Ji 2025 — Physical Color Calibration of Digital Pathology Scanners
Physical colour calibration of scanners, using a biomaterial calibrant slide and a spectrophotometric reference, substantially improved AI Gleason grading in three external hospital cohorts — with the largest gains where baseline agreement was worst.
Why this is the strongest paper in the calibration collection
It is the one that answers the question the others leave open: calibration obviously makes slides look more alike, but does it make a model work better somewhere else? This study tests exactly that, and it does so with the design External Validation asks for.
Design. Fully supervised and foundation-model-based systems were trained to detect and Gleason-grade prostate cancer on biopsies from the STHLM3 trial (n = 3651), then evaluated in three external cohorts (n = 1161 total) with and without calibration. Agreement with pathologists is Cohen’s linearly weighted κ.
Results (κ, uncalibrated → calibrated):
| Cohort | n | Fully supervised | Foundation model |
|---|---|---|---|
| Stavanger University Hospital | 860 | 0.439 → 0.619 | 0.547 → 0.670 |
| Karolinska University Hospital | 229 | 0.354 → 0.738 | 0.739 → 0.760 |
| Aarhus University Hospital | 72 | 0.423 → 0.452 | 0.424 → 0.459 |
The finding worth extracting
The foundation model was already more robust, so calibration bought it much less. At Karolinska the supervised system went from 0.354 to 0.738 — more than doubling — while the foundation model started at 0.739 and moved only to 0.760. Read the two columns together and the story is that physical calibration and foundation-model pretraining are substantially substituting for one another: both are ways of not being destroyed by a scanner you did not train on.
That has a direct planning consequence, and it is why this paper is worth citing rather than filing. Calibration hardware is a per-site cost; a robust encoder is a one-time modelling choice. Where a group already intends to build on Foundation Models in Pathology, the marginal value of physical calibration is smaller than this paper’s headline suggests — but where a supervised model is being deployed across sites, the headline is real.
The Aarhus cohort is the honest caveat. n = 72, and both systems barely moved (0.423 → 0.452, 0.424 → 0.459). One small cohort where calibration did almost nothing, reported alongside two where it did a great deal, is a reason to treat the effect as site-dependent rather than universal.
Method note
Calibration is physical: a biomaterial-based calibrant slide with a spectrophotometric reference measurement, applied at the scanner. This is not post-hoc software normalisation of the kind in Stain Normalisation — it changes what the scanner produces rather than correcting the image afterwards, which is the distinction Colour Calibration exists to draw.
Related: Colour Calibration — the physical intervention this paper evaluates. Related: Foundation Models in Pathology — the substitution effect above is the reason this paper belongs on that page too. Related: External Validation — three external cohorts and a reported null in the smallest is the design pattern worth copying.