As the number of COVID-19 infections are again spiking around the U.S., health care workers struggling to stay ahead have a tool with a novel approach to add to their arsenal in COVID-Net, an open source AI-based platform that uses radiological lung images to determine COVID-19-specific lung damage, as well as assess the degree of that damage.
The technology was developed in March, during the early days of the pandemic, but has been gaining more notice as an example of artificial intelligence in health care as more organizations have adopted it.
Although the nonprofit project is being led by Red Hat, Boston Children's Hospital and DarwinAI (a 3-year-old proprietary artificial intelligence startup headquartered in Waterloo, Ontario), it began as a collaboration between Canada's University of Waterloo and DarwinAI.
"COVID-Net was an initiative to try to contribute to the whirlwind of the pandemic in March," DarwinAI CEO Sheldon Fernandez told ITPro Today. "We open sourced it and we didn't want it to be commercial. We wanted it to just give practitioners another tool that they could leverage to battle COVID. That was the motivation."
Fernandez said that DarwinAI is "organically connected" with the hometown university and was already working with its researchers on other projects when COVID-19 became a global issue and researchers around the world turned to artificial intelligence in health care as a potentially valuable tool in understanding and diagnosing COVID-19.
He said that in seven days they were able to harness DarwinAI's technology to develop a platform that not only helped researchers ascertain that COVID-19 leaves unique visual markers, but also used machine learning techniques to train the system to distinguish between images of COVID-19-infected lungs and images of healthy lungs or lungs that were damaged by causes other than the coronavirus.
"When corona was significant and went from being a curiosity to suddenly stopping the world, very quickly we pivoted and asked if we could leverage our domain expertise and just move it to COVID," Fernandez said. "So we fed the system all these images and using our technology very rapidly created a system that started to figure out these fluctuations in the lungs and the respiratory system are markers for COVID, and the system got gradually smarter over time."
The process was slow and cumbersome, however, and couldn't handle a great number of images quickly enough to be particularly useful. In addition, the AI technology wasn't user-friendly and required people with engineering skills to operate it. To address these issues, with Red Hat's help, the COVID-Net team connected with Boston Children's Hospital, which had developed a container platform that's now called ChRIS that ran on OpenShift and OpenStack.
"They already had [COVID-Net] developed but didn't really have a way to deploy it, at least not simply and in such a way that would be useful for clinicians," Rudolph Pienaar, assistant professor in radiology at Harvard Medical School and staff scientist at Boston Children's Hospital, told ITPro Today. "That's where we came in. They were looking for some mechanism to deploy their application and also to build a user interface to their application to make it easy for folks like doctors to use."
Pienaar said that integrating the existing COVID-Net technology with the ChRIS platform was accomplished in about three days.
"What's taken more time is to build a compelling user interface to expose that functionality to an end user," he said. "Part of that was done conceptually over at DarwinAI, and then one of the Red Hat designers, Máirín Duffy, sort of productized or hardened that design."
Although the platform can be used as a diagnostic tool to determine whether a patient has contracted COVID-19, that's not a recommended primary use of the technology, since chemical-based tests using swabbed samples are more accurate. Pienaar said that one of the ways the platform is most useful is as a triage tool, to help doctors determine who has the most immediate medical need in facilities that are being overwhelmed with COVID-19 cases.
"The real benefit is that the human eye becomes tired," he added. "To me I think that's the main thing with these kind of medical diagnoses. A human is able to make extremely good and accurate diagnoses, but is subject to fatigue. A machine algorithm might not make as good predictions always – sometimes it will come close [or] it might be better – but it will always give the same results for the same input. It never gets tired. It will never become bored. It will be more reliable from that perspective."
Indeed, this benefit is a main driver of the use of artificial intelligence in health care.
The source code, documentation, dataset and scientific papers describing COVID-Net and its approach to AI in health care are available on the project's GitHub repository.