Implementation of the SDNet model to perform disentanglement of anatomical and modality information in medical images. For further details please see our paper, accepted in Medical Image Analysis.
The structure of this project is the following:
- configuration: package containing configuration parameters for running an experiment.
- layers: package with custom Keras layers
- loaders: package with data loaders
- models: package with the SDNet model and other Keras models
- model_executors: package with scripts for running an experiment
- callbacks: package with Keras callbacks for printing images and losses during training
To define a new data loader, extend class base_loader.Loader, and register the loader in loader_factory.py. The datapath is specified in parameters.py.
To run an experiment, execute experiment.py, passing the configuration filename and the split number as runtime parameters:
python experiment.py --config myconfiguration --split 0
The code is written in Keras version 2.1.6 with tensorflow 1.4.0 and experiments were run with a Titan-X GPU.
A tensorflow implementation is uploaded in https://github.com/GabrieleValvano/SDNet.
If you use this code for your research, please cite our paper:
@article{CHARTSIAS2019101535,
title = "Disentangled representation learning in cardiac image analysis",
journal = "Medical Image Analysis",
volume = "58",
pages = "101535",
year = "2019",
issn = "1361-8415",
doi = "https://doi.org/10.1016/j.media.2019.101535",
url = "http://www.sciencedirect.com/science/article/pii/S1361841519300684",
author = "Agisilaos Chartsias and Thomas Joyce and Giorgos Papanastasiou and Scott Semple and Michelle Williams and David E. Newby and Rohan Dharmakumar and Sotirios A. Tsaftaris",
keywords = "Disentangled representation learning, Cardiac magnetic resonance imaging, Semi-supervised segmentation, Multitask learning"
}