Please select the desired project time frame:

• July 2026
• January 2026
• July 2025
• January 2025
• July 2024
• January 2024
• July 2023
• January 2023
• July 2022
• January 2022
• July 2021
• January 2021
• July 2020
• January 2020
• July 2019
• January 2019
• July 2018
• January 2018
• July 2017
• January 2017
• July 2016
• January 2016
• July 2015
• January 2015
• July 2014
• January 2014
• July 2013
• January 2013
• July 2012
• January 2012
• July 2011
• January 2011
• July 2010
• January 2010
• July 2009
• January 2009
• July 2008
• January 2008
• July 2007
• January 2007
• July 2006
• January 2006
• July 2005
• January 2005
• July 2004
• January 2004
• July 2003
• January 2003
• July 2002
• January 2002
• July 2001
• January 2001

Start of funding 01.07.2019

Enhancing the Accuracy of Material Decomposition

---

In recent years, energy-resolved x-ray computed tomography (CT) imaging has shown clinical value thanks to its ability to differentiate materials in the depicted volume. This is possible, because elemental materials show different energy-dependent attenuation behaviors. Exploiting this phenomenon, we can compute the material composition of volumes from multiple images, acquired at different x-ray energies. These spectral images can directly be measured by modern CT systems through either energy-resolving detectors or multi-spectrum usage.

Traditional analytical approaches to material decomposition require accurate calibration of the imaging system and are therefore prone to deviations from the idealized system model, producing high levels of noise and other errors. We expect the use of learning algorithms to surpass these model-based approaches in quality and generalizability. The application of deep learning models to this problem has already shown promising results on simulated data and will now be implemented on measured data.

Final report:
As part of the funded project, the applicability of deep learning to material decomposition of dual-energy data was analyzed. With the ultimate goal of being able to apply the learned material decomposition to real scans, the simulation was first adapted to the CT, then a model was trained on the resulting simulation data and finally the results were analyzed.

In order to facilitate the application of the resulting models to real data, the CONRAD simulation environment was adapted in terms of trajectory, spectrum and noise behavior and suitable phantoms were modeled. The training data must map the value range of the input data as evenly as possible. Therefore, different constellations of path lands and materials were modeled and raw data generated from them.

A neural network of the U-Net architecture was chosen to break down the “dual-energy” projections into water and iodine length distributions. Due to its structure, this has proven to be successful in segmentation tasks in the literature. After successfully training the model on the generated data, the results were qualitatively examined. The results were promising, with a correlation between the quality of the decomposition and the occurrence of the corresponding lengths in the training data.

In order to enable generalization to measurement data, a geometric phantom was scanned with a CT device and a “ground truth” mask was created by subtracting successive scans after carefully removing individual pieces of material. These data can be used by future work in order to adapt the network that has been trained on simulation data to real data.