32 (single sections)	0.549 ± 0.025	0.543 ± 0.026	0.590 ± 0.015	0.591 ± 0.020
UNITOPATHO31,32 (WSIs)	0.750 ± 0.022	0.770 ± 0.025	0.723 ± 0.024	0.764 ± 0.023
TCGA-COAD33	0.862 ± 0.051	0.868 ± 0.093	0.925 ± 0.029	0.927 ± 0.056
Xu38	0.717 ± 0.053	0.728 ± 0.038	0.677 ± 0.084	0.725 ± 0.041
AIDA34	0.743 ± 0.046	0.760 ± 0.030	0.744 ± 0.047	0.752 ± 0.026
IMP-CRC35	0.706 ± 0.035	0.678 ± 0.048	0.707 ± 0.033	0.682 ± 0.048

Results for the performance of the CNN on WSI-level classification task for the Catania and Radboudumc datasets (upper part) and for the classification of images from publicly available datasets (lower part). The performance at WSI-level is evaluated with micro-accuracy and weighted F1-score. For each classification type, the average and the standard deviation (of the models involved in the k-fold cross-validation) are reported for each metric, including cumulative results for each dataset. The performance is reported for the CNNs trained using the automatically generated weak labels (SKET labels) and the manually created ground truth weak labels (GT labels).