Table 3.

Statistical dependence between the learned features with regard to the target task (MSI) and each bias quantitatively measured by the squared distance correlation ($dc$). Hidden interactions between the different biases are explored by subgroups by re-calculating the $dc$ for each project. The learned features from a baseline model are compared against a bias-ablated model with regard to their statistical dependency to the target task and each bias. In bold are highlighted the lowest distance correlations achieved for each bias. The bias-ablated model maximizes the association of the learned features with respect to the MSI task and minimizes the statistical mean dependence with the biases.

Model	Subgroup	MSI	Project Bias	Patient Bias	Glass Bias
		$\mathit{dc}$	$\mathit{dc}$	$\mathit{dc}$	$\mathit{dc}$
Baseline	All	0.08	0.17	0.08	0.07
Baseline	EPICOLON Project	0.04	-	0.02	0.07
Baseline	HGUA Project	0.46	-	0.29	0.29
Bias-ablated	All	0.21	0.02	0.02	0.02
Bias-ablated	EPICOLON Project	0.15	-	0.01	0.07
Bias-ablated	HGUA Project	0.62	-	0.35	0.33