Improving the Flow of COVID-19 Screening with X-rays and AI
Professor Moulay Akhloufi didn’t expect to join the effort against COVID-19. And while it may be difficult to imagine X-rays and artificial intelligence helping in the face of a world-paralyzing virus, that’s exactly what Akhloufi and his computer science lab—PRIME—at the Université de Moncton are hoping to do.
Akhloufi and two of his graduate students, Mohamed Chetoui and Andy Couturier, are teaching AI to recognize the signs of COVID-19 on X-ray images of patients’ lungs. So far, the results look promising.
PRIME stands for Perception, Robotics and Intelligent Machines. The group uses deep learning, the science of developing machines that can learn by example. “It consists of various types of algorithms that form artificial neural networks,” says Akhloufi. In simpler terms, they use math to mimic the human brain. The applications are as varied as modelling forest fires and autonomous navigation of drones.
“We’ll take data representing this or that category and train the network to differentiate them,” explains Akhloufi. By showing the machine images of lungs that are healthy, or have pneumonia or COVID-19, it gets progressively better at telling them apart.
When the pandemic hit Canada, Akhloufi’s team wasn’t overly affected. “We’re in computer science,” says Akhloufi. “We adapted very quickly to working from home.” With remote access to ACENET and Compute Canada’s computing power to pummel through their complex calculations, it was business as usual.
Chetoui, who was finishing his master’s, then had the idea to adapt his project to this new coronavirus disease. He had developed algorithms to recognize and give severity scores for diabetic retinopathy, a complication from diabetes causing damage to the blood vessels of the eye’s retina. A crossover to COVID-19 seemed feasible.
Around mid-February, Spain and Italy’s health systems were overloading, creating a shortage of materials necessary for the tests that diagnose the disease. Radiologists began using X-rays, CT scans and ultrasounds of patients’ lungs to make quicker diagnoses and made the images public.
With this first batch of images and others coming from Chinese studies, the team developed two algorithms. One gives the probability of infection and the other, a heat map of the affected areas. Near the end of March, they put up a website that allows medical professionals to upload and obtain readouts for their X-ray images.
“When the Institut du Savoir Montfort joined us, the project took a whole other dimension,” adds Akhloufi. The Hôpital d’Ottawa’s research institute supplied them with more images to improve the tool. Akhloufi says that predictions on a set of more than 5,000 images are now close to 98 per cent accurate and will only get better with more data.
Because most images come from patients presenting symptoms, it is too early to tell if AI will be as effective with asymptomatic people. Although a team at the University of California San Diego detected one case through a similar technique, they’ll need more targeted data to be sure it wasn’t just a lucky guess.
Akhloufi hopes AI will help improve the flow of COVID-19 screening by offering radiologists and other medical professionals a fast and reliable tool to inform their diagnoses. “The aim isn’t to replace radiologists,” he specifies. “When it comes to my health, I’d rather speak to another human.”