APS: A deep learning framework for MR-to-CT image synthesis

This repository is the official implementation of An Improved Deep Learning Framework for MR-to-CT Image Synthesis with a New Hybrid Objective Function.

Getting started

• Download or clone this repo to your computer.
• Run pip install -r requirements.txt to install the required Python packages.
• The code was developed and tested on Python 3.6.13 and Ubuntu 16.04.
• Note that this codebase runs on PyTorch Lightning library.

Training and testing the APS framework on your own dataset

Steps

1. Create a custom Dataset class for your own data. Follow the official PyTorch tutorial if you are unfamiliar with this step.
2. Integrate your custom Dataset class into main_train.get_training_dataloaders().
3. Run python main_train.py --gpus 1 to start training and testing the APS model on 1 GPU. Execute python main_train.py --help for the available input arguments.

Notes

• Ensure that the data are normalized using min-max formula and resized to the expected input image size (default is 288x288). The expected input image size is set via --image_size input argument of main_train.py.
• The __getitem__() of your custom Dataset class has to return a dict with the following keys:
  • ct: the CT image tensor.
  • in_phase: the MR in-phase image tensor.
  • ct_min: the smallest CT's HU value before the min-max normalization. This is needed for metric computations.
  • ct_max. the biggest CT's HU value before the min-max normalization. This is needed for metric computations.

Visual results

Citation

If you use this code for your research, please cite our paper.

@inproceedings{ang2022,
  title={An improved deep learning framework for MR-to-CT image synthesis with a new hybrid objective function},
  author={Ang, Sui Paul and Phung, Son Lam and Field, Matthew and Schira, Mark Matthias},
  booktitle={Proceedings of the IEEE International Symposium on Biomedical Imaging},
  year={2022}
}

Acknowledgement

Some parts of our code are inspired by F-LSeSim and CycleGAN.