This algorithm is better at finding tiny brain hemorrhages than human experts.
The technology, developed by scientists at UC San Francisco and UC Berkeley, could one day help doctors treat patients with traumatic brain injuries, strokes, and aneurysms.
Radiologists look at thousands of head scans a day, searching for itty-bitty abnormalities that can signal life-threatening emergencies.
That’s a lot of pressure.
Enter an algorithm.
It’s a lot more efficient (and potentially more accurate) for artificial intelligence to scroll through stacks of images, flagging those with significant abnormalities and passing them on to radiologists for further examination.
“We wanted something that was practical, and for this technology to be useful clinically, the accuracy level needs to be close to perfect,” study co-author Esther Yuh, associate professor of radiology at UCSF, said in a statement.
“The performance bar is high for this application, due to the potential consequences of a missed abnormality,” she continued. “And people won’t tolerate less than human performance or accuracy.”
Their final product, detailed in a paper published this week by the journal Proceedings of the National Academy of Sciences (PNAS), took just one second to determine whether an entire head scan contained any signs of hemorrhage.
It even found some small abnormalities that experts missed.
The algorithm also noted their location within the brain and classified them according to subtype—and did it all with “an acceptable level” of false positives.
Now it’s just showing off.
To make this clinically useful, though, the technology must achieve 100 percent accuracy.
“If a computer is pointing out a lot of false positives, it will slow the radiologist down, and may lead to more errors,” Yuh explained.
The key, according to Jitendra Malik, professor of electrical engineering and computer sciences at Berkeley, was choosing which data to feed into the model.
Using a fully convolutional neural network (FCN), researchers trained their algorithm on a relatively small number of images—4,396 CT exams, each image was packed with information.
The richness of this data, according to the team, helps prevent “noise” to create an extremely accurate system.
“We took the approach of marking out every abnormality—what’s why we had much, much better data,” Malik, a co-author of the study, said. “Then we made the best use possible of that data. That’s how we achieved success.”
Moving forward, researchers are applying the algorithm to CT scans from trauma centers across the country.
More on Geek.com:
- Watch: Human-Like Robot Hand Learns to Solve a Rubik’s Cube
- This Insulin-Delivering Pill Could Replace Injections for Diabetes
- World’s First 5G-Powered Remote Brain Surgery Performed in China
from Geek.com https://ift.tt/2BRsKEJ
via IFTTT
0 comments:
Post a Comment