The COVID-19 pandemic has killed more than 160,000 people in the U.S. alone and infected more than 20 million around the world. Until there's a vaccine, one of the few strategies for slowing the disease’s lethal race through communities is identifying and isolating people who are already infected. That takes testing that’s widely available, fast, and accurate. The U.S. is currently lagging on all three counts.

Wu Feng, a professor of computer science in the College of Engineering, and Guohua Cao, a visiting scientist in computer science who also has expertise in biomedical engineering, have landed a grant from the National Science Foundation for technology that could help fill in some of those gaps. The researchers are developing deep-learning-based algorithms that can boost the resolution of CT scans — an approach that searches for SARS-CoV-2, the virus that causes COVID-19, by looking not for the virus itself but for the telltale damage it leaves behind. 

Testing capacity in the U.S. has ramped up dramatically since March, and now hovers around 700,000 tests per day. But that’s still well below than the more than 4 million per day public health experts say are needed to suppress the pandemic. Meanwhile, conducting more tests has exposed a secondary problem: Labs don’t have the capacity to process them quickly. In some parts of the country, the wait time between getting tested and getting results stretches to more than two weeks. In those two weeks, people who aren’t yet aware they’re infected — and may not ever show symptoms — can pass the virus on.

“There is the potential for additional disease spread that’s occurring because the turnaround time is so long,” Feng said.

Speed and availability aren’t the only problems. The most common kind of test for SARS-CoV-2 detects the virus by using a method called a reverse transcription polymerase chain reaction, or RT-PCR, to scan for fragments of its RNA in a swab taken from the cavity between the patient’s nose and mouth. The accuracy of these tests is slowly improving, but they still miss a substantial portion of infections: in one recent study, around 38 percent of people who actually had the virus got a negative result from RT-PCR. “These people can unwittingly spread the virus, too,” said Feng.

But scavenging for genetic material isn’t the only way to identify an infection. COVID-19 inflicts damage all over the body, but its most devastating effects are often in the lungs. There, the virus invades the small sac-like alveoli, causing inflammation that fills the tissue with fluid. These clogged alveoli show up as hazy patches on CT scans called ground-glass opacities. Early studies in China suggested that these hazy patches, along with other hallmark features, were more reliable than RT-PCR in indicating whether a patient had the virus. 

These features can be subtle, however, especially early in an infection. Feng and Cao, who had already worked together on image-enhancing software for CT, reasoned that sharpening the scans of patients’ lungs could make COVID-induced changes easier to spot (and easier to distinguish from other respiratory diseases that cause similar inflammation). 

“The idea is that if we enhance these features so that they come out more definitively, the radiologist could better diagnose whether one might have COVID-19,” Feng said.

The algorithm creates those sharper scans by working backwards, taking advantage of knowledge about how a CT scanner creates its images.

The cross-sections we see in CT scans are assembled from hundreds of individual projections created as the X-ray generator in the machine travels around the patient’s body. The algorithm Feng and Cao developed artificially deconstructs the image into its precursor, a simpler combination of projections called a sinogram. It then refines the sinogram and converts it into an image that’s less blurry than the original. This cycle is repeated over and over again, eventually arriving at an image with dramatically improved resolution. All this happens in less than five minutes — a stunning improvement over the hours required to process a RT-PCR test in even the best-case scenario. 

The algorithm was honed through machine learning processes, which used pairs of high- and low-resolution CT images and a wealth of open-source COVID-19 data to train the software to accurately produce a crisp, detailed image from a blurry original. 

“If we understand the physics and the image formation processes, we can use that knowledge to enhance CT image quality,” Cao said. “You can combine the deterministic power of physics with the statistical power of artificial intelligence, and you get the best of both.” 

It’s a combination that could far outstrip the accuracy of RT-PCR, and do so with significantly less labor and materials. RT-PCR tests require truckloads of specialized swabs, tubes, and reagents to be deployed to testing sites and labs. All a CT scan requires is a CT scanner, a proven technology already available in hospitals and clinics around the country.

The researchers estimate that their method could boost the nation’s daily testing capacity by more than 400,000. It could also help doctors track the progression of the disease. 

“Hallmark features show up with increasing frequency as a person moves through COVID-19,” Feng explained. “It enables you to monitor it and treat it appropriately, which the PCR test is not able to do.”

Feng and Cao hope to share their approach as open-source software by the end of the summer. Because it works from existing CT images, it can be integrated easily into a hospital’s workflow. Updates can be distributed and downloaded virtually in real time.   

With a second NSF grant, Feng and Cao are developing a version that, over the longer term, could be integrated directly into a CT scanner, enhancing the resolution of the image during the scan itself.

“Our approach has applications beyond COVID-19,” Cao said. “It’s conceivable that our algorithms could be used to enhance medical images when a high-powered CT scanner isn’t available, which would be useful in a variety of medical contexts – not just COVID-19.”

In immediate response to the COVID-19 pandemic, Virginia Tech faculty, staff, and students have initiated numerous research projects with local and global salience. Learn more from the Office of the Vice President for Research and Innovation.

Written by Eleanor Nelsen.