Weakly supervised localization on chest radiographs is usually reported as a single overlap accuracy, but the authors of a new arXiv preprint argue that those numbers are not comparable across papers. They suggest that the metric, as commonly used, hides important differences in what each study actually measures.
The paper proposes a method called PACE, which relies on position-null controls that keep a readout's properties intact while testing its localization behavior. This is meant to give a more meaningful assessment than a standalone overlap figure.
Since this is a single preprint, there is no independent source to compare against; the claims reflect the authors' own analysis. The abstract is brief, so details of the method and results are not fully available from this source alone.