Implicit Neural Representations for Robust Joint Sparse-View CT Reconstruction (TMLR)

This is the official repository for the paper Implicit Neural Representations for Robust Joint Sparse-View CT Reconstruction.

Implicit Neural Representations for Robust Joint Sparse-View CT Reconstruction,
Jiayang Shi, Junyi Zhu, Daniel M. Pelt, K. Joost Batenburg, Matthew B. Blaschko TMLR 2024

We introduce a novel Bayesian framework for joint reconstruction of multiple objects from sparse-view CT scans using Implicit Neural Representations (INRs) to improve reconstruction quality. By capturing shared patterns across multiple objects with latent variables, our method enhances the reconstruction of each object, increases robustness to overfitting, and accelerates the learning process.

Node 1	Node 5	Node 9	Prior Mean

Setup

Create and activate a Conda environment:

conda env create -f environment.yml
conda activate inr4ct

Dataset

We provide 10 slices LungCT images in imgs.tif from LungCT for easy testing. You can also use your own data by loading it accordingly.

Run

To run the joint reconstruction code, use:

python main.py

We also provide implementations of several comparison methods for joint reconstruction:

• Meta-Learning (MAML)
• Federated Averaging (FedAvg)
• INR-in-the-Wild (INRWild)
• A single reconstruction method (SingleINR)

You can run these methods using:

python meta.py
python fedavg.py
python single_inr.py
python inr_wild.py

Hyperparameters

You can tune the hyperparameters to achieve optimal reconstruction performance. The β parameter controls the regularization strength in joint reconstruction. For noisy data, β should generally be set higher than in noiseless cases. For example, in the noisy experiment on the walnut dataset described in the paper, we used β = 1e-7.

For best results, you may also want to tune other parameters such as the number of assembly iterations, individual reconstruction iterations, and related settings.

Citation

If you find our paper useful, please cite

@article{shi2024implicit,
  title={Implicit Neural Representations for Robust Joint Sparse-View CT Reconstruction},
  author={Shi, Jiayang and Zhu, Junyi and Pelt, Daniel M and Batenburg, K Joost and Blaschko, Matthew B},
  journal={arXiv preprint arXiv:2405.02509},
  year={2024}
}