AutoencodersForUnsupervisedAnomalyDetection

The code behind my paper "Deep Autoeoncoding Models for Unsupervised Anomaly Detection in Brain MR Images"

This is the source code for my paper Deep Autoeoncoding Models for Unsupervised Anomaly Detection in Brain MR Images (Baur et al. https://arxiv.org/abs/1804.04488), accepted at the MICCAI 2018 BrainLesion Workshop, presented as an oral and as a poster.

The source code comprises my object-oriented Deep-Learning framework, developed on top of TensorFlow, as well as multiple files for training & evaluating the various Auto-Encoders and Generative Adversarial Networks described in the paper:

• DLMODEL.py: the base class for all your TensorFlow Deep Learning needs
• AE.py: inherits from DLMODEL, implements training & inference of Autoencoders
• AEGAN.py: inherits from DLMODEL, implements training & inference of Autoencoders + Adversarial Network
• VAE.py: inherits from DLMODEL, implements training & inference of Variational Autoencoders
• VAEGAN.py: inherits from DLMODEL, implements training & inference of a VAE-GAN
• Losses.py, custom_layers.py, utils.py: Utilities...
• architectures/*: This folder contains network architectures, which can be plugged in any of the above objects for training
• main_*.py: The files that will launch a training and evaluation of a specific model
• Experiment.py: THe file that describes the evaluation pipeline.

The dataset used for this work was a non-public, clinical brain MR dataset with both healthy subjects, and subjects with MS, and is therefore not provided. You are free to plug in your own dataset though, by simply taking a look at the class "Dataset.py". Of great importance is the implementation of the "next_batch(...)" method.

Training & Evaluation

1. pip install -r requirements.txt
2. python main_*.py

Cite Us

If you use some of the code in your work, please cite our paper:

@article{baur2018deep,
  title={Deep Autoencoding Models for Unsupervised Anomaly Segmentation in Brain MR Images},
  author={Baur, Christoph and Wiestler, Benedikt and Albarqouni, Shadi and Navab, Nassir},
  journal={arXiv preprint arXiv:1804.04488},
  year={2018}
}