Raymvp12/focusmil-camelyon16

Datasets used in FocusMIL paper

The Camelyon16 and Camelyon16-Standard-MIL-test datasets used in From Correlation to Causation: Max-Pooling-Based Multi-Instance Learning Leads to More Robust Whole Slide Image Classification (FocusMIL).

Code: FocusMIL-and-other-max-pooling-methods

These are pre-extracted patch features — you do not need the raw whole-slide images. Slide classification, patch-level metrics, and FROC localization all run directly off the files here.

For Camelyon17, see the AEM repository, whose publicly released features we use.

What's in here

Directory	data_tag	Backbone	Dim	Size
Camelyon16feat_256_10x/	c16_real	ResNet18 (ImageNet)	512	5.3 GB
CTransPath_feature/	c16_real_ctrans	CTransPath (SSL)	768	7.9 GB
camelyon16-standard-MIL-test/	c16_semi075	ResNet18, semi-synthetic	512	5.3 GB
froc_features/	—	per-slide feats + level-0 coords	—	4.3 GB
masks_c16_test.tar.gz	—	47 GT tumor masks (tarball)	—	60 MB

Patches are 256×256 at 10×. 23 GB total (27 GB on disk once the masks are extracted).

Usage

Download, mask extraction, directory layout, file format, and the training commands are documented in the code repo — see docs/DATASETS.md.

hf download Raymvp12/focusmil-camelyon16 --repo-type dataset --local-dir /path/to/camelyon16

MANIFEST.sha256 lets you verify nothing was truncated:

cd /path/to/camelyon16 && grep -v '^#' MANIFEST.sha256 | awk '{print $1"  "$3}' | sha256sum -c -

The Standard MIL Test (camelyon16-standard-MIL-test/)

A controlled probe of the standard MIL assumption, not a second copy of the data.

Attention-based and TransMIL-style methods pool instances through a learnable weighted combination, so they can exploit negative evidence — a pattern whose presence correlates with the negative bag label.

This benchmark plants exactly such a pattern. In the training set we introduce a poison by randomly selecting 20% of the patches in the normal slides and increasing the intensity of their green channel — a negative shortcut: a cue that correlates with the negative bag label but is causally irrelevant to tumour. In the test set we poison 20% of the patches in the tumour slides in the same way, so the shortcut now points the wrong way.

A method that leans on negative evidence is fooled and its slide AUC collapses; a max-pooling model, which predicts only from positive evidence, stays robust.

Any method whose slide AUC falls below 0.5 on this benchmark violates the standard MIL assumption — it is predicting from negative evidence rather than from positive evidence.

Format and labels are identical to c16_real; only the features change. Select it with the c16_semi075 data tag, and compare its slide AUC against c16_real.

Citation

@misc{liu2025correlationcausationmaxpoolingbasedmultiinstance,
      title={From Correlation to Causation: Max-Pooling-Based Multi-Instance Learning Leads to More Robust Whole Slide Image Classification},
      author={Xin Liu and Weijia Zhang and Wei Tang and Thuc Duy Le and Jiuyong Li and Lin Liu and Min-Ling Zhang},
      year={2025},
      eprint={2408.09449},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.09449},
}

Please also cite the original Camelyon16 challenge.

License

The underlying Camelyon16 data is released under CC0 1.0 by the challenge organizers; these derived features are distributed under the same terms.