Decomposing scanner batch effects and their impact on pathology foundation models

Abstract

Pathology foundation models (PFMs) are expected to represent the same tissue in the same way on any scanner, but their embeddings still carry scanner signals. Scanner-related image differences and the scanner sensitivity of PFMs have each been described. How tissue-dependent changes in color and spatial frequency relate to shifts in PFM representations, and what this means for correction, is less understood. We imaged 103 slides from 37 normal tissue types on six scanners (PanNormal; 95,780 matched locations) and used the real paired images as targets. We used the public PLISM dataset for external validation. First, we decomposed scanner differences into image properties. Color and spatial frequency changed together, by amounts that depended on tissue type. Second, corrections that matched these measured properties brought images closer to the target more reliably than representations. Pix2Pix reduced the image residual by 69% but increased UNI distance to the target by 27%. The same frequency matching helped one scanner and hurt another. Probes still distinguished corrected from real target embeddings after every correction. Third, changing one frequency band at a time showed that PFMs differ in how strongly each band moves them and whether it moves them toward the target. Across eight PFMs pretrained on whole-slide image tiles, the models that moved least under these small changes were the most robust to real scanner differences (Spearman {rho}=-0.90). Aligning images and aligning representations are therefore different goals. The gap between them depends on scanner, tissue, and model.