Can Generative AI Reinvent Radiology Reporting?: An Interview with Samir Abboud, MD

While some have predicted that artificial intelligence (AI) may replace radiologists, Samir Abboud, M.D., says generative AI will allow radiologists to maximize their time on image interpretation.

In a new study, recently published in JAMA Network Open, Dr. Abboud and colleagues evaluated the impact of a generative AI model for providing initial drafts of radiology reports for radiographs. In a total cohort including 23,960 radiographs, the researchers found that use of the generative AI model reduced documentation time for radiology reports by a mean of 29.4 seconds.

While the study noted a 15.5 percent increase in documentation efficiency, Dr. Abboud noted in a recent interview that he has seen “truly eye-popping” efficiency gains of up to 40 to 50 percent with use of the generative AI model for initial draft reporting.

“ … Once you've decided and formed your conclusions about what's going on with that patient, you look over the report that's been generated and decide if you agree or not, and then make any corrections that are necessary. It's just a much more natural way of doing things to my mind, and I think that's the only solution you have to the workforce shortage we have in radiology,” asserted Dr. Abboud, the chief of emergency radiology at Northwestern Memorial Hospital in Chicago.

In a total of 97,651 studies, the generative AI model provided a 72.7 percent sensitivity and a 99.9 percent specificity for unexpected pneumothorax cases that were clinically significant, according to the study authors.

Dr. Abboud attributes the specificity to additional filters incorporated into the Northwestern generative AI model that provide greater context than other AI software models for differentiating between cases of expected pneumothorax and incidental detection with clinical significance. This difference is a key advantage in a large hospital system with many ICU and postoperative patients, according to Dr. Abboud.

“When we ran the trial … the number of flags I was getting a week for unexpected pneumothoraces was one or two as opposed to getting 30 or 40 flags a day,” recalled Dr. Abboud, a clinical assistant professor of radiology at the Northwestern University Feinberg School of Medicine. “You knew right off the bat that if you were getting a flag, it was a real one, not just a stack of flags to then sort through further. The clinical relevance and the net benefit to our radiologists from that sort of approach has been quite a bit more profound than using the more traditional methods.”

(Editor’s note: For related content, see “Generative AI Model Reduces CXR Reading Time by 42 Percent,” “Study: AI Bolsters Sensitivity for Pneumothorax on CXR and Significantly Reduces Reporting Time” and “Can Innovations with AI Help Address the Impact of Staffing Shortages on Radiology Workflows?”)

For more insights from Dr. Abboud, watch the video below.