Harmonization and the Worst Scanner Syndrome

TL;DR

This paper demonstrates fundamental limitations of domain-invariance methods in medical imaging harmonization, showing they can restrict predictive accuracy and fail when labels are highly informative about the source domain.

Contribution

It provides theoretical insights into the unavoidable trade-offs and limitations of domain-invariance schemes in medical imaging harmonization.

Findings

Invariant predictors are limited by the least informative domain.

Highly informative labels about the source domain cannot be accurately predicted.

The results highlight fundamental constraints in harmonization methods.

Abstract

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.