Skip to content

Latest commit

 

History

History
46 lines (39 loc) · 1.75 KB

README.md

File metadata and controls

46 lines (39 loc) · 1.75 KB

CRET: Continuous Representation-based Reconstruction for Computed Tomography

This is the official implement of the paper "CRET: Continuous Representation-based Reconstruction for Computed Tomography"

Environment

  • python3
  • pytorch 2.1.2+cuda11.8
  • matplotlib, yaml, tqdm, timm, etc..

Dataset

  • Download AAPM-Mayo dataset at here
  • Our CRET requires sinogram data. To generate sinogram data, run forward projection as:
python sinogen.py
  • After generating sinogram and ground truth image as '.npy' extension, put these datasets as folows:
/Dataset
├── L067
│     ├── img
│     ├── sinogram
├── L096
│     ├── img
│     ├── sinogram
...

For training

training the reconstruction module corresponding to the step-1 & encoder as RDN (use 2 GPUs)

python main.py --config train/sino/train_rdn-cret --gpu 0,1 --save rdn_cret

training the restoration module corresponding to the step-2 (use 2 GPUs)

python main.py --config train/dual/train_rdn-cret+ --gpu 0,1 --save rdn_cret+

Checkpoints

  • Pre-trained modules are can be downloaded in here (rdn_cret.pt & rdn_cret+.pt)
  • Note that mask_info.pt is information about the sinogram mask for sinogram squeezing and should be placed at "data/mask/"

Demo

  • We provide a demo trial of our proposed CRET in demo.ipynb

Acknowledgements

Our code is built mainly with reference to EDSR and OPE-SR. We appreciate them for their invaluable contributions to the development community.