The code was tested in Keras with Tensorflow backend. The packages needed are listed in the requirements.txt (not all packages in file are used. Just did a pip freeze for my virtual environment.)

pip install virtualenv
virtualenv --no-site-packages vkeras
source vkeras/bin/activate
pip install -r path/to/requirements.txt

For each data-set:

• Convert each instance into a numpy array with dimensions [height, width, channels] and save in a pre specified folder.
• Create a csv file containing the following columns: ['instance_file_name', 'bag_name', 'bag_label', <additional information,>, 'Cross-validation_split'] for each instance. instance_file_name can be bag_name + a unique number, Cross-validation_split should be a number starting from zero.

• Adjust hyper parameters and paths defined in the file hyperparameters.py
• Run train_3d_evt_mil_cv.py

The AUC values for validation data, at each epoch, will be writen to a .csv file in ./outputs folder

If you find this work useful in your research, please consider citing:

@ARTICLE{8805413,
author={R. {Tennakoon} and G. {Bortsova} and S. {Ørting} and A. K. {Gostar} and M. M. W. {Wille} and Z. {Saghir} and R. {Hoseinnezhad} and M. {de Bruijne} and A. {Bab-Hadiashar}},
journal={IEEE Transactions on Medical Imaging},
title={Classification of Volumetric Images Using Multi-Instance Learning and Extreme Value Theorem},
year={2019},
volume={},
number={},
pages={1-1},
doi={10.1109/TMI.2019.2936244},}