École Nationale Supérieure de
Techniques Avancées

Abstract

In many biological applications the natural variability within images often largely overlaps the phenotypic variability between drug or disease conditions, making the last essentially invisible to the human eye. Unpaired image-to-image translation machine learning methods aim at learning a mapping from a source to a target domain and have successfully been used to transfer biological images between non-obvious conditions. This internship work is focused on exploring an emerging class of deep models –dubbed diffusion models– in the additional context of a low data regime, and adapting them to this domain transfer task.