Data preparation flow: from the whole dataset (i.e. the knowledge base) a validation set is extracted. The rest of the dataset is split through a k-fold partitioning process (in this image, we simplified the figure assuming k = 4 folds, while in reality we used k = 10). The training samples are then arranged, by permuting the involved augmented folds, while the test samples dof not include the artefact images generated by the augmentation process. These sets are finally stacked in order to evaluate the global training performances.