COVID-19 shown to leave unique lung fingerprint
Researchers at Lawson Health Research Institute (Lawson) have developed and tested an artificial neural network for diagnosing COVID-19. The AI system was trained to learn and recognize patterns in ultrasound lung scans of patients with confirmed COVID-19 infection at London Health Sciences Centre (LHSC) and compared them to ultrasound scans of patients with other types of lung diseases and infections.
“The AI tool that we developed can detect patterns that humans cannot. Lung ultrasound scans of patients with COVID-19, as well as other lung diseases, produce a highly abnormal imaging pattern, but it is almost impossible for a physician to tell apart different types of infections by looking at these images. There are details, however, that distinguish COVID-19 at the pixel level that cannot be perceived by the human eye,” explains Dr. Robert Arntfield, Lawson Researcher and Medical Director of the Critical Care Trauma Centre at LHSC.
“The neural network was able to identify the unique characteristics among different scans, and exceed human-level diagnostic specificity. Our study of over 100,000 ultrasound images showed that while trained physicians could, as expected, not distinguish between different causes of lung disease, the AI had nearly perfect accuracy in making the diagnosis. It’s almost like the AI sees a QR code that we cannot see, unique to the disease.”
LHSC is a global leader in the use of point-of-care ultrasound. It has become an important tool for the diagnosis and care of critically ill patients experiencing acute respiratory failure. The convenience, portability and low cost of using these machines makes them ideally suited for pandemic conditions.
This study is part of a collaborative effort by a small group of local clinicians driven to innovate and create technology that solves complicated problems with AI. “There are a lot of brilliant minds in our city, and I’m very proud that we were able to rapidly pull together a local team to design, develop and test a complex idea,” says Dr. Arntfield.
“We are already expanding on these findings with more research.” Lawson has recently approved Dr. Arntfield’s “Project Deep Breathe” which aims to go beyond COVID-19 and explore multiple conditions where lung ultrasound and AI can be paired together.
The study, “Development of a convolutional neural network to differentiate among the etiology of similar appearing pathological B lines on lung ultrasound: a deep learning study”, is published in BMJ Open.