AIDA Data Hub

A Swedish national repository of annotated whole-slide cohorts, free but gated behind a PhD requirement and an institutional data-sharing agreement — the most realistic source of a ready-made external validation cohort for this group.

What it is

AIDA (Analytic Imaging Diagnostics Arena) is a Swedish national initiative based at Linköping University, and its Data Hub publishes whole-slide cohorts with expert annotations under persistent DOIs. It is not a portal in the TCGA sense — there is no bulk download and no open tier. Each dataset is individually applied for.

The access conditions are the thing to read first, because they decide whether this is usable at all. In the hub’s own wording, an applicant must “hold at least a PhD degree in a relevant field” and be “an authorized signatory who can legally enter into data sharing agreements on the behalf of the institution”, with use restricted to “legal and ethical medical diagnostics research”. So it is free of charge but not frictionless: it needs a named qualified researcher and someone with institutional signing authority, which in practice means involving a legal or research office rather than filling in a form.

Two cohorts are documented here because both appear in work already filed in this repo.

LNCO2 — colon adenocarcinoma lymph nodes

Full title Regional lymph node metastasis in colon adenocarcinoma, second collection series (doi 10.23698/aida/lnco2, 2020; Maras, Lindvall and Lundström; slides from Region Gävleborg Clinical Pathology and Cytology, copyright Linköping University).

Slides	1,245 WSIs from 50 consecutive patients
Stain	H&E
Scanners	Aperio ScanScope at 20× and Hamamatsu NanoZoomer at 40×
Resolution	median 0.46 µm/pixel; median dimensions 49,920 × 38,016
Annotations	2,598 — lymph nodes labelled negative (948), positive (75) or unassessed (1,517), plus tumour polygons and exclusion regions for artefact

Two features make it unusually good for validating a metastasis detector. Patients are chronologically consecutive, so it is a real case series rather than a curated positive-rich set; and it carries explicit exclusion regions for artefact, which most public cohorts leave to the user to guess.

Two features make it demanding. Only 75 of 1,023 assessed nodes are positive — under 8%, so any accuracy figure on it needs Class Imbalance and Accuracy read first. And most annotations are “unassessed”, meaning the reference standard covers a minority of the nodes present.

DRSK — DROID skin

Skin data from the Visual Sweden project DROID (Linköping University, 2019, v1.1.0). 99 H&E WSIs, 49 abnormal and 50 normal, patients aged 20–93; 16,741 annotations across 13 abnormality categories (actinic keratosis, basal cell carcinoma, dermatofibroma and others) plus tissue structures — epidermis, dermis, subcutis — and surgical margins. 20× single plane on Aperio ScanScope AT and Hamamatsu NanoZoomer XR/XRL. 32.36 GB. Annotated by one physician and verified by a second pathologist.

The margin and tissue-structure annotations are the unusual part: most public skin sets label lesions only, and margins are what a report actually turns on.

Why it matters for my work

It is the shortest path to an external cohort. External Validation is this wiki’s second-largest known gap — no project here has an outside test set, and PinkKidney is explicitly one external cohort away from a publishable claim. Assembling a second institution’s slides locally means governance, scanning time and a collaboration. Applying to AIDA means a form and a signature, and the slides already exist with annotations attached.

It varies the right things. LNCO2 was scanned on two vendors at two magnifications by a different health system on a different continent from anything here. That is the Scanner and Stain Variability axis exercised properly rather than nominally, which is what makes it a real test and not a second internal one.

There is a worked precedent to learn from. Wang et al. used a 321-slide subset of LNCO2 as the external cohort for an agent trained entirely at Stanford (sources/papers/wang-2026-pathology-cot.md), and released per-slide predictions — so the cohort’s behaviour under a genuine domain shift is already partly documented, including the failure modes.

That precedent carries two warnings worth having before anyone applies.

The subset is not the cohort, and how it was chosen is not stated. The paper describes “321 WSIs from Sweden”; the hub says 1,245 slides from 50 patients. Recomputing from the released predictions gives 321 slides from 22 patients, 41 positive and 280 negative. The most likely explanation is that only nodes with an assessed label are usable and most are unassessed, which would make the subsetting sensible and simply undescribed — but it is an inference, not something either source states. [unverified] Anyone reusing that 321-slide split to compare against the published numbers needs to reconstruct it first, and nothing published says how.

The resolution shift is larger than the magnification labels suggest. The training slides were Leica Aperio 40× at 0.25 µm/pixel; LNCO2’s median is 0.46 µm/pixel — nearly half the linear resolution, so a fixed-pixel crop covers close to four times the tissue area. For any method whose high-power step is defined as a crop at native resolution, that is not a nuisance parameter, it is a change in what the model is looking at. It is the Patch Extraction argument in its sharpest form, and a plausible contributor to the navigation stage failing on 23% of slides there — see Model Abstention.

How it connects

External Validation — the gap this exists to close, and the reason a gated free cohort beats an ungated hypothetical one.

TCGA — the other public option, and the contrast: TCGA is open, huge, multimodal and a convenience sample assembled without imaging standardisation; AIDA cohorts are small, gated, consecutive and annotated by pathologists for a stated diagnostic task. Different tools.

Labquality EQA Staining Dataset — the third public resource here, and the controlled extreme: one block, 66 laboratories, one scanner. Between the three, staining variation, institutional variation and scanner variation can each be isolated.

Class Imbalance and Accuracy — LNCO2’s positive rate is under 8%, so any headline accuracy on it is close to meaningless without the base rate stated beside it.

Scanner and Stain Variability — two vendors at two magnifications in one cohort is what makes it a test rather than a second internal set.

Patch Extraction — 0.46 versus 0.25 µm/pixel decides what any fixed-size tile actually contains, and has to be resolved before a model trained here is run there.

Agentic Slide Navigation — the published precedent used LNCO2 to test whether a learned viewing policy survives a scanner change, which is the transferability question that matters most for that whole approach.

Mitotic Count — the DRSK margin and tissue-structure annotations are the kind of reference standard that area-calibrated measurement needs, and a reason to look at skin even though it is not a current project.

Open questions

Does an MD with a specialty and an academic appointment satisfy “at least a PhD degree in a relevant field”? Not answerable from the page. [unverified] It is the first thing to establish, because everything else is wasted effort if not, and the hub can be asked directly.
Who at the institution can act as authorised signatory for a data-sharing agreement? Unrecorded here, and the likelier bottleneck of the two.
Would LNCO2 serve as an external cohort for anything currently running, or only for a metastasis-detection project nobody has started? PinkKidney needs a renal external set, which this is not. Worth checking what else AIDA publishes before assuming the answer is no.
Is there a Turkish or European equivalent with lighter access terms? Nothing in this wiki records a survey of public WSI repositories, and TCGA, Labquality EQA Staining Dataset and this page are the only three entries.