MILAN — Despite the much-heralded promise of artificial intelligence (AI) to help neurologists and radiologists interpret CT and MR perfusion scans in people suspected of stroke, the automated approach still falls short of pinpointing a diagnosis compared with human expertise.
"I don't think there is any doubt that automated imaging processing is the future," Mark Parsons, MD, PhD, conjoint professor at the School of Medicine and Public Health at the University of Newcastle in Australia, said here at the 5th European Stroke Organisation Conference (ESOC) 2019.
Parsons believes automation provided by AI will raise the standard interpretation of complex imaging overall. "I think the future is clinical decision support," he said, and not a direct threat to physician job security.
"Radiologists may be worried it will replace them, but I think it won't," he said.
Along with colleague Andrew Brivard, Parsons published an editorial in 2012 calling for clinicians to move beyond non-contrast computed tomography such as the Alberta Stroke Program Early CT Score (ASPECTS).
"It is time to move into the 21st century and use more advanced imaging routinely in hyper-acute stroke assessment," they write. "ASPECTS becomes somewhat superfluous as we now have rapid, validated, automated infarct core and penumbra volumes with either MRI or perfusion CT."
"We might have been ahead of our time" in recommending such automation back in 2012, Parsons said at ESOC.
Several companies are now developing AI software for detecting features of stroke, such as large vessel occlusions, he said. Current technology allows automated maps of collateral flow measurements and reveals many pathologic features of ischemia. However, there is still a ways to go before AI decision support is routine in stroke care, Parsons said.
One goal is for AI to assist physicians in navigating a large amount of complex data from imaging studies. "The stroke physician has to learn new tricks and interpret the combination of qualitative and quantitative information now available," he said.
One clinical tip Parsons shared with attendees is to review the raw maps of the penumbra produced by AutoMIStar (Apollo Medical Imaging Technology Ltd), not just the summary map, for example. "The secret is to look at the raw data," he noted. "There are subtle things you won't pick up on the summary maps."
"Some of our patients arrive 4 to 6 hours [after symptom onset]. If they have longer penumbral life expectancy [on imaging], it might be appropriate to transfer them to thrombectomy," Parsons said.
Future Enhancements
Increased access to automated imaging to support telehealth; a more accurate core volume measurement; and prediction of core growth (especially for patients transferred to thrombectomy) are on Parson's wish list for the future.
It is also important that AI provides decision support and not the final call on diagnosis, he said. It could be that the computer generates the probability of a good, bad, or moderate outcome — "I do think that is the future." Regardless, the AI systems should allow expert imaging reprocessing.
One of the drawbacks of the automated perfusion imaging is "blind trust of its output," Parsons added.
During the discussion, session comoderator Olav Jansen, MD, PhD, professor in the Department of Neuroradiology at Schleswig-Holstein University Hospital in Kiel, Germany, asked how often AI fails. "In about 50% of cases, it's perfect, in 25% it's okay, and in another 25% it's inaccurate," Parsons replied. "But it varies from software to software, and we are getting better versions."
"At the moment, these automatic analyzers of stroke imaging are one of the most used AR [augmented reality] algorithms in radiology. It is being done often, but there are still a lot of mistakes," he told Medscape Medical News when asked to comment. "You need to control the AI."
He concurred with Parsons that physicians should not fear losing their jobs to AI at this point. "At the moment, we are safe."
Jansen described the current AI technology as analogous to safe-driving software on cars. "They help you drive along the way, but you would never close your eyes," Jansen said. "It's the same with this."
Parsons has disclosed no relevant financial relationships. Jansen disclosed that his institution has a relationship with iSchemaView, makers of the RAPID neuroimaging platform.
5th European Stroke Organisation Conference (ESOC) 2019.
Presented May 23, 2019.
For more Medscape Neurology news, join us on Facebook and Twitter
Medscape Medical News © 2019
Cite this: AI Not Quite Ready to Make the Call on Stroke Imaging - Medscape - Jun 10, 2019.
Comments