This is the implementation of the Contrastive Diffusion Model with Auxiliary Guidance for Coarse-to-Fine PET Reconstruction (early accpeted by MICCAI 2023🔥🔥🔥), which is the first work that applies diffusion model to PET Reconstruction.
We have shared our pre-processed datasets and final checkpoints. You can download it from the link below.
Our proposed framework has two modules, i.e., a coarse prediction module (CPM) and an iterative refinement module (IRM). The CPM predicts a coarse-denoised PET image from the LPET image(use unet to make a coarse prediction), while the IRM predicts the residual between the coarse prediction and the SPET image iteratively(use diffusion model to predict residual). To improve the correspondence between the LPET image and the RPET image, we adopt an auxiliary guidance strategy (b) at the input level(Add to both CPM and IRM unet) and a contrastive diffusion strategy (c) at the output level(In the loss of training).
The environment can be set up following the instructions below.
conda create --name diffpet python=3.8
conda activate diffpet
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
conda install tensorboardX wandb numpy tqdm SimpleITK opencv-python medpy
pip install scipy==1.5.4
git clone https://github.com/wyhlovecpp/PET-Reconstruction.git
cd PET-Reconstruction
The code structure is as follows.
├── assets
│ └── model.svg # model structure
├── config
│ └── sr_sr3_16_128.json # config file for training and inference
├── core
│ ├── logger.py # logger
│ ├── metrics.py # function for evaluation
│ └── wandb_logger.py # wandb logger
│
├── datasets
│ ├── __init__.py # dataloader
│ ├── LRHR_dataset.py # dataset(Support .mat format data reading, you can replace with your own)
│ └── util.py # dataset utils
├── model
│ ├── sr3_modules # main_model
│ │ ├── diffusion.py # diffusion model
│ │ ├── unet.py # unet structure
│ ├── __init__.py # init
│ ├── base_model.py # function for model
│ ├── model.py # train and inference
│ └── networks.py # Build CPM and IRM networks for model.py
├── inference.py # inference
└── train.py # train
We conducted most of our low-dose brain PET image reconstruction experiments on a public brain dataset, which is obtained from the Ultra-low Dose PET Imaging Challenge 2022. Each subject has a resolution of 128 × 128 × 128, and 2D slices along the z-coordinate were used for training and evaluation. After you've prepared your 2D-slice datasets, you need to change the datasets config to your data path.
"datasets": {
"train": {
"name": "FFHQ",
"mode": "HR", // whether need LR img
"dataroot": "C:\\Users\\Administrator\\Desktop\\deblur\\dataset\\train_mat",
}
To train the model, you can run the following command.
python train.py -opt config/sr_sr3_16_128.json
To inference the model, you can run the following command.
python inference.py -opt config/sr_sr3_16_128.json
We would like to thank the authors of previous related projects for generously sharing their code and insights: