pT1-HBTG

Bern/Radboud graph dataset of 626 tumour-budding hotspots from 575 pT1 colorectal cancer patients — bud and T-cell nodes with outcome-based risk labels — on Zenodo under CC BY-NC-SA but listed access-restricted, which is not the ‘publicly available’ its paper claims.

What it is

The pT1 Hotspot Tumor Budding T-cell Graph dataset, released with Studer 2023 — Tumor Budding T-cell Graphs (pT1 CRC resection risk) (MIDL 2023) at Zenodo DOI 10.5281/zenodo.7867085. Per the Zenodo record it contains:

Graphs in GXL — nodes are CNN-detected tumour buds and CD8+ T-cells inside the pathologist-selected ITBCC hotspot (0.785 mm²), with coordinates, type, and 384-d ImageNet-DINO features; edges carry inter-node distance in µm;
hotspot PNGs at full resolution and the 200×200-px patches the features came from;
class labels and the paper’s 5-fold patient-level CV splits as JSON.

Cohort behind it: 626 WSIs / 575 patients from eight pathology institutes, CD8–AE1/AE3 double IHC, one scanner (Pannoramic 250, 0.243 µm/px), 541 low-risk vs 85 high-risk with outcome-based labels — nodal status at resection, or ≥36 months recurrence-free follow-up.

Access, stated precisely

The paper says code and dataset are “made publicly available”. The Zenodo API lists the record as Restricted with licence CC BY-NC-SA 4.0 — metadata public, files request-gated. The record page itself (fetched 2026-08-23, second pass) states it plainly: “The record is publicly accessible, but files are restricted”, with a request form behind a Zenodo login whose stated condition is that “Users need to provide information on their research activities and affiliations, and what they inted [sic] to use the data for”. So: an approval step, not a download button — but one that demonstrably opens, since the record shows 80 unique downloaders (435 views) and 466.9 GB of data served as of that date. The contact person on the record is Heather Dawson. Non-commercial share-alike terms apply to whatever is granted.

Why it matters for my work

It is the budding cluster’s missing data. The estate ships a tumour-budding module in QuPath Atölye Extension and forks Bankhead’s budding scripts inside Digital Pathology Drafts, but owns no annotated budding material to run either against. These are real hotspots, selected per ITBCC by the group that wrote the ITBCC recommendations, with images included — the closest thing to a reference set the local budding tooling could be exercised on, subject to the access gate and the caveat that node positions are CNN detections (reviewed-subset precision 89.9% / recall 93.0%), not hand annotations.

It is also a different kind of release than any dataset filed here — graphs rather than images-plus-labels. The images are included, but the primary object is the derived spatial structure, which makes it directly reusable for graph-method work and only indirectly for detector training.

The precedent point KUCell carries applies here too, from a second direction: a pathology group packaging its own material for release, this time with outcome labels rather than dot annotations — and outcome labels are the expensive kind, since they need years of follow-up rather than hours of marking.

How it connects

Studer 2023 — Tumor Budding T-cell Graphs (pT1 CRC resection risk) — the paper it was released with, including the honest metric handling and the selection-over-144-configurations caveat that bounds its headline.

KUCell — the sibling precedent: another nearby group releasing annotated pathology data; dot-annotated cells there, outcome-labelled graphs here.

QuPath Atölye Extension — the local budding module this could exercise, if access is granted.

Cell Detection and Counting — the nodes are detector output, so every use of this dataset inherits the detectors’ error profile; treat the graphs as measured-with-noise, not ground truth.

External Validation — for any local budding or TIL work, this is an eight-institute outside cohort with outcomes; the paper itself never used its multi-site structure for validation, which does not stop a local user from doing so.

Class Imbalance and Accuracy — 541 versus 85: any model trained or evaluated on this inherits a 6.4:1 imbalance, and the release ships the stratified splits to handle it the way the paper did.

Second pass, same day (2026-08-23)

The same five links arrived in chat again a few hours after the first filing. The delivered PDF is byte-identical (md5) to the one already beside the note, so nothing new was filed — but the re-read settled four things the first pass had left open or got wrong, and since the paper note is hook-protected, they are recorded here, on the work’s editable page.

The paper note quotes the wrong SGG baseline row. The note gives the guideline baseline as TNR 31.4 ± 4.7 / TPR 88.3 ± 5.1 and calls it “the clean row”. Those numbers belong to the GraphSAGE-JK / Delaunay / type-only model — the row directly above. Table 2’s SGG Criteria Classification row reads average F1 28.2 ± 4.4, TNR 22.2 ± 4.3, TPR 85.0 ± 7.6, read visually from a rendering of page 7 (the text layer misaligns exactly as the note’s access line warned it might). Three corroborations independent of the rendering: the abstract’s “+20% specificity” only reconciles from 22.2 (best TNR 42.5 − 22.2 = 20.3; from 31.4 it would be 11); the paper’s selection rule discards models “with a TNR below the baseline”, yet Table 2’s selected rows include TNRs of 24.1, 27.9 and 30.0, impossible under a 31.4 baseline; and 31.4/88.3 appear verbatim as the GraphSAGE-JK row of appendix Table 7. The correction cuts in the paper’s favour twice: the guideline baseline is weaker on specificity than the note says, so the model’s improvement is larger — and the true baseline sensitivity of 85.0 is what makes the paper’s “no loss of sensitivity” (best model: 84.0) hold, where the wrong row’s 88.3 would have read as a four-point loss. Three body corrections are pending in the hook-protected note: the “clean row” sentence, “against the baseline’s 31.4” in the results paragraph, and criticism 4’s “whose TNR is 31.4%”.

The official citation exists. proceedings.mlr.press lists the paper as Medical Imaging with Deep Learning, PMLR 227:235–259 — completing the note’s cite_as:, which had (correctly) refused to guess a volume from the PDF header’s “88”. Frontmatter edit pending.

Two code-versus-paper discrepancies, from reading the GNN repo’s config and evaluation script rather than surveying its structure:

The ensemble is five models, not ten: config/bts_midl.json sets multi_run: 5 and util_scripts/softmax_to_results.py hardcodes the first five softmax columns (iloc[:, 0:5]), agreeing with §4.1’s “run 5 times with a different seed” — Figure 1’s caption (“an ensemble of 10 models”) is wrong.
The paper says 196 neurons per message-passing layer, three times, including the tuning grid [128, 196, 256]; the released config says nb_neurons: 192. Which value produced the published numbers is [unverified] — the authors’ Weights & Biases logs (project names sit in the repo README) could settle it.

The OpenReview reviews remain unread — both URLs still dead-end at the browser-verification gate, confirmed again from this machine.

Open questions

What the restriction requires is now answered (second pass, 2026-08-23): a Zenodo login plus a justification form — research activities, affiliation, intended use (see Access above). Still open: who approves, and how fast. The 80 granted downloads say requests succeed; nothing on the record says how quickly.
Are the raw WSIs available at all, or only hotspots and patches? The record lists hotspot PNGs and patches; nothing suggests whole slides. For budding-module testing the hotspots suffice; for anything upstream (detection, hotspot selection itself) they do not.
Could the local budding module’s output be compared against these graphs’ bud nodes as a cheap agreement study? It would need only the hotspot images and the GXL node lists — but the reference is itself CNN output, so the comparison would measure agreement between two detectors, not accuracy against truth.