Intratumoral Heterogeneity
Biomarker expression that varies across different regions of the same tumour, so the answer you get depends on which piece you looked at.
What it is
Heterogeneity turns a biomarker result from a property of the patient into a property of the sample. Two implications follow, and both are measurement problems rather than biology problems.
First, sampling: a biopsy samples a small fraction of the tumour, so a heterogeneous marker will be scored differently on biopsy versus resection, and differently between two blocks of the same resection. This is why HER2 scoring rules differ by specimen type — the biopsy rule accepts focal positivity precisely because a small sample cannot rule out positivity elsewhere.
Second, quantification: once expression varies spatially, a single score is a summary of a distribution, and you have to decide which summary. Percentage of positive tumour area, the maximum score in any field, the number of distinct clones — these can rank cases differently. Whole-slide imaging makes this tractable for the first time, because you can annotate the whole tumour area and compute the distribution rather than eyeballing a few fields.
Heterogeneity can be quantified as an entropy, and stratifying by it changes which method wins. sources/papers/uysal-2026-patient-level-gleason.md does this at patient level: the spread of ISUP grade groups across a patient’s prostate biopsy cores is summarised as an entropy, splitting a 106-patient cohort into 37 homogeneous (zero entropy) and 69 heterogeneous cases. Every model performed worse on the heterogeneous group, which is unremarkable. The useful result is that the best strategy inverted: majority voting across cores won where the cores agreed, and compositing all the tissue into a single image and predicting once won where they did not. So a method benchmarked on homogeneous cases does not transfer to heterogeneous ones — a sharper version of this page’s central point. Heterogeneity does not merely add noise to a measurement; it changes which measurement is correct. See Prediction Aggregation.
A model inherits the sampling problem whole, and it can be measured directly. sources/papers/koyun-2026-slide-selection-uncertainty.md holds a trained model, its weights, its patients and their labels fixed, and varies only which of a patient’s whole slide images is fed to it. On 122 breast cancer patients with a mean of 3.28 slides each, patient-level AUC runs from 0.46 to 0.97 depending on the section chosen — near-chance to near-ceiling, on the same people, for the same question. This is the first quantity on this page attached to a model rather than to a reader, and it says the sampling problem is at least as large for automated scoring as for manual scoring.
Two qualifications keep it honest, and the paper supplies both. First, the design cannot separate biology from technique: block, section depth, tumour content, fixation, freezing artefact, folds, section thickness, staining and scanner all vary along with the tissue, and their contributions are not decomposed. What is measured is the combined consequence of analysing a different valid section, which is the operationally relevant quantity but is not “heterogeneity” in the biological sense. Second, excluding tumour-poor slides narrowed the effect without removing it — the best-to-worst gap fell from 0.346 to 0.308 for a recurrence surrogate and from 0.366 to 0.267 for HER2 — so poor tumour representation contributes but does not explain. Notably, the within-patient spread of predictions grew with the number of slides available rather than shrinking, which is what a sampling account predicts and a noise-averaging account does not.
Why it matters for my work
This is the core of HER2 Intratumoral Heterogeneity. That project’s premise is that HER2 scoring in gastric and gastroesophageal carcinoma is harder than in breast, and heterogeneity is one of the three named reasons (alongside fixation variability and a different scoring rubric). The project maps heterogeneity using WSI-level QuPath annotations and separately measures how much of the observed interobserver disagreement it explains.
The chapter records a directly practical consequence: slide-level tumour percentage is needed for downstream heterogeneity metrics, so tumour-area annotation cannot be skipped even when the HER2 score alone is the question. That is a data-collection decision that is expensive to retrofit.
How it connects
Interobserver Agreement — heterogeneity is a legitimate source of disagreement; separating it from observer error is the analytical challenge in the HER2 project.
OncoPath — carries an ihcheterogeneity analysis, found on reading the module’s source on 2026-07-26. It is the most direct overlap between a tool and a project in the whole estate, and whether HER2 Intratumoral Heterogeneity actually uses it is recorded nowhere. [unverified]
QuPath Annotation Workflow — the GeoJSON annotation pipeline that makes spatial quantification possible at all.
Whole Slide Imaging — heterogeneity mapping is only practical once the entire section is digitised rather than sampled by eye.
Biomarker Cut Points — a heterogeneous marker makes any single threshold less stable, since the value being thresholded depends on the sampling scheme.
HER2 Gastric Cohort — the assembled cohort in which the group measures this, and whose mixed biopsy/resection composition is itself a heterogeneity-sampling problem.
Spatial Proteomics — the instrument that measures this directly instead of inferring it from serial single-marker sections, with every marker co-registered on the same cell.
Prediction Aggregation — where heterogeneity stops being a nuisance parameter and becomes the thing that selects the method, evidenced in Uysal 2026 — Patient-Level Gleason Grading from Individual Sections.
Koyun 2026 — Slide Selection Introduces Sampling-Induced Prediction Uncertainty — puts a number on the sampling half of this page for a model rather than a reader: same patient, same weights, different section, AUC 0.46 to 0.97. Read it for the honesty about what it cannot separate — biological heterogeneity from fixation, depth and staining — which is the same confound this project has to handle in HER2.
External Validation — a model tested only on one section per patient has not been tested against this page’s central fact, whatever cohort that section came from.
Open questions
- Which heterogeneity metric will the HER2 project use as its primary readout? Not recorded. [unverified]
- Does the biopsy-versus-resection divergence subset overlap with the interobserver disagreement subset? If the same cases drive both, that is a strong and publishable finding.
- Is there scope to extend the same method to other heterogeneous markers in the group’s portfolio, for example PD-L1?