Meningioma H&E Deep Learning

An attempt to reproduce, then retrain on local cases, a published model that predicts meningioma molecular groups and recurrence risk from H&E alone — labels that normally require DNA methylation.

Purpose

The target study is Landry, Nassiri et al., Deep learning for H&E-based meningioma molecular classification and outcome prediction (Lancet Digital Health 2026, doi 10.1016/j.landig.2026.100986). It trains five models that take only a whole-slide H&E image and predict:

  1. Molecular group — MG1 immunogenic, MG2 NF2-wildtype, MG3 hypermetabolic, MG4 proliferative
  2. Recurrence risk group — high or low
  3. Chromosome 1p loss
  4. Chromosome 22q loss
  5. Chromosome 1q gain

What makes this worth a page rather than a bookmark is the kind of claim being reproduced. These are not morphological labels a pathologist could assign by eye and the model is learning to imitate. They are molecular labels whose ground truth comes from DNA methylation arrays, and the claim is that the information survives into the H&E appearance well enough to be recovered. If it holds locally, it substitutes a slide the department already has for an assay it does not routinely run.

There is currently no code — the repository holds a design spec, a superpowers plan, the source paper, and a Zenodo access request to the original authors. That is an honest state for a project four weeks old and it is the right order of work.

Data used

Stated plainly in the spec, which is unusually disciplined about what is and is not in hand:

In hand now H&E WSIs only, 50–200 cases
Arriving later DNA methylation .idat, clinical outcomes (PFS, grade, resection, age), pathologist molecular reports
Goal Validate the published model, then retrain — both
Compute Local NVIDIA GPU workstation, Windows 11
Reproducibility “Pragmatic” — scripted stages, pinned environments, caching, seeds

The cohort size is the binding constraint and the spec does not hide it. The published models were trained on a multi-institution methylation-labelled cohort; 50–200 local cases is a validation set, not a training set, for a four-class molecular target. The “validate then retrain” ordering is therefore the right one — and the honest expectation is that validation is achievable and retraining is not, at least until the cohort grows.

The methylation ground truth is what makes this hard. Without .idat files there is no local label for molecular group at all, so until those arrive the only testable predictions are the ones with an independent clinical readout — recurrence risk against actual follow-up.

Methods

The paper’s pipeline, recorded in the spec: WSI → 512×512 tiles at 20× → colour normalisation → UNI foundation-model features (1024 per tile) → MLP classifier (1024 → 1024 → 2 or 4) → 5×5 nested cross-validation → tile predictions averaged to patient level → AUC, balanced accuracy, log-rank and Cox.

Three things about that pipeline are worth drawing out because they connect to decisions already recorded elsewhere in this wiki:

  • UNI is the encoder, and it is already on the Foundation Model Cache drive. So the feature-extraction half of this pipeline can be run today without downloading anything, and PinkKidney has already built a harness that runs UNI over a cohort and caches per-slide features. The marginal cost of starting is low.
  • Colour normalisation is a step in the published pipeline, which puts this project directly into the argument on Stain Normalisation and Colour Calibration — and squarely into the Khan finding on Foundation Models in Pathology that UNI embeddings shift under staining variation. Reproducing on local slides from a different laboratory is exactly the condition where that matters.
  • Tile-averaging to patient level is the weakest link in the design, and the spec inherits it from the paper rather than choosing it. Averaging assumes the signal is diffuse across the section; if the molecular phenotype is regional, mean pooling dilutes it. This is precisely the aggregation step that a slide-level model would replace — see the second generation described on Foundation Models in Pathology, where TITAN and CHIEF are already local.

Delivery is intended as a Quarto book that doubles as an executable pipeline, updated as the work proceeds — the same pattern as PembeBobrek.

Current state / open questions

Design spec dated 2026-06-07, marked “draft for review”, five commits, nothing since 2026-06-08. Private repository. A Zenodo access request to the original authors is logged in docs/correspondence/, which suggests the published weights or data are gated and the reproduction depends on an answer that may not come.

  • The validation is blocked on labels, not on code. Without methylation .idat there is no molecular ground truth locally, so the only near-term testable target is recurrence risk against clinical follow-up. Worth deciding explicitly whether to start there rather than waiting.
  • Nothing records the scanner or magnification of the local cohort. The paper’s pipeline is specified at 20×, and this wiki’s standing gap about unrecorded acquisition metadata applies here before any tiling is done — see the Known gaps list. Getting it right at the start is cheaper than discovering a mismatch after feature extraction.
  • [unverified] whether the published model weights are obtainable at all. If only the method is available and not the trained model, “validate” becomes “reimplement and train”, which is a different project with a much larger cohort requirement.
  • No external cohort is contemplated, and for a 50–200 case single-institution study that is the usual limitation — External Validation applies with full force to any claim that comes out of this.

Related: Foundation Models in Pathology — UNI is the encoder this depends on, and the tile-averaging step is the one a slide-level model would replace. Related: Foundation Model Cache — UNI is already local, so feature extraction can start without any download. Related: WHO CNS — Who Cites Who — the CNS classification context these molecular groups sit inside. Related: External Validation — a single-institution cohort of this size makes this the binding question for any result. Related: Stain Normalisation — a required step in the published pipeline, and the place a cross-laboratory reproduction is most likely to lose signal. Related: PinkKidney — the existing feature-extraction harness that would save building one.

Derived from: repository read 2026-07-29 — docs/superpowers/specs/2026-06-07-meningioma-he-dl-cohort-design.md (read in full for §1–2), docs/superpowers/plans/2026-06-07-phase1-foundation.md, docs/correspondence/2026-06-07-zenodo-access-request-zadeh.md, literature/; git log and gh repo view for currency and visibility.