Unsupervised Domain Adaptation for Low-dose CT Reconstruction via Bayesian Uncertainty Alignment

This is the official PyTorch implementation of our paper, coined [Unsupervised Domain Adaptation for Low-dose CT Reconstruction via Bayesian Uncertainty Alignment] (https://ieeexplore.ieee.org/document/10593806).

If you have any problems, please email me ([email protected]).

Usage

Environment

please see the requirement file

Dataset

 - Source Domain: 2016 NIH AAPM-Mayo Clinic Low-dose CT Grand Challenge Dataset (25% of normal dose)
 - Target Domain: 
      - Simulated 5% dose of source domain
      - Simulated 50% dose of source domain
      - ISICDM 2020 Challenge Dataset (<25% dose)

Due to the restricted license of AAPM-Moyo dataset, we CANNOT directly share the above datasets with you. Two options for you:

Get the datasets from our: - You can get the license for the AAPM-Moyo dataset by following the instructions at AAPM Moyo challenge website . If you have get the access right from AAPM, you can email us ([email protected]) to provide the given proof. Then, we will share with you a download link including all the above datasets.

If you get the dataset,the dataset saving structure will be

 - /your_path/
 
    └── AAPM
           ├── 1200000_1mm(50%)/
 
           ├── 110000_1mm (5%)/
 
           ├── full_1mm/
           
           ├── quarter_1mm (25%)/
      
    └── ISICDM

Simulate by yourself: We provide a simulation file that you can use to process your dataset.

Pretrained Source Domain Model and VGG19 Model

1. We provide a source-trained model in 'pretrained_model/' folder. This model is trained on AAPM-Moyo 2016 dataset. You can also use your own pre-trained model. Please set the correct model path at init.py.

2. We also provide the pretrained VGG19 model for perception loss computation at here. Please download the pretrained model and put it to the 'pretrained_model/' folder

Training

1. All hyperparameters are in init.py. Some important settings include

• "self.target": target domain, including 'AAPM_5', 'AAPM-50', and 'ISICDM'
• "self.baseline_type:" choose different models for training. Our model is named "ours_new" (corresponds to DA_Denoiser in model.py). You can choose other baseline methods but need to set the self.baseline = True
• "self.data_root_path": path for dataset
• "self.root_path": path for training results

2. Run train.py

LDCT image denoising/reconstruction benchmarking methods for unsupervised domain problems

We provide wide LDCT image denoising/reconstruction benchmarking methods for unsupervised domain problems. You can choose different model names at self.baseline_type:

• ClycleGAN: self.baseline_type = 'clycle'
• Noise2noise: self.baseline_type = 'n2n'
• CCDnet: self.baseline_type = 'RMF'
• UDA: self.baseline_type = 'uda'
• WGAN-VGG: self.baseline_type = 'WGAN'