Abstract

Image-to-image (I2I) synthesis aims at learning the mapping between two visual domains. Due to the scarcity of paired datasets, learning such transformation may be challenging. With a growing number of publications each year, unpaired I2I synthesis has drawn attention from the research community and has found a great application field in medical imaging. Disentangled representation learning constitutes a significant portion of these methods. By relying on the factorization of an image into independent variation latent codes, this approach offers greater control over the result of the synthesis than GAN-based ones. However, disentangling latent representations is not a trivial task and may be influenced by particular inductive bias. In this work, we propose to apply disentangled representation learning to the unpaired synthesis of high-resolution dynamic MRI. We study the impact of both the entanglement module and the addition of a segmentation auxiliary task on the result of the synthesis and the disentanglement of the representations. Our results demonstrate that the choice of the entanglement module greatly influences the learning of a good representation, and that the addition of a segmentation auxiliary task leads to better synthesis performances. Our code is available at https://github.com/cScavinner/Unpaired_image_synthesis