Abstract
Domain shift refers to change of distributional characteristics between
the training (source) and the testing (target) datasets of a learning
task, leading to performance drop. For tasks involving medical images,
domain shift may be caused because of several factors such as change in
underlying imaging modalities, measuring devices and staining
mechanisms. Recent approaches address this issue via generative models
based on the principles of adversarial learning albeit they suffer from
issues such as difficulty in training and lack of diversity. Motivated
by the aforementioned observations, we adapt an alternative class of
deep generative models called the Energy Based Models (EBMs) for the
task of unpaired image-to-image translation of medical images.
Specifically, we propose a novel method called the Boundary Preserving
Twin EBMs (BPT-EBM) which employs a pair of EBMs in the latent space of
an Auto-Encoder trained on the source data. While one of the EBMs
translates the source to the target domain the other does vice-versa
along with a novel boundary preserving loss, ensuring translation
symmetry and coupling between the domains. We theoretically analyze the
proposed method and show that our design leads to better translation
between the domains with reduced langevin mixing steps. We demonstrate
the efficacy of our method through detailed quantitative and qualitative
experiments on image segmentation tasks on three different datasets
vis-a-vis state-of-the-art methods.