‘Like magic’: AI bests radiologists in correctly diagnosing collapsed lung

New assistive technology developed by one Canadian university could serve as a “computational second opinion” for busy radiologists assessing difficult-to-diagnose collapsed lungs, experts said this week.

University of Waterloo, Ontario, researchers have developed the new software, which uses AI to scour tens of thousands of chest x-rays. It then compares a patient’s images with those from the database to determine if the individual has a collapsed lung.

The tool has shown early promise, helping researchers correctly diagnose the condition—also known as pneumothorax—about 75% of the time, compared to 50% for medical specialists using only x-rays. They are presenting those findings at the Evolution of Deep Learning Symposium in Toronto, which kicked off Wednesday, Oct. 16.

“Our results are very exciting,” Antonio Sze-To, a postdoctoral fellow at Waterloo, told the university’s news website. “The AI we use works almost like magic—and it will help radiologists save lives.”

Pneumothorax is a painful condition that can snowball if not treated quickly enough. Radiologists typically diagnose its more severe form with ease, but minor instances are “extremely challenging” to pinpoint, Waterloo experts noted. Missing that critical diagnoses can waste clinicians’ time and energy and potentially put patients at risk.

With that in mind, the school’s experts have worked alongside the University Health Network research organization to develop a remedy. Backed by the AI-focused nonprofit Vector Institute, they hope to eventually increase its diagnostic accuracy to 90%. Those involved are also planning to integrate the AI tool into Coral Review, a quality-assurance software system used by UHN-affiliated hospitals that allows doctors to compare their decisions to those made by their peers.  

Waterloo researchers believe the technology would be useful in diagnosing other conditions that require a chest x-ray, such as pneumonia and COPD. They anticipate that such computational second opinions will become commonplace in medicine in the near future.

“There is no question systems like this will be in place in hospitals within the next two years. People are pushing for it and the technology is there,” Hamid Tizhoosh, a professor of systems design engineering and director of the Laboratory for Knowledge Inference in Medical Image Analysis, told the Waterloo News.