http://arxiv.org/abs/2403.16361

Author: Ziheng Deng, Jun Zhao, School of BME, Shanghai Jiao Tong University

This repository is the official implementation of RSTAR4D. The main contributions of the paper are:

• We identified the rotational streak artifacts (RSA) in 4D CBCT image.
• We proposed the RSTAR-4DNet to effectively reduce the RSA in the spatiotemporal domain.
• We effectively trained a 4D CNN with limited computational resources and only dozens of 4D training samples.

This work has been accepted by IEEE Transactions on Radiation and Plasma Medical Sciences (IEEE TRPMS, early access).

Codes and pre-trained model are available now. Please check the /githubcode for the codes. Example data can be download from: example_data

Let me know if there is any bug.

1. According to the respiratory signal, the projection data from a full-circular scan are sorted into 10 respiratory phases.

   

2. Within each phase, there are limited projection angles for CT reconstruction -> sparse-view reconstruction -> streak artifacts.

   

3. The distribution of streak artifacts is closely related to the projection sampling pattern.

   

4. The respiratory motion is quasi-periodic and sequential -> rotational projection sampling map.

3+4 -> RSA

Projection Sampling Map	Dynamic CBCT image (with RSA)

Here are some results of our RSTAR-Net tested on real clinical CBCT scans.

Case	Uncorrected Image	Corrected Image
Case1
Case2
Case3