Read it, then run it
This page is the read-only chapter intro. To actually execute the cells (load the dataset, train a model, render GradCAM heatmaps on your own images), open the Colab notebook in a free T4 runtime. The curriculum is designed to be executed, not read.
Chapter 9: PCam Challenge¶
The PatchCamelyon Challenge¶
You have 90 minutes to achieve the highest test AUC on PatchCamelyon.
Rules¶
- Use any architecture from Chapters 3–4, the foundation model from Chapter 8, or define your own (Chapter 10)
- Use any hyperparameters, augmentations, and training strategies
- Must produce GradCAM visualization on your 5 best and 5 worst predictions
- Ranked by test AUC - highest AUC wins
Evaluation Criteria¶
| Weight | Criterion |
|---|---|
| 50% | Test AUC |
| 25% | GradCAM analysis quality |
| 25% | Methodology (did you try multiple approaches?) |
This is a competition, not an exam. Experiment boldly, track what works and what doesn't, and be prepared to discuss your findings.
About PatchCamelyon (PCam)¶
PatchCamelyon is derived from the Camelyon16 Grand Challenge, one of the most influential challenges in computational pathology.
The task: Classify 96×96 pixel RGB histopathology patches as containing metastatic tissue (positive) or normal tissue (negative) in lymph node sections.
Clinical relevance: Pathologists spend significant time reviewing lymph node biopsies for metastatic cancer cells. An accurate classifier could dramatically reduce workload and improve detection of small metastases that might be missed during manual review.
Key differences from MedNIST: - Binary classification (2 classes, not 6) - RGB color images (3 channels, not 1 grayscale channel) - 96×96 pixels (different from MedNIST's 64×64) - More challenging - subtle differences between metastatic and normal tissue