DatasetGAN2-ADA

Code supporting the paper Enhanced deep-style interpreter for automatic synthesis of annotated medical images (link)

1. How does it work?

2. Requirements

• Python 3.8 is supported.

• Pytorch >= 1.4.0.

• The code is tested with CUDA 10.1 toolkit with Pytorch==1.4.0 and CUDA 11.4 with Pytorch==1.10.0.

• All results in our paper are based on NVIDIA Tesla V100 GPUs with 32GB memory.

• Set up python environment:

  virtualenv env
  source env/bin/activate
  pip install -r requirements.txt
  

• Add the project to PYTHONPATH:

  export PYTHONPATH=$PWD
  

3. Creating execution environment

Detailed instructions will be added soon.

4. Download datasets

For convinience, we provide the preprocessed ATLAS 2.0 image inferior-superior layers used in the experiments, the projected images and latents, and the synthetic images generated by DatasetGAN2-ADA.

Create a folder datasets/ inside the project and extract the following files inside the folders: Folder Name | Link | Description | |———|————–|—————-| |ATLAS_2_0_processed|link| ATLAS 2.0 preprocessed layers (train folder) and synthetic images generated by DatasetGAN2-ADA (test folder) without filtering low US score and anomalies.| |synthetic_ATLAS_2_0_layers|link|Metadata, masks and additional info generated by DatasetGAN that are not inside ATLAS_2_0_processed/class_80_100/test.| |result_images| link |Result images obtained at the end of the experiment. Refer to Tab 4 for the nomenclature.|

After downloading and adding all the folders, run the following commands to populate synthetic_ATLAS_2_0_layers:

cp ./datasets/ATLAS_2_0_processed/class_80_100/test/*.tiff ./datasets/synthetic_ATLAS_2_0_layers/synthetic_with_metadata/

5. Reproduce the experiments

5.1. Only the results and plots with the provided preprocessed and synthesized images

1 - Download the image datasets following the instructions above

2 - Execute the jupyter notebook Article_experiments.ipynb

5.2. Run all the experiments

1 - Train StyleGAN2-ADA with processed_ATLAS_2_0_layers images. Be aware that rerunning the training of StyleGAN2-ADA might produce different results. We share the StyleGAN2-ADA checkpoint for convenience.

2 - Generate projections of a few training images. We share the image projections along with their latentes and original masks for convenience: PENDING

3 - Select a few projections and use them along with original masks to Train DatasetGAN2-ADA’s interpreter. We share the DatasetGAN2-ADA checkpoint for convenience:

4 - Generate Synthetic images and masks with DatasetGAN2-ADA.

5 - Execute the jupyter notebook Article_experiments.ipynb

6. Run experiments with your own dataset

You can use DatasetGAN2-ADA on your own dataset by reproducing the item 5.2. on your dataset. DatasetGAN2-ADA was originally prepared for grayscale MRI. Thus, a few minor adjustments should be done to make it compatible with RGB images.

Citations

Please use the following citation if you use our data or code:

@article{PachecoDosSantosLimaJunior2025,
  author = {Pacheco dos Santos Lima Junior, M. S. and Ortiz-de-Lazcano-Lobato, J. M. and Fernández-Rodríguez, J. D. and López-Rubio, E.},
  title = {Enhanced deep-style interpreter for automatic synthesis of annotated medical images},
  journal = {Neural Computing and Applications},
  year = {2025},
  doi = {10.1007/s00521-025-11516-8},
  url = {https://doi.org/10.1007/s00521-025-11516-8}
}