Code for our MICCAI 2021 paper "AMINN: Autoencoder-based Multiple Instance Neural Network Improves Outcome Prediction of Multifocal Liver Metastases".
by Jianan Chen (chenjn2010@gmail.com), Helen M. C. Cheung, Laurent Milot and Anne L. Martel (anne.martel@sunnybrook.ca).
AMINN refers to autoencoder-based multiple instance neural network that is deveopled to address multifocality, i.e. to incorporate features from all lesions for prediction/classification:
- We jointly train an autoencoder to reconstruct input features and a multiple instance network to make predictionsby aggregating information from all tumour lesions of a patient
- We incorporate a two-step normalization technique to improve the training of deep neural networks, built on the observation that the distributions of radiomic features are almost always severely skewed.
- Experimental results empirically validated our hypothesis that incorporating imaging features of all lesions improves outcome prediction for multifocal cancer.
- Our code is written in Keras, using Tensorflow as backend. If using radiomic features as input, one run of training takes few minutes on a moderate GPU or CPU.
The paper has been early accepted by MICCAI 2021. For more details, please refer to our paper. Link
main.py: Trains a AMINN model. The performance of the model along with hyperparameters used to train the model are saved in a output csv file.
myargs.py: Defines arguments used for training the model, including model related parameters, training parameters and system parameters.
network.py: Contains code for training and testing. Defines the AMINN model.
pooling_method.py and layer.py: Defines max, log-sum-exponential (lse), average pooling and attention-based pooling, respectively.
dataset.py and utils.py: Contains scripts that are used for creating datasets for AMINN from csv files (radimoics features + clinical variables).
data/input.csv: This file is provided as an example of input data. We have modified the value of the radiomic features and survival outcomes so the file cannot be used for training (It will throw a 'Divided by zero' error).
If you find AMINN useful in your research, please consider citing:
@article{chen2020aminn,
title={AMINN: Autoencoder-based Multiple Instance Neural Network for Outcome Prediction of Multifocal Liver Metastases},
author={Chen, Jianan and Cheung, Helen and Milot, Laurent and Martel, Anne L},
journal={arXiv preprint arXiv:2012.06875},
year={2020}
}
If you have any questions about this code, I am happy to answer your issues or emails (to chenjn2010@gmail.com).
python main.py --pooling=ave --epoch=100 --runs=10 --folds=3 --lr=1e-4 --recon_coef=1.0 --fp_coef=1.0 The work conducted by Jianan Chen was funded by Grants from the Martel lab.
I would also like to thank the following repositories for supporting and inspiring this work.