In the U.S., more than 80 million CT scans are performed annually. A CT (computer tomography) scan is a common, effective way for doctors to peer inside the human body to diagnose, treat and monitor disease. However, a CT scan doesn’t tell the whole story. While it reveals tissue morphology, it does not provide any information about the elemental composition of tissue. One workaround is to use contrast agents like iodine, but even that methods has pitfalls—contrast-enhanced structures may have similar density to bone or calcified plaques, making them difficult to distinguish.
Thus, doctors have turned to dual-energy CT, which gathers two datasets to produce images that reveal both tissue shape and information about tissue composition. But dual-energy CT is expensive, high-tech and often requires a higher dose of radiation than conventional CT scans.
So, engineers at Rensselaer Polytechnic Institute turned to deep learning to help.
“With traditional CT, you take a grayscale image, but with dual-energy CT you take an image with two colors," explained Ge Wang, professor of biomedical engineering at Rensselaer Polytechnic Institute and author of a new paper on the subject. "With deep learning, we try to use the standard machine to do the job of dual-energy CT imaging."
Using dual-energy CT-derived virtual monoenergetic (VM) images, Wang and team taught the ResNet deep learning network to map single-spectrum CT images to virtual monoenergetic images at pre-specified energy levels.
According to their results, described in the journal Patterns, the trained neural network delivered high-quality approximations of dual-energy CT-derived VM images with a relative error less than 2% for the testing dataset. Additionally, structural information, especially texture features, were well preserved by the machine learning method.
The researchers then used the learned VM images to generate multi-material decomposition (MMD) images—hoping for close approximations to clinical images directly produced by dual-energy. Again, the results demonstrated high-quality material-specific images. The study points to a specific example in which a bone image was clearly separated from the reconstructed VM image, shedding light on a calcification in the abdominal aorta that was not clearly visible through conventional CT scans. Ultimately, the method enables multi-material decomposition into three tissue classes, with accuracy comparable with dual-energy CT.
“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," said Wang.
The researchers say their deep learning method is well-suited for computing proton stopping power for proton therapy planning, as well as photon-counting micro-CT for in vivo preclinical applications. For preclinical imaging, the neural network could significantly reduce scanning time and radiation dose by learning from multiple photo-counting micro-CT datasets and then reconstructing those to produce dual-energy CT images.