Over the past week, companies around the world announced a flurry of AI-based systems to detect COVID-19 on chest CT or X-ray scans. Already, these deep learning tools are being used in hospitals to screen mild cases, triage new infections, and monitor advancing disease.
AI-powered analysis of chest scans has the potential to alleviate the growing burden on radiologists, who must review and prioritize a rising number of patient chest scans each day, experts say. And in the future, the technology might help predict which patients are most likely to need a ventilator or medication, and which can be sent home.
“That’s the brass ring,” says Matthew Lungren, a pediatric radiologist at Stanford University Medical Center and co-director of the Stanford Center for Artificial Intelligence in Medicine and Imaging. “That would be the killer app for this.”
Some companies are selling their tools, others have released free online versions, and various groups are organizing large crowdsourced repositories of medical images to generate new algorithms.
“The system we designed can process huge amounts of CT scans per day,” says Hayit Greenspan, a professor at Tel-Aviv University and chief scientist of RADLogics, a healthcare software company that recently announced [PDF] one such AI-based system. “The capability for quickly covering a huge population is there.”
Four COVID-19 lung CT scans (top) with corresponding colored maps showing coronavirus abnormalities (bottom). Images: RADLogics
While many of the systems tout impressive numbers to diagnose COVID-19 from chest scans—RADLogics, for example, is reporting [PDF] up to 98 percent accuracy—there is little likelihood these AI tools will supplant standard nucleic acid tests as the primary diagnostic tool for coronavirus infection. A swab for a nucleic acid test can be taken in a car or any isolated location, while a chest scan is performed in an enclosed space with staff nearby. That technique presents challenges including the patient’s radiation dose and increased exposure to staff and facilities, and the need for staff to use valuable personal protective equipment. The American College of Radiology, the U.S. Centers for Disease Control and Prevention, the Royal College of Radiologists in the United Kingdom, and many other national and international organizations explicitly recommend against the use of CT as a first-line screening test.
“I wouldn’t use this as a primary screening tool, but I would use it for opportunistic screening,” such as flagging suspicious CTs or X-rays on patients who received imaging for unrelated medical reasons, says Lungren.
Early evidence that chest scans might be useful in the fight against COVID-19 emerged in a series of papers published in February in the journal Radiology. Teams in China and the United States found that the lungs of patients with COVID-19 symptoms had certain visual hallmarks, such as ground-glass opacities—hazy darkened spots in the lung diffuse enough that they don’t block underlying blood vessels or lung structures—and areas of increased lung density called consolidation. Those characteristics became more frequent and more likely to spread across both lungs the longer a person was infected.
Last week, RADLogics published preprint research [PDF], meaning it has not yet been peer-reviewed by other scientists, validating its AI-powered system trained on multiple international datasets. After a chest scan, the system can give immediate alerts if a patient needs to be seen. The tool also tracks a patient’s progress by providing a numerical “Corona score”—a measurement of disease severity—which can be used to quantify the disease over time. The software is currently being deployed in China, Russia, and Italy.
“We see two ways we can contribute: identification, and monitoring and predicting patient status,” says Greenspan. The latter might help in situations with limited hospital resources: By training the tool on multiple chest scans from individual patients over time, Greenspan hopes the system will soon be able to predict which patient will need a ventilator soon or who no longer requires one. The group is working in an Italian hospital to identify specific treatments that could be tied to the Corona score as patients progress or recover.
In late February, Chinese technology giant Alibaba Group announced an AI algorithm from its research unit, DAMO Academy, that can diagnose suspected cases within 20 seconds with 96 percent accuracy. On a company news site, DAMO member Xu Minfeng said the algorithm is being used in 26 hospitals in China, where it has already helped diagnose more than 30,000 cases.
Alibaba’s algorithm is said to be trained on more than 5,000 confirmed coronavirus cases and, like RADLogics’ system, also tracks treatment responses, such as detecting signs of improvement including a reduction of white mass in the lungs, said Minfeng. DAMO is working with partners to bring the AI system into the cloud, where medical staff could directly upload CT scans using their smartphones or laptops to get instant results, he added.
On Monday, 30 March, Seoul-based medical AI software company Lunit released its AI-powered software for chest X-ray analysis online free of charge. The software is already assisting with screening in coronavirus care centers in South Korea, and especially as a tool to triage patients for radiologists with case overloads, according to a press release from the company. It has also been installed in one of the largest hospital networks in Brazil to screen patients with mild symptoms.
As more companies and teams announce AI tools for chest scans, cooperation will be critical, says Lungren. “It is important we all share our methods and data so we can build off one another’s work.”
Accordingly, there are several major efforts to compile large open data repositories of images and associated data from hospitals and societies around the world. On 30 March, the Radiological Society of North America, for which Lungren is the task force chair, announced the creation of an international research and education initiative for COVID-19-related imaging data, and will collaborate with the similarly large European Imaging COVID-19 AI initiative to coordinate efforts.
COVID-Net, a nother open-source project designed to collect and analyze chest X-rays, recently reported a rapidly growing dataset and the development of a neural network tailored for COVID-19 risk stratification.
Megan is an award-winning freelance journalist based in Boston, Massachusetts, specializing in the life sciences and biotechnology. She was previously a health columnist for the Boston Globe and has contributed to Newsweek, Scientific American, and Nature, among others. She is the co-author of a college biology textbook, “Biology Now,” published by W.W. Norton. Megan received an M.S. from the Graduate Program in Science Writing at the Massachusetts Institute of Technology, a B.A. at Boston College, and worked as an educator at the Museum of Science, Boston.