Colour Calibration
Making a scanner produce known, reference-anchored colour by measuring it against a physical standard — as opposed to accepting whatever it produces and correcting the image afterwards.
What it is
Colour in pathology is not decoration; it is the signal, because stains are the only reason tissue structure is visible at all. Clarke & Treanor 2017 — Colour in Digital Pathology: A Review makes the point that the profession lived with wide slide-to-slide colour variation for a century without much investigating it, because a pathologist at an eyepiece adapts without noticing — and that digitisation breaks that arrangement from both ends. It adds a layer of variation (optics, illumination, sensor, display, each with its own colour behaviour), and it exposes the existing variation to algorithms that read colour and intensity numerically and cannot adapt.
Two genuinely different interventions get called “fixing the colour”, and keeping them apart is the whole point of this page:
Physical calibration happens at the scanner, before the image exists. A calibrant slide of known spectral properties is scanned, compared against a spectrophotometric reference measurement, and the scanner’s output is corrected so that a known colour comes out as that colour. Bautista, Hashimoto and Yagi’s calibration-slide method is the classic form — nine colour patches, a correction matrix derived from scanned-versus-inherent colours, applied per pixel, reducing CIELAB colour difference between the same slide imaged on two different scanners (J Pathol Inform 2014;5:4, doi 10.4103/2153-3539.126153).
Software normalisation happens afterwards, on the image, and is covered separately in Stain Normalisation. It maps images onto a common appearance without ever knowing what the true colour was.
There is a third mechanism that sits between the two and is easy to overlook, because it is neither a correction nor a transformation: the ICC profile, a colour-space mapping embedded in the slide file itself. It records how the scanner’s native colours relate to a universal colour space, so a viewer that reads it can render what the tissue would look like down a microscope. Nothing is corrected and nothing is normalised — the information simply travels with the file and is applied at display time.
Its weakness is that it only works if the software honours it, and a great deal does not. Chauveau 2025 — Converting WSI from DICOM to SVS-like TIFF records that QuPath has no out-of-the-box ICC profile support, which is why that author describes it as software “dedicated more to WSI analysis rather than WSI viewing”, and illustrates the difference on an Aperio GT450 DX image — a scanner for which ICC support is called essential. The group runs QuPath on Aperio scanners, so the colour on screen in QuPath is not the colour the scanner intended. For annotation and analysis that is tolerable. For any colour judgement made by a human from a QuPath view, and for any work on this page, it is a confounder sitting upstream of the thing being measured.
The measurement machinery underneath both is colour deconvolution — separating an RGB image into per-stain optical density channels, originally Ruifrok and Johnston 2001 and revisited by Landini, Martinelli and Piccinini (Bioinformatics, doi 10.1093/bioinformatics/btaa847). Optical density rather than raw intensity is the right domain because stain absorbance is what is physically additive; scoring IHC on optical density rather than eyeballed intensity is the same idea applied to quantification.
Why it matters for my work
Because it is the only one of the two that has been shown to move a model’s agreement with pathologists across sites. Ji 2025 — Physical Color Calibration of Digital Pathology Scanners calibrated scanners physically and then measured AI Gleason grading in three external cohorts — the design External Validation asks for. The fully supervised system’s weighted κ went from 0.354 to 0.738 at Karolinska and 0.439 to 0.619 at Stavanger. That is not a cosmetic improvement.
The second half of that result is the one that changes planning here: the foundation model started far more robust and gained far less (0.739 → 0.760 at Karolinska). Physical calibration and foundation-model pretraining are, to a large extent, two ways of buying the same thing — survival on a scanner you did not train on. A group already building on Foundation Models in Pathology should expect a smaller marginal return from calibration hardware than the headline numbers imply, and should still expect a large one if it deploys supervised models across sites.
Because absolute stain intensity is the wrong quality metric, and this is easy to get wrong when building QC tooling. Dunn 2025 — International Study of H&E Stain Variability had 247 labs stain the same circulated tissue and found little correlation between H&E intensity and expert assessor score — but the haematoxylin-to-eosin ratio did trend with it. Quality is a relationship between the two stains, not an amount of either. Any automated H&E quality check built on intensity thresholds is measuring something the experts demonstrably do not care about.
Because calibration is also a scale problem, not only a colour one. The same logic — put a known reference in the frame and solve for the transform — is what makes macroscopic measurement from photographs possible; see Macroscopic Specimen Photography.
How it connects
WSI Format Conversion — where the ICC profile is most at risk: it is one of the several things inside a slide file that a conversion either carries across or silently drops, and it is the one most often lost.
Stain Normalisation — the software alternative; this page is the physical one, and the distinction matters because they have different costs and only one of them can be applied retrospectively to slides already scanned. Scanner and Stain Variability — the problem both interventions exist to solve, stated from the modelling side. Foundation Models in Pathology — a partial substitute for calibration, per the Ji result; this is the cross-link that should change a purchasing decision. WSI Quality Control — where the Dunn intensity-ratio finding lands operationally. Macroscopic Specimen Photography — the same known-reference-in-frame principle applied to scale rather than colour. Ki-67 Proliferation Index — an index whose reproducibility is limited by staining and scoring rather than biology, so it is downstream of everything on this page.
Open questions
- Would physical calibration change anything for this group specifically? The Ji result suggests the answer depends on whether the models deployed are supervised or foundation-model-based. Given the direction of travel here, the honest expectation is a modest gain — but nobody has measured it on Memorial scanners, and the Aarhus cohort in that same paper (n = 72, κ 0.423 → 0.452) shows sites where it does almost nothing.
- Is there a usable cut-point for the H&E intensity ratio? Dunn reports a trend, not a validated threshold. Turning it into an automated check needs a fitted cut-point and a reference standard, and the study deliberately does not supply one.
[unverified] - Nothing in the group’s own tooling currently calibrates anything physically. The stain work here is all post-hoc; see Stain Quality. Whether that is a deliberate choice or simply the path of least resistance has never been written down.