Skip to main content

UW researchers find more precise way to detect COVID-19 pneumonia

October 16, 2020 By Emily Kumlien

Using cutting-edge artificial intelligence technology, University of Wisconsin‒Madison investigators have developed a far more precise way to identify cases of COVID-19-induced pneumonia.

Using a custom artificial intelligence algorithm called CV19-Net, the UW research team dug into a vast resource database of tens of thousands of COVID-19 chest X-rays to show its method could identify pneumonia caused by COVID-19 at a sensitivity of 88 percent, according to Guang-Hong Chen, professor of medical physics and radiology at the UW School of Medicine and Public Health.

From the tens of thousands of X-rays available, the team pared down the number of X-ray images to train the algorithm and then evaluated its performance over 5,900 X-rays from approximately 3,000 patients between Feb. 1 and May 3.

To compare to diagnoses generated by the human eye, Chen’s team asked three expert thoracic radiologists experienced with COVID-19 pneumonia X-ray images to examine 500 chest X-ray images from the CV19-Net database. The three radiologists were able to correctly perform diagnosis with accuracy of 76 percent, 68 percent and 72 percent. In contrast, the CV19-Net algorithm examined the images and achieved a diagnostic accuracy of 84 percent.

“It is clear: Based on the data, we conclude that artificial intelligence can identify COVID-19 pneumonia better than the human eye,” Chen says.

The results of the research were recently published in the journal Radiology.

Chen and the research team that includes SMPH’s Ran Zhang, assistant scientist in medical physics, and Scott Reeder, professor of radiology and medical physics, and other researchers and clinicians at both SMPH and Henry Ford Health System in Detroit. Reeder is also a UW Health radiologist.

The team is currently determining how to utilize this new technology to help health care workers in the field identify COVID-19 cases in just minutes using X-ray techniques rather than more costly and less available CT scans, Reeder says.

Such an algorithm could even be deployed into the X-ray machine itself so that the detection of COVID-19 pneumonia could be made before the images are transmitted to the radiologist’s computer screen, he says.

“The algorithm could even page the radiologist to alert them to review the case in a real-time manner, so that a diagnosis and report can be made within just a few minutes,” Reeder says. “Indeed, it would be a straightforward extension to even generate a preliminary report, before the radiologist has even reviewed the X-ray images.”

Chen and his team are working with scientists at Epic, a Verona, Wisconsin, company that provides health record software to hospital systems, and UW Health to develop the clinical use of the algorithm. The algorithm would produce a COVID-19 risk score immediately after a chest X-ray image is taken, Chen says.

A next step would be to create a more universal algorithm for COVID-19 screening, he said — not just for COVID-19 cases with pneumonia findings, but also for people with mild or no pneumonia findings.

“Once developed, this could become a fully automatic tool for COVID screening,” Chen says. “Again, it just underscores the power and potential of artificial intelligence in medical practice.”

The work received funding from the Wisconsin Partnership Program.