An artificial intelligence has been developed capable of recognizing Covid-19 patients from lung x-rays: ten times faster than the human eye in analyzing images, it produces results with an accuracy of 82%, therefore even 6 times higher % compared to that achieved by radiologists.
This is demonstrated in the journal Radiology by a study conducted by researchers at Northwestern University, who have made the algorithm public to allow anyone to develop it further in view of its potential use in the clinical setting.
"We do not intend to replace the current test method with swabs," says artificial intelligence expert Aggelos Katsaggelos.
If anything, the goal is to offer a complementary tool to screen patients admitted to hospital for reasons other than Covid-19.
"Radiographs are routine, safe and inexpensive tests. Our system - adds Katsaggelos - would take a few seconds to examine a patient and determine if he should be isolated".
This would reduce the risk of contagion for healthcare professionals and other hospitalized patients.
"It can take hours or days to get the swab results," says cardiologist Ramsey Wehbe.
"Artificial intelligence cannot confirm whether a patient is positive or not, but it can report it to speed up triage even before the swab results arrive."
The DeepCOVID-XR algorithm has indeed learned to recognize the Covid-19 imprint on the lungs after examining more than 17,000 chest x-rays.
Once the training was complete, the system was tested on 300 new radiographs and its performance was compared to that of five specialist radiologists.
Doctors took two to three and a half hours to evaluate the radiographs, achieving an accuracy of 76-81%.
Artificial intelligence, on the other hand, took only 18 minutes to reach an accuracy of 82%.