Pathology foundation models are typically trained on hematoxylin and eosin (H&E) stained slides, yet they are increasingly used as frozen feature extractors for immunohistochemistry (IHC) tasks. The mismatch between training and application domains can be especially problematic in IHC grading, where staining protocols and imaging conditions vary across hospitals.
The new arXiv preprint proposes a correction strategy that operates at two levels: adjusting the input images before they enter the model, and adjusting the feature representations the model produces. By targeting both spaces, the method aims to reduce the domain shift that arises when a model trained on one type of histology is applied to another.
Because the abstract is brief, the paper's exact evaluation details are not available in the source. The key contribution, as stated, is a dual-space correction approach for cross-hospital IHC grading using foundation models.