Elías-Cabot AI triage

AI-based triage and decision support in mammography and digital tomosynthesis for breast cancer screening: a paired, noninferiority trial

Patient / Population Intervention / Exposure Comparison Outcome

In 31,301 women undergoing routine screening mammography or digital breast tomosynthesis, partially autonomous AI triage that treated AI-classified low-risk examinations as normal and double-read the rest with AI support reduced radiologist screen-reading workload compared with standard double-blind reading, with higher cancer detection but a recall rate that failed non-inferiority.

31,301

Design

Prospective paired non-inferiority study (non-randomised)

Endpoint

Radiologist workload, cancer detection rate and recall rate (co-primary)

Relevance

ResultWorkload −63.6%; cancer detection 6.3 → 7.3 per 1,000 (+15.2%, 95% CI 6.6–24.4%; P<0.001); recall rate +14.8% (95% CI 9.0–20.6%), not non-inferior. Workload reduction similar for digital mammography (−62.1%) and tomosynthesis (−65.5%); detection and recall rose only with digital mammography.

Elías-Cabot E, et al. Nat Med. 2026;32(4):1296-1305. 10.1038/s41591-026-04277-x

Discussion & critique

First prospective test of letting AI sign off low-risk screens without any human reading: nearly two-thirds of the workload disappeared and detection rose, but recalls rose too — the trade-off any autonomous AI programme must justify. Paired, non-randomised design (both strategies applied to every examination), a single AI workflow, and screen-detected cancer rather than interval cancer or mortality as the outcome.