Studer 2023 — Tumor Budding T-cell Graphs (pT1 CRC resection risk)

Graph neural networks over tumour buds and T-cells in the ITBCC hotspot of pT1 colorectal polyps, asking whether the patient actually needed the colectomy — honest metrics under 6:1 imbalance, a released graph dataset, and a headline picked from 144 configurations with no external cohort.

Graph neural networks over tumour buds and T-cells in the ITBCC hotspot of pT1 colorectal polyps, asking whether the patient actually needed the colectomy — honest metrics under 6:1 imbalance, a released graph dataset, and a headline picked from 144 configurations with no external cohort.

Read status: full text (camera-ready PDF) plus a survey of the companion repo. One mechanical caveat: the main results table’s text layer extracts with columns misaligned, so per-configuration numbers below are quoted only where a row is unambiguous; the paper’s own prose carries the headline claims.

The clinical question, which is the good part

Less than 15% of pT1 CRC patients have nodal metastasis, yet the cited Bern registry data show 55% underwent colectomy, and Zwager et al. put >90% of “high-risk” patients as node-negative — because guidelines recommend surgery on one unfavourable histological factor. The paper asks whether the spatial interplay of two established factors — tumour budding and CD8+ T-cell infiltration at the invasive front — can rescue specificity without giving up sensitivity. The label is unusually honest for this literature: did the patient, in retrospect, need the resection — N+ at resection or recurrence within ≥36 months of follow-up, not a surrogate.

What they built

From each WSI: buds and T-cells are detected by CNNs (buds: the Bokhorst student–teacher model, public on Grand Challenge; T-cells: U-Net with Gaussian-filter/regional-maxima post-processing, run after colour deconvolution because it was trained on single-stain IHC). A pathologist selects the ITBCC hotspot (0.785 mm² at the invasive front); every detection inside becomes a node. Three graph structures are compared — plain Delaunay; Delaunay-Star (buds triangulated, T-cells attached to all buds within 100 µm); Hierarchical (T-cells to nearest bud within 100 µm, then clique within each bud’s group, buds fully connected) — with node labels stacked from: type (bud/T-cell), zero-centred coordinates, and a 384-d ImageNet-DINO ViT-16 embedding of a 200×200 crop. Edges carry distance in µm. GraphSAGE / GIN / GATv2, with and without jumping knowledge, with and without the four SGG clinical variables appended at the classification head — 144 setups, 5-fold patient-level CV, 5 seeds ensembled by averaged softmax, patient-level evaluation via max over a patient’s graphs.

The four fields, all actually stated

Dataset	626 WSIs / 575 patients, eight pathology institutes (Bern plus, per the acknowledgments, Bamberg, Bayreuth, Baselland, Scott&White among others); 541 low-risk vs 85 high-risk
Stain / scanner / magnification	CD8–AE1/AE3 double IHC; Pannoramic 250 at 0.243 µm/px — one scanner for the whole cohort
Validation	5-fold CV, patients non-overlapping across folds, class-stratified; metrics chosen for imbalance (per-class-average F1, TNR, TPR — accuracy never quoted)
External validation	None — named as future work in the conclusion

Results, as carefully as the extraction allows

The SGG-guideline baseline (the clean row): TNR 31.4 ± 4.7, TPR 88.3 ± 5.1, average F1 28.2 ± 4.4. The paper’s prose: the best model (GraphSAGE-JK on Hierarchical graphs with type + coordinates + ViT-16 features plus clinical data) improves both TNR and TPR over that baseline, and the conclusion claims a 20% decrease in falsely-identified high-risk patients with no loss of sensitivity. Best-row TNR values in the table sit in the low-40s against the baseline’s 31.4. Two observations the authors themselves make and deserve credit for: the simplest representation (Delaunay, type-only labels, no image features) performs well with the lowest fold-variance, and they explicitly invoke Ockham’s razor for it — noting it also sidesteps staining-variation bias entirely, since no pixel ever enters the graph; and GIN configurations overfit visibly (fold SDs up to ±22 TNR points).

Reading it critically

1. The headline is the best of 144 configurations, selected on the same cross-validation it is reported from. The selection rule (top-3 by F1 and mean recall, then highest TPR, discarding TNR below baseline) is applied post hoc to the full grid; there is no nested CV and no held-out selection set. With fold SDs of ±5–22 points, the winner of a 144-way race is optimistic by construction. The honest reading: some graph configuration beats the SGG baseline under CV; the specific “+20%” awaits confirmation on data that did not pick the winner.
2. Eight institutes in, zero institutes held out. Folds are split by patient, not by site, so every fold trains and tests on the same eight labs’ staining. One scanner removes scanner variance but the stain variance of eight labs is in both halves of every split — a leave-one-site-out arm would have cost nothing but compute and would have been the paper’s own internal external-validation. Its absence is the gap between this and a deployable claim. See Scanner and Stain Variability and External Validation.
3. Graphs inherit detector error silently. Node sets come from CNNs whose reviewed-subset precision/recall is 89.9%/93.0% — every missed bud is a missing node, every false T-cell a spurious one, and no sensitivity analysis to detection quality is run. The paper’s own defence (graph creation “is agnostic to the way lymphocytes and tumor buds are detected”) is an architecture point, not an error-propagation one.
4. Absolute performance is far from clinical. Average F1 around 40% and TNR around 40% at TPR ~85% would still send the majority of low-risk patients to surgery. The paper’s framing (better than the current guideline baseline, whose TNR is 31.4%) is fair; any stronger reading is not supported.
5. What it does right under imbalance is worth copying: per-class-average F1, TNR and TPR reported separately, accuracy never mentioned, and the imbalance (6.4:1) stated up front — Class Imbalance and Accuracy as practised.
6. “Publicly available” versus Zenodo’s gate. The abstract and conclusion say code and dataset are public; the Zenodo record is CC BY-NC-SA but listed access-restricted. Presumably request-gated — [unverified], and it changes what “available” means for anyone planning to use it. Recorded on the dataset page.

Why it earns space here

1. Tumour budding is a live local interest with no filed literature until now — the estate ships a budding module in the QuPath workshop extension and forks Bankhead’s budding scripts, and Bern (Zlobec, Dawson) is the ITBCC’s home turf. This is the budding cluster’s first paper note, and it comes with a dataset.
2. It is the wiki’s first worked example of cell-graph representation learning — spatial structure fed to a GNN rather than summarised into counts or persistent-homology features; the natural comparator for Persistent Homology’s “make arrangement computable” idea, arrived at by explicit edges instead of topology.
3. The metric discipline under imbalance is a model for the group’s own imbalanced problems, and the selection-over-144 flaw is the equally instructive counterweight.

Related: pT1-HBTG — the released dataset, filed as its own page with access conditions. Related: Class Imbalance and Accuracy — this paper as the worked positive example of metric choice under 6:1 imbalance. Related: External Validation — problem 2 above: multi-site data pooled instead of held out, so externality was available and unused. Related: Persistent Homology — the other route from cell arrangement to a number; explicit graphs versus topological summaries is a choice nobody here has written down. Related: Cell Detection and Counting — the detectors whose output becomes the nodes, and whose error the graphs inherit. Related: KUCell — the other public dataset filed here from a nearby research culture; different task, same precedent of a pathology group releasing its own annotated material.

The companion GNN repository (7 commits, last 2023-05-15, single author, Apache-2.0, PyTorch-Geometric + Lightning + W&B) gets no separate project page per the §6 gate and the curvereg precedent: it is the paper’s implementation, surveyed rather than adopted. The separate BT-graph-creation repository — the WSI-to-graph pipeline, which is the piece the group could actually reuse against its own budding material — was not read and is recorded as a pointer only.