A novel deep autoencoder architecture is proposed for the analysis of histopathology images. Its purpose is to produce a disentangled latent representation in which the structure and colour information are confined to different subspaces so that stain-independent models may be learned. For this, we introduce two constraints on the representation which are implemented as a classifier and an adversarial discriminator. We show how they can be used for learning a latent representation across haematoxylin-eosin and a number of immune stains. Finally, we demonstrate the utility of the proposed representation in the context of matching image patches for registration applications and for learning a bag of visual words for whole slide image summarization.
Enhancing Dimensionality Reduction Techniques for Deep Neural Network Visualization