TCGA
The public pan-cancer atlas that most computational pathology benchmarks are built on — free, large, matched across modalities, and carrying confounders that its size does not fix.
What it is
The Cancer Genome Atlas is a US National Cancer Institute programme that profiled roughly 11,300 cases across 33 cancer types, and released the result publicly through the Genomic Data Commons. Its defining feature is not size but matching: the same patient has gene expression, microRNA expression, DNA methylation, copy number, clinical follow-up and a digitised diagnostic H&E slide. Very few resources pair molecular and image data on the same person at this scale, and none of the alternatives are free.
Coverage is uneven across modalities, and any study using more than one has to decide what to do about that. The counts below are the ones MultiSurv worked with after its own filtering, out of 11,081 patients with usable follow-up:
| Modality | Patients | Notes |
|---|---|---|
| Clinical | 11,081 | Follow-up time, event, stage, demographics |
| DNA methylation | 10,257 | Array beta values; ~1.4 TB raw |
| Copy number | 10,325 | Gene-level scores, coded loss / neutral / gain |
| miRNA expression | 9,616 | ~557 MB raw |
| mRNA expression | 9,605 | FPKM-UQ or raw counts; ~5.8 GB |
| Whole-slide images | 8,376 | Diagnostic slides; ~13 TB raw |
Access is through the GDC Data Portal, using a manifest file plus the GDC Data Transfer Tool. Clinical data is easier to pull with the TCGAbiolinks R package. The image and expression data used here are open tier; raw sequence data is controlled access and needs dbGaP authorisation. [unverified] as to current authorisation requirements — the MultiSurv manifests date from 2019.
TCGA holds two kinds of slide, and everything above is about one of them. The 8,376 slides in the table are diagnostic FFPE sections. Beside them sit the flash-frozen tissue slides — cut from the specimens that fed the molecular assays — which are more numerous per case, generally poorer histology, and a different domain for any model: Lerousseau 2020 — MIL Tumor Segmentation (percentile pseudo-labels) trained on 6,481 frozen slides and then scored at or near chance on FFPE tissue in its own external test. A method or number derived from TCGA slides is not interpretable until it says which kind.
Two things from that paper are worth knowing as TCGA facts rather than paper facts. The barcode carries a tumour/normal label for free: the sample-type field (position four of the name; codes below 10 are tumour, 10–19 solid-tissue normal) labels every slide with no pathologist involved, which is how a 6,481-slide weakly supervised corpus was assembled at zero annotation cost. And a derived resource exists for the frozen slides: that paper’s release ships ensemble-generated tumour maps for the entire snap-frozen breast, kidney and bronchus/lung TCGA collections — raw tile probabilities and 0.3-thresholded binaries at 112 µm granularity — usable as a free tumour-region filter when tiling frozen slides. Caveats travel with it: ensemble output, not annotation; frozen sections only; the release states its own count as both 6461 and 6481; AGPL-3.0 as the repository declares it.
Why it matters for my work
It is the default answer to “where do I get data to try this on”. For any method the group wants to evaluate before committing scanner time or cohort assembly — a fusion architecture, a survival model, a tile encoder — TCGA lets that happen this week rather than after an ethics submission. It is also the cohort most published comparators were built on, so matching its setup is often the only way to compare against them.
The caveats matter more than the convenience, and four of them are specific enough to plan around.
Slide magnification is not uniform. TCGA diagnostic slides include both 40× and 20× scans, and some have no readable magnification field at all. MultiSurv dealt with this by discarding every 20× slide and every slide with an unreadable aperio.AppMag — a real methodological control, and part of why its image cohort is 8,376 rather than the ~11,700 slides available. Any patch-based work has to make the same decision explicitly, because a patch of fixed pixel size covers different amounts of tissue at the two magnifications. See Patch Extraction.
Slides come from many contributing institutions, and institution is a confounder. Cases were submitted by dozens of tissue source sites, each with its own H&E protocol and scanning setup, and the site a case came from correlates with both the cancer type and the patient population. So institution-specific colour and texture are learnable signal that has nothing to do with biology — the mechanism described on Scanner and Stain Variability, with an unusually large surface here because nothing about TCGA’s assembly was standardised for imaging. MultiSurv applied no stain normalisation — only colour augmentation (ColorJitter) at training time, absent from its validation and test pipelines. Augmentation asks a model to tolerate colour variation; normalisation removes it. The two are often conflated and only the second addresses a site confounder at inference. [unverified] as to the published magnitude of this effect; the mechanism is what has been verified here.
It is in the pretraining corpus of half the published tile encoders, which makes a TCGA benchmark an in-domain test rather than a held-out one. This is now checkable rather than assumed. Bareja et al. 2026 (sources/papers/bareja-2026-foundation-model-benchmark.md) tabulate the declared pretraining corpus of 16 pathology encoders:
| Declares TCGA | Declares a proprietary corpus, no TCGA named |
|---|---|
| cTransPath (+PAIP), Kaiko, HIPT, Lunit (+TULIP), Phikon, GPFM (+PAIP, +49 public sets), Phikon-v2 (+GTEx, +CPTAC, +TCIA), H-optimus-mini | Virchow, UNI (+GTEx), Hibou, H-optimus-0, UNI2, Virchow2, Prov-GigaPath |
EXAONEPath declares only “Public”. Two things to take from this. All five of that benchmark’s top performers are in the right-hand column — so for the encoders most likely to be reached for here, TCGA is not declared training data, though undisclosed proprietary corpora make that a weaker assurance than it looks; the paper’s own Discussion says exposure “cannot be ruled out for all models.” And Phikon-v2 declares CPTAC, which is worth knowing because CPTAC is the usual substitute when someone wants a TCGA-like cohort that is external — it is not external for that model.
The design lesson is the one that benchmark builds on: report TCGA results separately from everything else rather than pooling them, because the two measure different things. See External Validation.
The slides are diagnostic, tumour-rich sections, not a representative sample of practice. They were selected to confirm a diagnosis and yield material for molecular profiling. A model trained on them has not seen the normal, the equivocal, or the poorly fixed — which is most of what arrives in a real service.
Clinical fields are inconsistent across entities and partly recorded during follow-up. Stage is the clearest case: MultiSurv’s preprocessing needed a dedicated cleanup step and still ended with 20 distinct stage strings across the 33 diseases, which cannot be treated as one ordinal scale. More subtly, fields such as “received pharmaceutical treatment” and “received radiation treatment” are populated over the follow-up period, so using them as baseline predictors leaks post-baseline information — MultiSurv does exactly that, and it is one of the stronger criticisms on the paper note.
Follow-up is heavily censored and uneven by entity, which is a survival-analysis problem rather than a data problem, but it decides whether a pan-cancer outcome model means anything. See Concordance Index for why pooling 33 diseases into one discrimination statistic flatters the result.
How it connects
Multimodal Fusion — TCGA is essentially the only public resource where a patient-level multimodal model can be built at all, which is why nearly every paper in that area uses it.
Concordance Index — the pan-cancer pooling that TCGA invites is the exact setting where a concordance index measures cohort heterogeneity instead of model skill.
Patch Extraction — the mixed 40×/20× problem has to be resolved before any tiling, and this is where that decision is recorded.
Scanner and Stain Variability — multi-institution submission makes TCGA a worst case for site signal, and the reason a stain decision cannot be deferred.
Labquality EQA Staining Dataset — the deliberate opposite design: one block, 66 laboratories, one scanner, so staining variation is isolated rather than tangled with everything else. Useful as the controlled complement to TCGA’s uncontrolled variation.
External Validation — a model developed on TCGA and never tested elsewhere has been validated against a single, unusual, multi-institution convenience sample; the “external” framing that vendor evaluations get does not apply here.
Foundation Models in Pathology — TCGA slides are in the declared pretraining corpus of half the published tile encoders, so evaluating one of those on a TCGA-derived task is not a clean held-out test. The table above says which.
Encoder Benchmarking — the practical consequence: any local encoder comparison has to keep a TCGA tier separate from a genuinely external one, or it cannot tell representation quality from recall of the pretraining set.
AIDA Data Hub — the gated alternative: small, consecutive, pathologist-annotated Swedish cohorts instead of a large uncurated convenience sample. The better choice when the question is external validity rather than scale.
Lerousseau 2020 — MIL Tumor Segmentation (percentile pseudo-labels) — the source for the frozen-slide half of this page: labels from the barcode, tumour maps for the whole frozen breast/kidney/lung collection, and the measured warning that frozen-trained models mostly fail on FFPE.
Weakly Supervised Learning — what a 6,481-slide zero-annotation corpus buys: the barcode label is exactly the weak label that MIL formulations consume.
Open questions
Which public tile encoders include TCGA in their pretraining data?Answered for 16 of them — see the table above. The residual question is narrower and harder: the seven models declaring a proprietary corpus disclose nothing about whether TCGA-derived material sits inside it, and those are exactly the seven worth using. Nothing short of a vendor statement settles that. [unverified]- Would a TCGA pilot be an acceptable way to de-risk a local fusion or survival project before requesting Memorial data? It costs nothing but disk and time.
- Storage: the full image set is ~13 TB. Is there anywhere here that could hold even a single-entity subset? Unrecorded. [unverified]