search

thomasneff / Generative-Adversarial-Network-based-Synthesis-for-Supervised-Medical-Image-Segmentation

Code for the paper 'Generative Adversarial Network based Synthesis for Supervised Medical Image Segmentation'
MIT License
18 stars 13 forks source link

readme

Generative-Adversarial-Network-based-Synthesis-for-Supervised-Medical-Image-Segmentation

Code for the paper 'Generative Adversarial Network based Synthesis for Supervised Medical Image Segmentation'

This modification adds the ability to generate pixel-wise segmentations to the GAN. Currently, it assumes that the images are grayscale, therefore the GAN model only handles 2 image channels (1 for the image, 1 for the segmentation)

For more details, check out our paper.

For citations (bibtex):

Neff, Thomas and Payer, Christian and Stern, Darko and Urschler, Martin (2017). 
Generative Adversarial Network based Synthesis for Supervised Medical Image Segmentation.
In Proceedings of the OAGM&ARW Joint Workshop, pp. 140-145.

Credits

Credits to sugyan for their tf-dcgan implementation, which this code is based on.

How to use/modify

In train.py, the 'idlist_*' variables defined at the start of 'train' need to point to text files containing a list of image/segmentation filenames, one for each line.

The '*_folder_path* variables need to point to the folder containing the files defined in the idlists.

(You can easily change the data loading, the dcgan model just takes image batches as a tensor as input.)