AI for Pathologists

Learn the AI methods of digital pathology with small networks trained in your browser and real mutation data

What this site is for

The QuPath workshop for pathologists teaches measurement on slides. This site is its sequel: it shows how the AI methods beyond those measurements work (weak supervision, foundation models, shortcut learning) and when they fail. No code is needed; every simulation runs in the browser. All simulations train on synthetic data and none produces a clinical interpretation.

Two kinds of material

Series

Interactive simulations

Small networks trained in the browser, each showing one method together with one failure mode. In the style of Nucleus Net: Slides → Train → Test.

Simulation 1 · ready

Slide Bag Net

Multiple-instance learning (MIL): an attention network that learns whether a tumour carries a given mutation from slide-level labels only. Pooling, attention maps, the lab batch-effect trap.

Simulation 2 · ready

Stain Shift

Macenko stain vectors, normalising lab B onto lab A, and stain augmentation; the same trap retrained with four preprocessing choices and a results table.

Module · ready

Mutation data with cBioPortal

The real counterpart of the simulation's synthetic labels: mutation frequency, co-occurrence and survival in TCGA cases, and how the slide-level label file is built.

Suggested order

1. Slide Bag Net, 40 minutes. Run the simulation in the order of the lecture recipe on its page; answer the self-check questions.
2. Stain Shift, 35 minutes. Same slides; two cures for the trap.
3. Mutation data with cBioPortal, 60 minutes. The web interface first (no code), then optionally the API part (Python).
4. Series page: which simulations come next and which lesson each one carries.

Scope boundary

Nothing on this site diagnoses, proposes a threshold or produces a clinical decision. The simulations run on synthetic data; the cBioPortal module uses public research data (TCGA) only. The clinical meaning of any number belongs to the pathologist.