Deep Variational Clustering Framework for Self-labeling of Large-scale Medical Images

The proposed framework is composed of two networks (see Fig 1). The encoder network q with parameters of φ computes qφ(z|x) : xi → zi. The encoder maps an input image xi ∈ X to its latent embedding zi ∈ Z in a lower dimensionality space compared to the input space X. The decoder network p parametrizes by θ, pθ : zi → x(cid:48)i, and reconstructs xi from its latent embedding zi.