Mitotic Count
Counting mitotic figures per defined tissue area — the grading component that breaks most visibly when the microscope is replaced by a screen.
What it is
A count of mitotic figures within a defined area of tumour, used as a proliferation measure in breast grading and, with different conventions, well beyond it. Three separate decisions sit inside what looks like a single number:
- Where to count — hotspot selection.
- How much to count — the area, and how that area is defined.
- What counts — recognising a mitotic figure versus apoptotic or pyknotic mimics.
Each of the three is a distinct source of variability, and on WSI each degrades for a different reason. That is why “the mitotic count is unreliable” is too coarse a statement to act on.
Why it is harder on WSI
From Rakha 2026 (sources/papers/rakha-2026-wsi-breast-grading.md), which identifies mitotic assessment as the principal challenge in digital breast grading:
Hotspot identification. Scanning a whole section for the most proliferative region is a different visual task by mouse than by stage, and there is no established procedure for doing it reproducibly.
Area calibration. The microscope’s high-power field is an optical quantity; the screen’s “high-power view” is not the same thing and varies with monitor size, resolution, and zoom level. Field-of-view equivalence does not survive the transition. The stated fix is to define the counting area in mm², optimally 2–3 mm², and to use calibrated high-power screen fields. The abstract does not give the derivation behind that range. [unverified]
No z-axis. A single-plane scan removes focusing through the section, which is exactly how a pathologist resolves an ambiguous figure at the microscope. Borderline objects that would be adjudicated by fine focus stay ambiguous on a flat image.
The consequence, stated as recognised in the review:
Mitotic counts are systematically underestimated on WSI relative to light microscopy, and this matters most in cases near grading thresholds.
The direction is established. The magnitude is not given in the abstract, nor is the proportion of cases whose grade changes because of it. [unverified] — and that missing number is the one a department would need to decide whether to change anything.
Why it matters for my work
It is a bias term in Aiforia Breast, not just a caveat. If the algorithm reads WSI and the reference standard was read on glass, a systematic modality difference is embedded in the comparison. Where the readout involves proliferation, disagreement is partly the modality. The project page records neither the readout nor the reference standard’s modality. [unverified]
It is a measurement that depends on metadata the pipeline may be discarding. Counting per mm² requires µm/pixel. The metadata-qupath tooling reads scanner resolution from SVS, but the ecosystem notes state anonymisation deletes scanner metadata — an open question already logged on Whole Slide Imaging. For most purposes losing it is a provenance annoyance; for area-calibrated counting it makes the measurement impossible.
The tooling to fix it is already installed. QuPath Annotation Workflow draws annotations and reports their area in mm² directly, so “count within this 2 mm² region” is a QuPath annotation rather than a new capability. If the group is grading digitally, that is the cheapest possible standardisation.
It generalises past breast. Mitotic counting drives grading and risk stratification in several other systems the department signs out. Whatever is decided about screen calibration and counting area applies to all of them, so it is worth deciding once.
How it connects
Nottingham Grading — the scoring system this is one component of, and the reason the undercount can change a diagnosis rather than just a number.
Interobserver Agreement — hotspot choice, area definition, and figure recognition are three separable disagreement sources; a study that reports one kappa for “mitotic count” cannot tell you which of the three to fix.
Biomarker Cut Points — a continuous count banded into an ordinal score, so near-threshold cases are where a systematic bias converts into a changed grade.
Whole Slide Imaging — µm/pixel from scanner metadata is what makes an mm² area computable at all.
QuPath Annotation Workflow — already measures annotation area in mm², so the counting-area recommendation is implementable today.
Scanner and Stain Variability — same underlying theme, extended: the display is part of the instrument chain, not just the scanner.
Intratumoral Heterogeneity — hotspot selection is a sampling decision, so mitotic count inherits the general problem that the answer depends on which region you chose.
Video-Based WSI Mosaicing — a second, worse version of the µm/pixel problem. A scanner at least records physical pixel size; a camera mosaic has none unless someone computes it, and both in-house implementations default it to a placeholder rather than refusing. A default is more dangerous than a missing value here, because a missing value stops the calculation while a plausible default completes it and returns a number that is out by roughly tenfold in area.
Ki-67 Proliferation Index — the other breast proliferation measure. Both estimate how fast the tumour divides and both are near-threshold-sensitive, but they fail in different places: mitotic count through hotspot choice, area calibration and the missing z-axis; Ki-67 through staining and denominator choice. Worth holding side by side rather than treating either as the proliferation readout.
Open questions
- What is the magnitude of the WSI-versus-glass undercount, and what fraction of cases migrate grade because of it? Not in the abstract, and the number that would decide practice.
[unverified] - Is scanner resolution preserved anywhere downstream of anonymisation? If not, area-calibrated counting cannot be done on the archived cohort at all.
- Would a same-case glass-versus-WSI counting round be feasible in-house? It is the direct measurement, it reuses meddecide, and it answers the first question locally.
Is automated mitosis detection in scope for any current project?Answered, and more comprehensively than expected. QuPath Atölye Extension ships eight mitosis modules as one-click QuPath entries — PHH3-based, FCOS, RetinaNet, KongNet, an EfficientNet atypical- mitosis detector, a Sanofi atypical-mitosis detector, and a tiatoolbox route — plusyardimci-mitoz-karsilastir, a module whose whole purpose is to compare them against each other. So the capability exists, it is already packaged for pathologists, and the comparison harness is built. What is missing is a result: no benchmark output is recorded in that repository. Running that comparison on local material would answer a real question and is most of the way done — and it would pair naturally with the glass-versus-WSI counting round proposed above, since both need the same manually counted reference.- Area calibration is preserved by at least one anonymisation route, which partly answers the question below it. The same extension’s anonymisation wizard redacts
date,timeandscanscope idfrom the TIFFImageDescriptionbut leavesMPPandAppMaguntouched — so pixel size and magnification survive that particular tool, and area-calibrated counting remains possible on slides it has processed. That is one tool, not a guarantee about the archive; see De-identification for why the extract-before-redact ordering still matters.