Radiologists are under increasing pressure to be more and more efficient. And with that pressure, their interest in artificial intelligence (AI) is growing. The hope is that AI might help them get things done.
But before these tools are adopted, radiologists have much to consider (Listen to the PODCAST: Hear And Now: Artificial Intelligence in Radiology). Radiologists need to understand both how AI works — and how it can fail, said Bradley J. Erickson, M.D., a professor of radiology and director of the Radiology Informatics Laboratory at the Mayo Clinic in Rochester, Minn.
“We need to understand the basic principles … so (we) can understand how it might fool (us) or how it might give a spurious result,” Erickson said.
Much of the attention so far has focused on using AI in the interpretation of medical images. AI is also being groomed for other applications, for example, to choose tools for doing interpretations; or fetching and orienting images from prior exams; or accelerating the reporting process.
Whether or how far radiologists can trust these tools — particularly ones built to help interpret images — is an important issue. A simple explanation of what these tools are looking at could go a long way toward building trust, Erickson said.
“If we understand what the algorithm is looking at, and it is looking at sensible things, that gives us some confidence,” he said.
Rather than building “black boxes,” Erickson urges developers of radiological AI applications to build “gray” ones in. But regardless of how gray or transparent these boxes become, there will always be an element of faith to their use, he said.
“I’m not sure that we need to know (everything about an AI tool),” Erickson said. “We just need to know that someone knows it.”
Editor's note: in celebration of the International Day of Radiology on November 8, which will also begin ITN's extensive coverage of RSNA, contributing editor Greg Freiherr begins the coverage with this exclusive podcast and accompanying blog.