Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study (ScreenTrustCAD)

Primary publication of ScreenTrustCAD (NCT04778670), a prospective, population-based, paired-reader, non-inferiority study at Capio Sankt Göran Hospital, Stockholm. Consecutive women aged 40 to 74 without breast implants in population screening were included. The primary outcome was screen-detected breast cancer within three months of mammography; the primary analysis assessed non-inferiority (margin 0.15 relative reduction in diagnoses) of double reading by one radiologist plus AI against standard double reading by two radiologists, with single reading by AI and triple reading by two radiologists plus AI as further comparisons.

From 1 April 2021 to 9 June 2022, 58,344 women underwent screening and 55,581 were included. 269 women (0.5 percent) had screen-detected cancer on an initial positive read. One radiologist plus AI was non-inferior to two radiologists (261 versus 250 detected cases; relative proportion 1.04, 95 percent CI 1.00 to 1.09). AI alone (246 versus 250; 0.98, 0.93 to 1.04) and two radiologists plus AI (269 versus 250; 1.08, 1.04 to 1.11) were also non-inferior. The authors conclude that replacing one radiologist with AI gave a 4 percent higher non-inferior detection rate and suggest controlled implementation with risk management and real-world follow-up. Funded by the Swedish Research Council, the Swedish Cancer Society, Region Stockholm and Lunit.