When the work of radiologists is augmented with artificial intelligence, the results outperform those of both radiologists and AI working independently, a recent study led by researchers from Stanford (Calif.) University School of Medicine found.
The study, which was published in npj Digital Medicine and also included collaborators from Durham, N.C.-based Duke University Medical Center and Unanimous AI, compared the accuracy of two deep learning models, a group of expert radiologists and a combination of the two when tasked with diagnosing pneumonia from chest radiographs.
While the AI models were found to outperform the humans when each worked separately, a combined "human-in-the-loop" model was found to produce the greatest diagnostic accuracy of all, by "harnessing the advantages of both while at the same time overcoming their respective limitations," according to the report.
"The clinical significance of this could imply that, in a landscape of increasing clinical volumes, complexity of cases and medical record documentation, physicians could leverage deep learning to improve operational efficiency," the study's authors wrote, while "deep learning algorithms could provide automated rapid diagnosis for high-confidence cases as a triage tool, so that physicians could spend less time on high-confidence cases evaluated by an AI model and more time on relatively complex cases."