A deep learning algorithm can accurately detect acute intracranial hemorrhage (ICH) on head CT on par with highly trained neuroradiologists, in some cases identifying subtle abnormalities overlooked by the radiologists, new research shows.
Only a "handful" of artificial intelligence (AI) applications in medical image interpretation have achieved this level of accuracy, Esther Yuh, PhD, of the University of California, San Francisco (UCSF), told Medscape Medical News.
The study was supported by the California Initiative to Advance Precision Medicine and was published online October 21 in the Proceedings of the National Academy of Sciences.
Head CT is the "workhorse" medical imaging modality for diagnosing neurologic emergencies, such as acute traumatic brain injury, stroke, and aneurysmal hemorrhage, the investigators note.
"However, these gray scale images are limited by low signal-to-noise, poor contrast, and a high incidence of image artifacts. A unique challenge is to identify tiny subtle abnormalities in a large 3D volume with near-perfect sensitivity," they write.
Yuh, with colleagues at UCSF and UC Berkeley, developed a neural network called PatchFCN and trained it using 4396 CT scans.
They compared the algorithm's performance to that of four American Board of Radiology–certified radiologists on an independent test set of 200 randomly selected head CT scans (25 positive and 175 negative).
The algorithm was highly accurate. It achieved a receiver operating characteristic area under the curve of 0.991 for identification of CT scans that were positive for acute ICH, and it exceeded the performance of 2 of 4 radiologists, the researchers report.
"In addition, PatchFCN achieved 100% sensitivity at specificity levels approaching 90%, making this a suitable screening tool for radiologists based on an acceptably low proportion of false positives," they write.
The algorithm can perform pixel-level delineation of abnormalities and classify abnormalities into different pathologic subtypes.
The potential benefits of the algorithm include "accuracy, or reduction of errors in image interpretation and speed ― getting results to emergency department ordering physician more quickly," Yuh told Medscape Medical News.
There are also potential cost and efficiency benefits.
"Medical imaging utilization continues to increase every year, and radiologists can potentially be more efficient and handle more if supported by the AI," said Yuh.
"All of these potential benefits, of course, need to be validated in a 'real-world' situation," she said.
The researchers are now applying their algorithm to CT scans from trauma centers across the United States.
"The fact that this algorithm was able to do better than 2 out of 4 radiologists is pretty powerful," Cyrus Raji, MD, PhD, assistant professor of neuroradiology at Washington University in St. Louis, Missouri, who wasn't involved in the research, told Medscape Medical News.
"Where I see this algorithm as really helping us as radiologists is in identifying and flagging hyperacute findings and prioritizing those findings for us to read first. It can provide a way of ranking CT scans and knowing, when faced with many scans, which are most important to look at first, based on what the algorithm flags," said Raji, who has received consulting fees from Brainreader ApS.
"What you will see in the coming years and what's already starting," he added, "is a wave of artificial intelligence products and companies that will go for FDA [US Food and Drug Administration] clearance for application in the clinical setting."
Proc Natl Acad Sci. Published online October 21, 2019. Full text
Medscape Medical News © 2019
Cite this: Artificial Intelligence Rivals Experts in Diagnosing Brain Bleeds - Medscape - Oct 23, 2019.