CT scans, or Computed Tomography, of the brain are the gold standard of tests when it comes to examining the brain for evidence of stroke or other life-threatening brain-related emergencies. But the images aren’t always top quality, making errors or oversights by radiologists a distinct possibility.
A new AI program, however, may be here to help.
According to a study recently published in the Proceedings of the National Academy of Sciences, researchers have developed an AI-driven “neural network to achieve accuracy levels comparable to that of highly trained radiologists, including both identification and localization of abnormalities that are missed by radiologists.”
"This AI can evaluate the whole head in one second," said senior researcher Dr. Esther Yuh, an associate professor of radiology at the University of California, San Francisco, in a review of the report. "We trained it to be very, very good at looking for the kind of tiny abnormalities that radiologists look for.”
That “one second” can mean a lot, Yuh explained, given that any delay in treating a stroke can result in a more debilitated patient due to the death of more brain cells.
For the study, the researchers accessed a library of nearly 4,440 CT scans to train their AI to look for brain bleeding, training it “to the point that it could trace detailed outlines of abnormalities it found, demonstrating their location in a 3-D model of the brain being scanned. They then tested the algorithm against four board-certified radiologists, using a series of 200 randomly selected head CT scans.”
While the AI slightly outperformed two radiologists, and slightly underperformed against the other two, it found some small abnormalities that the experts missed while also providing detailed information that doctors would need to determine the best treatment.
Moreover, Yuh emphasized, "Instead of having a delay of 20 to 30 minutes for a radiologist to turn around a CT scan for interpretation, the computer can read it in a second.”
The computer program also provided this information with an acceptable level of false positives, Yuh said, which ideally would minimize how much time doctors would need to spend reviewing its results.
"Doctors train for years to be able to read these correctly," Yuh noted.