Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study.
Level 2 - randomized trial
Individual randomized controlled trial
PubMed 37541274 · doi:10.1016/S1470-2045(23)00298-X
What was done
Women aged 40–80 years eligible for screening mammography across four sites in Sweden were randomized 1:1 to AI-supported screen reading (intervention) or standard double reading without AI (control). In the AI arm, an automated system (Transpara v1.7.0) triaged examinations by malignancy risk score: scores 1–9 received single reading and score 10 received double reading, with AI risk scores and computer-aided detection marks available to radiologists. Participants and acquiring radiographers were masked, while reading radiologists were not. This prespecified safety analysis in the modified intention-to-treat population evaluated early screening performance (cancer detection rate, recall rate, false-positive rate, positive predictive value of recall, cancer type) and radiologist workload after 80,000 participants were enrolled.
What was found
The analysis included 80,020 women (39,996 AI-supported; 40,024 standard screening). AI-supported screening detected 244 cancers versus 203 in standard screening, yielding cancer detection rates of 6.1 per 1000 (95% CI 5.4–6.9) versus 5.1 per 1000 (95% CI 4.4–5.8), a ratio of 1.2 (95% CI 1.0–1.5; p=0.052). Recall rates were 2.2% (95% CI 2.0–2.3; 861 recalls) with AI versus 2.0% (95% CI 1.9–2.2; 817 recalls) without AI. False-positive rates were identical at 1.5% (95% CI 1.4–1.7) in both groups. PPV of recall was 28.3% (95% CI 25.3–31.5) for AI versus 24.8% (95% CI 21.9–28.0) for control. In the AI group, 184 of 244 (75%) cancers were invasive and 60 (25%) were in situ, compared to 165 of 203 (81%) invasive and 38 (19%) in situ in the control group. Total screen readings fell from 83,231 to 46,345, representing a 44.3% workload reduction.
Why it matters
This provides randomized trial evidence that integrating AI into mammography workflows can substantially alleviate radiologist workload without compromising cancer detection or increasing false positives.
Limits
This report is an interim safety analysis evaluating early detection metrics rather than the trial's primary endpoint of interval cancer rates, which requires 2-year follow-up. The study evaluated a single proprietary AI algorithm within Swedish screening centers, potentially limiting generalizability to other software or screening systems. Race and ethnicity data were not collected, and radiologists were unblinded to arm allocation.
Cited by
- supports A Swedish trial of over 100,000 women demonstrated that AI-assisted mammography screening detected 25% more cancers than radiologists alone with no increase in false positives.