Applicability of segmentation-based stain accuracy metrics to unpaired datasets

Ascertain whether segmentation-based stain accuracy metrics that rely on pixel-level paired annotations between hematoxylin and eosin and immunohistochemistry images are applicable to unpaired datasets used for evaluating virtual immunohistochemistry.

Background

The proposed accuracy framework computes overlap between brown DAB masks from real and virtual IHC and thus requires pixel-aligned H&E–IHC pairs, which are resource-intensive to obtain.

Because many virtual staining models are trained and evaluated on unpaired data, the authors highlight uncertainty about using these segmentation-based metrics in that setting and note that direct stain-accuracy evaluation is infeasible without ground truth.