AI triage could reduce second-reading workload by about 77% in the French breast cancer screening program, according to a retrospective study published in Radiology: Artificial Intelligence.
The study evaluated whether AI could identify negative screening mammograms that could bypass second reading in France’s national screening workflow.
Researchers analyzed 55,589 screening mammograms from 42,419 women ages 50-74. The exams were performed from January 2015 to December 2019 and were initially classified as BI-RADS 1 or 2.
Second-reading outcomes were compared with a commercial AI system using a predefined binary threshold of 5 or higher.
Second readers recalled 183 of 55,589 exams, or 0.33%. Those recalls yielded 12 cancers, with a positive predictive value of 6.6% and a cancer detection rate of 0.22 per 1,000 exams.
The AI system classified 42,606 exams, or 76.6%, as low risk. It classified 12,983 exams, or 23.3%, as non-low risk.
One cancer was detected in the AI-low group, compared with 11 cancers in the AI-non-low group.
Interval cancer rates were also higher in the AI-non-low group than in the AI-low group, at 2.16 versus 0.47 per 1,000 exams.
The authors said excluding AI-low exams from second reading could focus radiologist review on higher-risk cases. They also said prospective validation is needed.
“AI triage could potentially reduce second-reading workload by approximately 77%,” the authors wrote.
The study was authored by Christophe Tourasse, MD; Benoît Mesurolle, MD; Maud Ottavy, MD; Patricia Soler-Michel, MD; and Patrice Taourel, MD, PhD.
Sources
About the author
RadiologySignal.com writersEditorial Team
Radiology Signal Staff covers developments across medical imaging, radiology AI, imaging informatics, clinical research, and radiology business. The team monitors primary sources, peer-reviewed studies, company announcements, society updates, and healthcare industry news to deliver concise reporting for imaging professionals.
