by John R. Fischer
, Senior Reporter | October 22, 2020
A new deep learning algorithm could help conventional CT scanners generate images that match the quality of those produced with dual-energy CT at less expense.
While traditional CT scans show the shape of tissues, they are unable to provide informative details about the composition of tissues, even when using contrast agents such as iodine. To solve this problem, engineers at Rensselaer Polytechnic Institute trained a neural network to produce more complex images using single spectrum CT data.
“With such spectral information, we can do a better job for tissue characterization and material decomposition, for which more expensive CT scanners are intended, including dual-energy CT and photon-counting CT scanners,” Ge Wang, Ph.D., director of the biomedical imaging center at Rensselaer Polytechnic Institutes, told HCB News.
Numed, a well established company in business since 1975 provides a wide range of service options including time & material service, PM only contracts, full service contracts, labor only contracts & system relocation. Call 800 96 Numed for more info.
Conventional CT images are produced in grayscale, suppressing all the spectral information of the image. In contrast, dual-energy CT gathers two data sets to produce images that show both tissue shape and information on its composition. This technology, however, requires a higher dose of radiation and is more expensive due to the additional hardware needed.
The network was developed by Wenxiang Cong, a research scientist at Rensselaer, to computationally map a color-blinded grayscale image into respective red, green and blue components that share details of tissue composition. He and Wang trained the network to do so using images produced by dual-energy CT, alongside co-authors from Shanghai First-Imaging Tech and researchers at GE Research. They found it was able to produce high-quality approximations with a relative error of less than 2%.
Wang says that dual-energy CT and photon-counting CT scanners “can provide ground truth data to train Cong's network so that the network can extract spectral information from data collected with a relatively cheaper CT scanner that just produces energy-integrating data. Further evaluation and optimization is certainly needed in a task-specific fashion so that our prototyped network can be eventually translated into clinical practice.”
Wang and Cong are collaborating with radiologist Dr. Mannudeep K. Kalra and his team at Massachusetts General Hospital for further testing of their approach.
The findings were published in Patterns