Harmonization and the Worst Scanner Syndrome
This work highlights fundamental limitations in harmonization schemes for medical imaging, which is incremental as it formalizes intuitive but previously unstated constraints.
The paper demonstrates that for harmonization or domain-invariance methods, prediction accuracy is inherently limited by the domain with the least information, and if labels are highly informative about the source domain, they cannot be accurately predicted by invariant predictors, with these results applied to medical imaging harmonization.
We show that for a wide class of harmonization/domain-invariance schemes several undesirable properties are unavoidable. If a predictive machine is made invariant to a set of domains, the accuracy of the output predictions (as measured by mutual information) is limited by the domain with the least amount of information to begin with. If a real label value is highly informative about the source domain, it cannot be accurately predicted by an invariant predictor. These results are simple and intuitive, but we believe that it is beneficial to state them for medical imaging harmonization.