butkej / mil4cyto

Released code for the paper 'End-to-end Multiple Instance Learning for Whole-Slide Cytopathology of Urothelial Carcinoma'

MIL4Cyto

Released code for the paper 'End-to-end Multiple Instance Learning for Whole-Slide Cytopathology of Urothelial Carcinoma' by Joshua Butke, Tatjana Frick, Florian Roghmann, Samir F. El-Mashtoly, Klaus Gerwert and Axel Mosig.

Accepted to and presented at MICCAI 2021 Workshop Computational Pathology (COMPAY 2021).

Overview

This repo contains the Python code of our experiments, however there is no data included. Still, this implementation might serve as a starting point for those interested in applying Attention-based Multiple Instance Learning to problems of cytopathology.

Citation

Please cite our paper, if this work is of use to you or you use the code in your research:

    @inproceedings{butke2021end,
        title={End-to-end Multiple Instance Learning for 
            Whole-Slide Cytopathology of Urothelial Carcinoma},
        author={Butke, Joshua and Frick, Tatjana and Roghmann, Florian 
            and El-Mashtoly, Samir F and Gerwert, Klaus and Mosig, Axel},
        booktitle={MICCAI Workshop on Computational Pathology},
        pages={57--68},
        year={2021},
        organization={PMLR}
    }

Requirements

Packages:

• Pytorch (>= 1.6.0)
• OpenCV (4.4.0)
• sklearn (0.23.0)
• matplotlib (3.3.0)

Hardware:
We used a cluster equipped with 4 NVIDIA V100 GPUs, which is reflected in joshnet/custom_model.py where blocks of layers are assigned to dedicated cards.

Further Reading

I highly recommend to check out the original paper and implementation of Ilse et al. for Attention-based MIL, that can be found here, as well as the Li et al. paper 'Deep Instance-Level Hard Negative Mining Model for Histopathology Images'. The second one introduced Hard Negative Mining that I adopted. However they never released any code, so I implemented their improvements as best as I could.

Contact

If you have any questions you can contact me at [email protected], however we do not gurantee any support for this software.

Acknowledgements

This work was supported by the Ministry for Culture and Science (MKW) of North Rhine-Westphalia (Germany) through grant 111.08.03.05-133974 and the Center for Protein Diagnostics (PRODI).