A research team at London’s Lawson Health Research Institute is working to determine whether trained artificial intelligence can be used to diagnose COVID-19 by comparing lung ultrasound scans of positive patients to those without the disease.
As part of the project, team members will train an artificial neural network, designed to loosely mimic how brains compute, to detect small patterns in the ultrasounds that humans wouldn’t otherwise see, said researcher Dr. Robert Arntfield.
Researchers say that ultrasound scans of patients with COVID-19-related pneumonia produce a “highly abnormal imaging pattern,” but one that isn’t unique to the disease.
“We’re hoping to establish whether there is a uniqueness to what COVID-19 puts on the lungs as opposed to other similar diseases, whether it be influenza or other causes of pneumonia,” said Arntfield, who is also medical director of LHSC’s Critical Care Trauma Centre, in an interview Wednesday with 980 CFPL’s Devon Peacock.
Such a breakthrough would, in theory, allow health-care workers to link COVID-19 to a patients’ lung problems much more quickly compared to a standard swab test, he says.
The A.I. technology being utilized is not entirely unlike the A.I. systems deployed by social media websites and smartphones to distinguish between faces and objects in photographs.
“The difference here is we’re actually trying to harness the power of the computer to see if there’s some difference between a COVID lung and a non-COVID lung. Humans are actually not able to do this task,” Arntfield said.
“We believe, or we’re hopeful, that at the pixel level, at a level that exceeds human vision and human cognition, that the machine will actually detect a pattern in all of that noise.”
Researchers are currently in the process of training the neural network by showing it a large quantity of lung ultrasound scans from patients who have been critically ill from COVID-19, as well as scans from patients with other types of lung infections taken before the pandemic.
“In so doing, we’re training it to develop the eventual capacity to receive what’s called a test set, or a validation phase, where we will show it these pictures without the labels on them and look to it to perform accurately by sorting the images,” Arntfield said.
The research team hopes to have results in the next two to three weeks, with publication to come soon afterward. According to Lawson, many on Arntfield’s team have backgrounds in computer programming and wrote the code of the neural network being used.
Arntfield says such technology could be deployed to other areas of medical diagnostics, and in some cases already has.
“There’s been lots of work already… particularly in imaging, such as CAT scans or the X-rays in showing the capacity for machines to often exceed human performance in recognizing either subtle findings, or as a backstop to shore up diagnostic accuracy.
“The future for A.I. in medicine is wide open and is very exciting.”
Comments