Awesome

3D Shape Consistent GAN

PyTorch implementation of "Shape-consistent Generative Adversarial Networks for multi-modal Medical segmentation maps". Paper: https://arxiv.org/abs/2201.09693

Prerequisites

• Linux or macOS
• Python 3
• CPU or NVIDIA GPU + CUDA CuDNN

Getting Started

Installation

• Clone this repo:

git clone https://github.com/orhir/3D-Shape-Consistent-GAN
cd 3D-Shape-Consistent-GAN

• Install PyTorch and 0.4+ and other dependencies (e.g., torchvision, visdom and dominate).

  • For Conda users, you can create a new Conda environment using conda env create -f package-list.txt.

• Dataset link: https://drive.google.com/drive/folders/1F-qmV1t7X33i0ucPYPLXWin2RZJcoPdO?usp=sharing

Augmentation

• Create spatial augmentation

  python createAug.py <PATH_TO_DATASET> train <NUM_ITERS> <OUTPUT_FOLDER_NAME
  

Train

• Train a model:

  • Phase 1:

    python train.py --dataroot <PATH_TO_DATASET> --model cycle_gan --crop_size_z 32 --crop_size 256 --only_seg --max_dataset_size 200 --name phase_1 --train_phase 1 [--four_labels]
    

  • Phase 2:

    python train.py --dataroot <PATH_TO_DATASET> --model cycle_gan --crop_size_z 32 --crop_size 256 --load_seg --load_name phase_1 --max_dataset_size 200 --name phase_2 --train_phase 2 [--four_labels]
    

  • Phase 3:

    python train.py --dataroot <PATH_TO_DATASET> --model cycle_gan --crop_size_z 32 --crop_size 256 --load_all_networks --load_name phase_2 --max_dataset_size 200 --name phase_2 --lambda_seg_from_syn 0.5 --train_phase 3 [--four_labels]
    

• To see more intermediate results, check out ./checkpoints/MODEL_NAME/web/index.html.

• To view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097.

Test

python test.py --dataroot test_dataset_path/ --model cycle_gan --load_name phase_3 --crop_size 128 --crop_size_z 64 [--four_labels]

• The test results will be saved to a html file here: ./results/[load_name]/latest_test/index.html.

Acknowledgement

Part of the code is revised from the PyTorch implementation of CycleGAN.

Note

• The repository is being updated