Self-Contained Boosting of 4D Cone-Beam CT

This repository contains the official code for the following Medical Physics paper:

@article{Madesta2020,
  author = {Frederic Madesta and Thilo Sentker and Tobias Gauer and Ren{\'{e}} Werner},
  title = {Self-Contained Deep Learning-Based Boosting of 4D Cone-Beam {CT} Reconstruction},
  journal = {Medical Physics},
  year = {2020},
  month = Aug,
  publisher = {Wiley},
  doi = {10.1002/mp.14441},
  url = {https://doi.org/10.1002/mp.14441},
}

Installation

Build Docker

Internally, this project uses RTK for reconstructing CBCT images. As compiling CUDA-related stuff can be cumbersome, we provide a Docker image with batteries included. In order to build this image you need to perform the following steps:

0. You need a machine with a NVIDIA GPU and installed NVIDIA drivers
1. Install Docker and NVIDIA Container Toolkit (if not already installed)
2. Build the Docker image by executing cd docker && ./build_docker.sh. This will take some time.

Phantom data set

We privide a 4D CBCT phantom data set for test purposes.

Details:

• 4D CBCT Scanner: Varian TrueBeam
• Phantom: Dynamic Thorax Phantom: Model 008A
• The following scans are included:
  1. SI amplitude of insert: ±10mm, pattern: sin, period: 5.0s
  2. SI amplitude of insert: ±10mm, pattern: cos**4, period: 5.0s
  3. SI amplitude of insert: ±10mm, pattern: sin, period: 2.5s
  4. SI amplitude of insert: ±10mm, pattern: cos**4, period: 2.5s
  5. SI amplitude of insert: ±10mm, pattern: sin, period: 7.5s
  6. SI amplitude of insert: ±10mm, pattern: cos**4, period: 7.5s

Usage

The following scripts are included inside the scripts folder:

• prepare_varian.py:
  This script will prepare Varian TrueBeam 4D CBCT raw data, i.e.

  • extract needed files
  • convert projection files to single projection stack
  • air normalize projection stack
  • create RTK geometry
  • extract respiratory curve (phase and amplitude) from the projections

• reconstruct.py:
  This script will reconstruct 4D CBCT raw data using RTK. Especially, it can handle the Varian TrueBeam 4D CBCT data extracted by the previous script (prepare_varian.py). Of course any raw data can be feeded into this reconstruction pipeline as long as it is in the right RTK format.

• train.py:
  This script will train the 4D CBCT boosting network on the reconstructed data (you can use the provided phantom data set for test purposes). In the end, the trained

• model is applied to the 4D CBCT phase images.