TrialsgrambetaRCTs are beautiful

ScreenTrustCAD

2023

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

In 55,581 women in population-based mammography screening, double reading by one radiologist plus AI was non-inferior to standard double reading by two radiologists for screen-detected breast cancer within 3 months.

Result

Efficacy: Screen-detected breast cancer within 3 months of mammography (non-inferiority margin: 0.15 relative reduction)relative proportion 1.04 (95% CI 1.00-1.09)

Below 1 favours standard double reading by two radiologists, above 1 favours double reading by one radiologist plus AI.

Event rates
Absolute difference
0.1 points

The absolute difference is derived from the event rates.

Also reported: AI alone 246 vs 250 (0.98, 95% CI 0.93–1.04) and triple reading by two radiologists plus AI 269 vs 250 (1.08, 95% CI 1.04–1.11) were also non-inferior.

Original publication
Subtopic
AI-supported double reading
Design
Prospective paired-reader non-inferiority study (non-randomised), Sweden
Relevance
Supporting: context rather than must-know.
Source
Dembrower K, et al. Lancet Digit Health. 2023;5(10):e703-e711.doi 10.1016/S2589-7500(23)00153-X

Discussion & critique

First prospective evidence that AI can replace one of two radiologists without losing cancers: the radiologist-plus-AI reading model that MASAI then tested in a randomised design. Paired-reader design means every woman was read both ways, so there is no randomisation; a single hospital, partly industry-funded, and an outcome limited to screen-detected cancer at 3 months: interval cancers and overdiagnosis are unmeasured.