Northwestern University researchers are touting a new AI platform that can detect COVID-19 in X-ray images of the lungs.

DeepCOVID-XR, a machine-learning algorithm, performed better than a team of specialized thoracic radiologists by spotting COVID-19 in X-rays about 10 times faster and 1%-6% more accurately, according to a news release.

Researchers involved think physicians could put the AI system to use in the rapid screening of patients who are admitted to hospitals for reasons other than the coronavirus, leading to faster detection of the virus and potentially protecting healthcare workers and other patients by triggering an isolation period for the patient in a faster fashion.

Authors of the study also theorize that the algorithm might be capable of flagging patients for isolation and testing who are not otherwise under investigation for COVID-19. The study, highlighting a method that won’t replace actual testing but can help determine a patient’s condition, was published today in Radiology.

“It could take hours or days to receive results from a COVID-19 test,” Dr. Ramsey Wehbe, a cardiologist and postdoctoral fellow in AI at the Northwestern Medicine Bluhm Cardiovascular Institute, said in the release. “AI doesn’t confirm whether or not someone has the virus. But if we can flag a patient with this algorithm, we could speed up triage before the test results come back.”

Chest X-rays of COVID-19 patients return images of what researchers describe as “patchy and hazy” lungs, rather than clear, healthy organs. Pneumonia, heart failure and other illnesses can look similar on the lungs, meaning a trained eye is required to differentiate between COVID-19 and something less contagious.

The researchers used 17,002 chest X-ray images, 5,445 of which came from COVID-19-positive patients from sites across the Northwestern Memorial Healthcare System. They tested DeepCOVID-XR against five experienced radiologists on 300 random test images, with each radiologist taking approximately 2.5 to 3.5 hours, while the AI algorithm took 18 minutes. Accuracy from the radiologists ranged from 76% to 81%, while the algorithm had 82% accuracy.

Limits do exist within the system, as not all COVID-19 patients show signs of illness, meaning the AI system wouldn’t flag the patient as positive. However, neither would a radiologist, hence why radiologic diagnosis would not replace testing, but rather supplement it in certain cases.

The researchers said they have made the algorithm publicly available in the hopes that others train it with new data. It is still in the research phase, but could be used in a clinical setting down the road.