Figure 2.

Network architecture: (a) The deep neural network architecture is composed of three modules, FE learns features, F, that successfully classify the input in the outcome y using $MSI$ while being invariant (statistically independent and conditioned by $\rho$) to the biases variables, $b_n$, using the adversarial components BE and the adversarial loss. (b) The bias variables $b_n$, responsible for multiple batch effects, influence both the output y (i.e., ②, MSI status classification) and the input X, from which feature F is extracted (i.e., ①). The MSI classifier deems to find the relation ③ to enable prediction of the output labels while the adversarial components aim to remove the direct dependency between F and $b_n$. Figure adapted from [21] by renaming the modules and adding multiple adversarial components to the architecture. http://creativecommons.org/licenses/by/4.0/.