sh4174 / 3DStyleGAN

3D StyleGAN2 for Medical Images

3D StyleGAN for Three Dimensional Medical Images

3D StyleGAN for the generative modeling of full 3D medical images.

Please see the official repo (https://github.com/NVlabs/stylegan2) of StyleGAN2 from NVIDIA for the original code and its license.

Note

We plan to clean and update the code repository on October when we present this work at the DGM4MICCAI workshop at MICCAI 2021.

Requirements

The requirements of the original code + (TF 1.14 --> TF 2.4.0), Python 3.8, nibabel

Conda environment file is included in the conda_en directory.

Create TFRecord with 3D Medical Images (NIFTI or NPZ)

python dataset_tool.py create_from_images3d [TFRecord_Folder/TFRecord_Name] [NIFTI Data Folder] --shuffle 1 --base_size 5 6 7

base_size: The size of the base layer of a generator. e.g. [ 4, 4, 4] for 2^x images; [5,6,7] for 160x192x224 (5x32, 6x32, 7x32) or 80x96x112 (5x16, 6x16, 7x16) images (shown in the paper).

Train 3D StyleGAN

python run_training.py --num-gpus=4 --data-dir=[TF_Record_Folder] --config=[Training_Config] --dataset=[TFRecord_Name] --total-kimg=6000

[Training_Config] needs to be filled by the name of prefixed configuration in run_training.py

The hyperparameters can be changed in run_training.py with a configuration name.

Image Generation

python run_generator.py generate-images --network=[Trained_Network_Path] --seeds=66,230,389,1518,1020,11,1104,1120,1031 --truncation-psi=0.0

Style Mixing

python run_generator.py style-mixing-example --network=../trained_networks/2mm_f96.pkl --row-seeds=3181 --col-seeds=1104,1120 --truncation-psi=0.0 --col-styles=6-9

Please contact Sungmin Hong (HMS/MGH, [e-mail]) and Razvan Marinescu (MIT, [e-mail]) if you have any questions.