Thursday, December 4 | 10:10 a.m.-10:20 a.m. | SSNR13-5 | Room N228
This session addresses the challenge of high false-positive rates associated with using AI to detect intracranial hemorrhage (ICH).
A research team that included experts from an American College of Radiology ARCH-AI Recognized Center studied the issue of false-positive ICH to characterize incidence, understand common causes, and mitigate the problem.
The study focused on a single commercially available AI platform used across all sites of a multi-institutional hospital network. All noncontrast head CT scans with positive AI notifications for ICH were reviewed over a 12-month period, with studies performed between 10 p.m. and 7:30 a.m. excluded.
A review of 14,707 total emergency department CT scans in 2024 found that 632 were flagged as positive for ICH by AI. Of those, 243 (38%) were false positives, corresponding to a positive predictive value of 72.2%, according to presenter Emily Mungovan and colleagues.
An AI widget and CT heat maps were used to identify the flagged regions; these showed that unusual hyperdense areas could trigger a positive notification, according to the group. At the institution, radiologists can review AI notifications and flagged imaging via a desktop widget without opening the full study. The widget may offer several benefits.
"Familiarity with common false positive sources and efficient use of this tool may enable radiologists to triage cases more effectively and expedite true positive reporting," the group wrote in the abstract.
This work is important because high false-positive rates may distract from or delay the identification of true-positive ICH. Join the session to hear specific examples of false positives flagged by AI and about the widget aid.
