Waterloo startup DarwinAI and researchers from the University of Waterloo are confident its open-source project, COVID-Net, and a sprinkle of AI could help radiologists spot COVID-19 in patients much faster and provide deeper insights into patients’ symptoms.
While the Canadian Association of Radiologists has made it clear that the final diagnosis of COVID-19 infection should be confirmed by a positive RT-PCR test, not a CT chest scan, Dr. Alexander Wong, lead researcher on the project, said the additional information from CT chest scans can be hugely beneficial for radiologists.
“One of the biggest issues right now is that chest x-rays have become quite an important factor in the screening process. In many cases, it is used alongside viral testing because viral testing alone could take a while. And also, while specificity is great, sensitivity is at around the 70 per cent range, so having X-ray to help give additional information, as well as additional insights on severity is important,” Wong told the publication.
The ultimate goal, he added, is to build an AI that could help radiologists and clinicians distinguish COVID-19 and other illnesses that appear in medical imagery.
“Fundamentally we’re building deep neural networks that are then trained by being exposed to the plethora of chest x-rays, with both COVID positive as well as COVID negative so that these networks are able to learn to differentiate between COVID-19 intersections and other forms of infections,” Wong explained.
Wong, who is also the Canada Research Chair in AI and medical imaging and an associate professor in the department of systems design engineering at the University of Waterloo, said the project came together in seven days.
Ever since word about the project spread last month, the dataset has grown significantly. The project started with a dataset featuring nearly 6,000 posteroanterior chest radiography images across roughly 2,900 patient cases gathered from public sources. That number has since ballooned to approximately 16,760 chest x-rays across 13,700 patient cases.
The other key focus is risk stratification, which is using AI to be able to predict risk levels so that individuals can have better-individualized care based on their risk level, as well as to help improve the patient population management, indicated Wong.
“Being able to understand the risk as well as the severity of a patient allows clinicians and doctors to triage as well as manage patients better, such as understanding if they need ventilators, other forms of treatment or they could be self-isolated at home,” he said.
The instructions and scripts on generating this enhanced collection have been posted to the company’s GitHub repo, Sheldon Fernandez, chief executive officer of DarwinAI wrote in a press release.
Currently there are two models that have been made available, COVIDNet-Large and COVIDNet-Small. Each have different fundamental trade-offs between them, Wong explained. COVIDNET-Large is more sensitive when it comes to detecting COVID, but also requires more computing resources, while the smaller one has much better efficiency and speed that can be more appropriate especially in situations where one needs to run things on the edge, or in a local machine.
Wong also told the publication that they are working with a lot of different clinical partners to get as much data as possible.
“One thing to note is that they are not production-ready. But the goal here is that, by driving this open-source initiative, everyone can work together towards clinical viability. So, the goal is to keep improving upon it as more data comes in and as more and better models can be built based around that data,” he said.
