Labquality EQA Staining Dataset

A public set of WSIs in which the same three tissue blocks were stained by 66 laboratories in 11 countries and scanned on one scanner — inter-laboratory staining variation with everything else held constant.

A public set of WSIs in which the same three tissue blocks were stained by 66 laboratories in 11 countries and scanned on one scanner — inter-laboratory staining variation with everything else held constant.

What it is

Built from a Labquality external quality assessment round (Helsinki, Finland). A tissue microarray section containing three 6 mm punch biopsies — normal skin, kidney and colon, from anonymised FFPE material — was cut at 3 µm, and the unstained sections were dispatched to EQA participants. Each laboratory applied its own routine diagnostic H&E protocol. 66 laboratories in 11 countries returned slides.

All returned slides were digitised centrally on a single Hamamatsu NanoZoomer-XR, 20× objective, 0.46 µm/pixel (the published analysis resampled to 10×).

Content 66 WSIs, one per laboratory; each contains all three tissue punches
Tissue Normal skin, kidney, colon — no tumour
Varies Staining laboratory and protocol only
Held constant Tissue block, section thickness, scanner, magnification, digitisation
Access Free, FAIR-compliant server — doi:10.23729/4cb0b2ed-b074-4e5e-8cd6-ff5242c77fd0
Code Normalisation implementations at doi:10.5281/zenodo.12344369

Why it matters for my work

It is the only dataset here that isolates a single variable. Most public cohorts vary scanner, laboratory, protocol and biology simultaneously, which is why “how much does staining alone matter?” is normally unanswerable. This one answers it by construction, and that design is worth studying independently of anything it was used to conclude.

It measures the consultation condition. Konsultasyon receives cases from outside laboratories with slides stained there. Sixty-six laboratories’ worth of routine H&E is a direct measurement of the spread that referred-in material carries.

It is a free external test set. Any normalisation or augmentation choice the group makes can be checked against it without collecting anything — see Stain Normalisation. It is also a ready-made source of realistic colour variation for augmentation design.

It costs nothing to obtain. No data transfer agreement, no PHI exposure, nothing to de-identify — it is normal tissue from anonymised blocks, already public.

Limitations that decide what it can be used for

  • One section per laboratory. Intra-laboratory drift cannot be separated from a laboratory’s systematic bias, so the 66-way spread mixes both and is an upper bound.
  • Normal tissue only. No tumour, so it says nothing about chromatin or eosinophilia in neoplastic tissue — which is where diagnostic work happens.
  • Adjacent sections of one block. Near-identical morphology across images. Excellent for isolating stain; it also means methods that exploit morphological uniformity look better here than they would on a real cohort. See the caveat on Stain Normalisation.
  • Single scanner. By design. Combined scanner-plus-stain variation — the real-world condition — is out of scope.
  • Three tissue types. Skin, kidney, colon.

How it connects

Stain Normalisation — the method page whose evidence base this dataset is.

Scanner and Stain Variability — this dataset supplies the stain half of that page’s two stacked variance sources, cleanly separated from the scanner half.

Konsultasyon — referred-in slides stained in outside laboratories are the in-house instance of exactly this variation.

External Validation — a free external cohort with a known, single axis of variation is an unusually clean way to probe whether a pipeline is colour-dependent.

Foundation Models in Pathology — used in the source paper to show UNI-2 embeddings shift under staining variation, so it doubles as a robustness probe for any encoder.

Open questions

  • Has anyone here downloaded it? Nothing in the repo indicates so. [unverified]
  • Could the group contribute to a future EQA round of this kind? Memorial participating would put in-house staining on the same axis as the other 66 laboratories, which is a cheap way to find out where the department sits in that distribution.
  • Is a comparable EQA dataset available for IHC rather than H&E? That would be more directly useful given the group’s HER2 and biomarker work. Not known. [unverified]