Table 2.

The embedding is important for the downstream classifiers.

	Plain ResNet embeddings	BYOL ResNet embeddings
MIL	0.85	0.67
Barcode Analysis	0.65	0.68

This table denotes the performance of the MIL and barcode method for both the plain ResNet embedding and the BYOL-trained ResNet embedding. For each combination the macro F1 score is provided for the classification across all tumor types. While the plain ResNet embedding works best for the MIL approach, the barcode approach works best with the BYOL-trained embedding.