Introduction to Pathology MIL — Course Package

A self-contained course on Multiple Instance Learning (MIL) for computational pathology, aimed at ML engineers. Covers the full pipeline from gigapixel whole-slide images to interpretable, weakly-supervised slide predictions, with generic methods (ABMIL / CLAM / TransMIL) and modern foundation-model encoders (UNI2, Virchow2, Prov-GigaPath, CONCH).

Format: 50-minute lecture + 10-minute Q&A · one hands-on end-to-end TCGA notebook · NotebookLM quiz.

Contents

milcourse/
├── deck/
│   └── Pathology_MIL_Course.pptx     ← 49-slide deck (upload to Google Drive → opens as Google Slides)
├── notebooks/
│   ├── pathology_mil_tcga.ipynb      ← ONE end-to-end notebook (download→encode→train→infer→heatmap→eval)
│   ├── build_notebook.py             ← regenerates the notebook (embeds the helper modules)
│   ├── mil_models.py                 ← reference MIL aggregators (mean/max, ABMIL, CLAM-SB)
│   ├── mil_tcga.py                   ← TCGA pipeline: GDC download, WSI seg/patch, Midnight-12k, cache
│   ├── mil_utils.py                  ← leakage-safe CV, metrics, train/eval loops
│   └── requirements.txt
└── quiz/
    └── MIL_quiz_source.md            ← 20-question quiz + study guide (NotebookLM source)

Slides

deck/Pathology_MIL_Course.pptx — drag it into Google Drive and it opens as a fully editable Google Slides deck (for sharing/co-editing). Speaker notes are included on every content slide. Sections: Background · MIL formulation · Data preparation · Foundation encoders · Model architecture · Training protocols · Inference · Post-processing · Evaluation & visualization · Conclusion + quiz.

Notebook — real TCGA pipeline (single, end-to-end)

notebooks/pathology_mil_tcga.ipynb runs the whole pipeline in one Colab runtime on real TCGA whole-slide images (no synthetic data) for NSCLC subtyping — TCGA-LUAD vs TCGA-LUSC:

GDC download → tissue seg → 20× patching → Midnight-12k features → train → infer → heatmaps → evaluate

Because everything runs in one runtime, each section reuses the previous section's in-memory variables — there's no cross-notebook hand-off. The expensive download+encode is cached (to Google Drive if mounted, else locally under pathology_mil_tcga/), so re-running after a runtime restart skips it. The build step is idempotent — already-cached slides are skipped.

Encoder: Midnight-12k (kaiko-ai/midnight, MIT, ViT-g, 3072-d) — open, no gating, and the top non-gated pathology foundation model (beats gated H-optimus-0 / Prov-GigaPath / UNI on Kaiko's benchmark; owkin/phikon and owkin/phikon-v2 are lighter open alternatives).

To run: open the notebook in Colab → Runtime → Change runtime type → GPU → Run all. The first cell installs deps, pulls the helper modules from this repo (git clone, or uses the local copy when run inside the repo), and optionally mounts Drive for the cache. ⚠️ The download+encode step does real work and takes a while.

Note: the Colab setup clones this repo to get mil_tcga.py / mil_models.py / mil_utils.py, so push the repo to GitHub (public) first. If the repo is private, mount Drive and copy the notebooks/ folder there, or run the notebook from a local clone instead.

# local (non-Colab) use also works:
cd notebooks && pip install -r requirements.txt   # + the OpenSlide system library

Knobs (cell "1 · Configuration"): PER_CLASS (slides per class, default 50 ≈ 100 slides — lower to go faster), MAX_PATCHES (per-slide patch cap, default 2000), ENCODER (default kaiko-ai/midnight).

Quiz

Upload quiz/MIL_quiz_source.md (optionally with the deck and notebooks) to NotebookLM to generate an audio overview, flashcards, a study guide, and an auto-quiz. The file also contains a standalone 20-question quiz with a full answer key.

A note on the numbers

The notebooks now compute metrics on real TCGA data, but the default cohort is small (~100 slides at the default PER_CLASS=50, a teaching set). For real claims, increase PER_CLASS further and validate on an external cohort (a different hospital/scanner). Performance figures on the slides remain illustrative diagrams, not measured benchmarks.