Computed tomography (CT) is a popular medical imaging tool, visualizing and quantifying internal structures for screening, diagnosis, therapy planning and treatment monitoring. Conventional clinical CT scans generate a spectrally integrated attenuation image that shows tissue morphology, but does not directly provide any information regarding tissue composition.
Dual-energy CT (DECT) systems, which acquire two spectrally distinct datasets, can reconstruct virtual monoenergetic (VM) and material-specific images that provide information about tissue composition. Compared with conventional CT, however, DECT is more expensive and complex, and often requires an increased radiation dose.
For patients without access to DECT scans, a way to approximate the information provided by DECT using a single-spectrum CT scan could improve clinical diagnoses. With this aim, Wenxiang Cong and a team led by Ge Wang at Rensselaer Polytechnic Institute demonstrated a deep-learning approach that can produce VM images from single-spectrum CT scans.
“With traditional CT, you take a greyscale image, but with dual-energy CT you take an image with two colours,” explains Wang. “With deep learning, we try to use the standard CT machine to do the job of dual-energy CT imaging. We hope that this technique will help extract more information from a regular single-spectrum X-ray CT scan, make it more quantitative and improve diagnosis.”
Wang and his team – in collaboration with Shanghai First-Imaging Tech and GE Research – modified a residual neural network (ResNet) to transform clinical single-spectrum CT images to VM counterparts. After training the network on clinical DECT images, they used it to produce VM images at two energy levels (80 and 110 keV) from single-spectrum images. These VM images showed excellent agreement with those produced by DECT reconstruction using two different spectra.
The model was able to produce high-quality approximations to linear attenuation coefficients with a relative error of less than 2%. The structural similarity between the two types of VM images was up to 0.99, showing that structural information, especially texture features, was well preserved by the machine learning method.
Next, the team used the deep-learning-based VM images to generate material-specific images of three tissue types: adipose, muscle and bone. The resulting images were of high quality and closely approximated those produced by DECT using projection datasets from two X-ray spectra.
The researchers note that the bone image could be clearly separated from the VM images. Notably, a calcification in the abdominal aorta that was inconspicuous in the original polychromatic image was visible in the synthesized VM image and in the bone image. This reveals one of the potential clinical utilities of the DL method.
Having demonstrated that a conventional CT dataset coupled with deep learning can deliver a close approximation of DECT images, the researchers suggest that it is potentially feasible to use conventional CT to perform some important tasks currently achieved using DECT – thereby eliminating the hardware cost associated with a DECT scanner.
One example application is the determination of proton stopping power for use in proton therapy planning, where it is important to accurately represent material concentrations of common tissues along the therapeutic beam. The team also highlight deep-learning-based VM imaging as an alternative to photon-counting micro-CT for in vivo preclinical applications.
The full details of the research are published in Patterns.