Stain Normalisation Does Not Consistently Improve Matched-Location Retrieval with Pathology Foundation Models
- 1University College London
Abstract
Pathology foundation models provide reusable representations for computational pathology, but staining- and scanner-induced variation may compromise the robustness and reproducibility of embedding-based analyses across acquisition settings. Stain normalisation is widely used to mitigate such colour variation, yet whether it improves retrieval with modern self-supervised pathology encoders remains unclear. We test whether stain normalisation improves cross-scanner and cross-staining image retrieval using the PLISM whole-slide image subset, comprising 13 co-registered images of serial sections prepared under different staining conditions and digitised using seven scanners (91 whole-slide images). We sample 400 matched locations across the 91 images, yielding 36,400 tile instances. We compare two DINOv2-based encoders, Phikon-v2 (1,024-dimensional) and Midnight-12k (3,072-dimensional), with Reinhard and Macenko normalisation against an unnormalised baseline. Performance is measured using paired top-1 nearest-neighbour retrieval accuracy in two settings: cross-scanner retrieval of the same physical section digitised by different scanners, and cross-staining retrieval across co-registered images of serial sections, which differ in staining condition and may also contain section-level morphological variation. Neither normalisation method consistently improves top-1 retrieval relative to no normalisation, while Macenko reduces accuracy by up to 18.3 percentage points. An exploratory UMAP of unnormalised Midnight-12k slide-level mean embeddings showed partial overlap between scanner groups in a two-dimensional projection. These results challenge the routine use of stain normalisation in foundation-model retrieval pipelines and support validating preprocessing choices for each encoder and retrieval setting.
Related
- Pathology foundation-model embeddings reveal a morphological landscape of breast lesions beyond diagnostic categories
- Generalizing Amyloid β Pathology from Postmortem Human Brains with Vision Foundation Models for Lesion Quantification
- Towards reliable morphological subtyping and cancer detection on Barrett's esophagus pathology through ensembling foundation models.
- Foundation Model-Enabled Pathology-Informed MRI Model for Preoperative Prediction of Vessels Encapsulating Tumor Clusters in Resectable Hepatocellular Carcinoma: A Multicenter Retrospective Study.
- Evaluating the Transferability of Pathology Foundation Models Across Cancer-related H&E Neurodegeneration-related Immunohistochemical Classification Tasks
BibTeX
@misc{meng2026stain,
title = {{Stain Normalisation Does Not Consistently Improve Matched-Location Retrieval with Pathology Foundation Models}},
author = {Y. Meng and B. Chai and T. Chakraborti},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.13.758918},
url = {https://www.biorxiv.org/content/10.64898/2026.10.13.758918v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Stain Normalisation Does Not Consistently Improve Matched-Location Retrieval with Pathology Foundation Models
AU - Y. Meng
AU - B. Chai
AU - T. Chakraborti
PY - 2026
DA - 2026/10/13
PB - bioRxiv
DO - 10.64898/2026.10.13.758918
UR - https://www.biorxiv.org/content/10.64898/2026.10.13.758918v1
AB - Pathology foundation models provide reusable representations for computational pathology, but staining- and scanner-induced variation may compromise the robustness and reproducibility of embedding-based analyses across acquisition settings. Stain normalisation is widely used to mitigate such colour variation, yet whether it improves retrieval with modern self-supervised pathology encoders remains unclear. We test whether stain normalisation improves cross-scanner and cross-staining image retrieval using the PLISM whole-slide image subset, comprising 13 co-registered images of serial sections prepared under different staining conditions and digitised using seven scanners (91 whole-slide images). We sample 400 matched locations across the 91 images, yielding 36,400 tile instances. We compare two DINOv2-based encoders, Phikon-v2 (1,024-dimensional) and Midnight-12k (3,072-dimensional), with Reinhard and Macenko normalisation against an unnormalised baseline. Performance is measured using paired top-1 nearest-neighbour retrieval accuracy in two settings: cross-scanner retrieval of the same physical section digitised by different scanners, and cross-staining retrieval across co-registered images of serial sections, which differ in staining condition and may also contain section-level morphological variation. Neither normalisation method consistently improves top-1 retrieval relative to no normalisation, while Macenko reduces accuracy by up to 18.3 percentage points. An exploratory UMAP of unnormalised Midnight-12k slide-level mean embeddings showed partial overlap between scanner groups in a two-dimensional projection. These results challenge the routine use of stain normalisation in foundation-model retrieval pipelines and support validating preprocessing choices for each encoder and retrieval setting.
ER -