MixMIL

Code for the paper: Mixed Models with Multiple Instance Learning

Accepted at AISTATS 24 as an oral presentation & Outstanding Student Paper Highlight.

Please raise an issue for questions and bug-reports.

Installation

Install with:

pip install mixmil

alternatively, if you want to include the optional experiment and test dependencies use:

pip install "mixmil[experiments,test]"

or if you want to adapt the code:

git clone https://github.com/AIH-SGML/mixmil.git
cd mixmil
pip install -e ".[experiments,test]"

To enable computations on GPU please follow the installation instructions of PyTorch and PyTorch Scatter. MixMIL works e.g. with PyTorch 2.1.

Experiments

See the notebooks in the experiments folder for examples on how to run the simulation and histopathology experiments.

Make sure the experiments requirements are installed:

pip install "mixmil[experiments]"

Histopathology

The histopathology experiment was performed on the CAMELYON16 dataset.

Download Data

To download the embeddings provided by the DSMIL authors, either:

• Full embeddings: python scripts/dsmil_data_download.py
• PCA reduced embeddings: Google Drive

Microscopy

The full BBBC021 dataset can be downloaded here.

Download Data

• We make the featurized cells available at BBBC021
• The features are stored as an AnnData object. We recommend using the scanpy package to read and process them
• The weights of the featurizer trained with the SimCLR algorithm can be downloaded from the original GitHub repository

Citation

@inproceedings{engelmann2024mixed,
  title={Mixed Models with Multiple Instance Learning},
  author={Engelmann, Jan P. and Palma, Alessandro and Tomczak, Jakub M. and Theis, Fabian and Casale, Francesco Paolo},
  booktitle={International Conference on Artificial Intelligence and Statistics},
  pages={3664--3672},
  year={2024},
  organization={PMLR}
}