No two hospitals acquire imaging data in exactly the same way, due to local preferences and machine specificities. Under these heavy domain shift conditions, how do we use machine learning to derive reproducible imaging biomarkers that will work across sites? In this talk, I will focus on machine learning based solutions to this problem such a paired or unpaired image-to-image translation or invariant representations, and discuss other approaches including quantitative imaging and standardization. The talk will be illustrated with real-world examples with various pathologies.