← All case files
On the recordLancet RCT · Jan 2026

AI triage, human read

Sweden’s breast-screening trial: AI sorts the scans, radiologists make the call

MASAI screening trial · Public healthcare · Sweden

In a randomised trial of about 106,000 women, AI decided which mammograms needed one radiologist and which needed two, and flagged suspicious areas. Cancer detection rose 29%, screen-reading workload fell 44% and — the result that mattered most — 12% fewer cancers surfaced between screenings.

+29%

cancers detected v standard double reading

−44%

radiologists’ screen-reading workload

−12%

interval cancers (82 v 93) at two years

AI routes and flags; a radiologist still reads every scan.

More cancers found, fewer missed, with less reading.

The record

  • MASAI randomised about 106,000 women aged 40–74 in Sweden between April 2021 and December 2022 either to AI-supported screening — AI triaging each exam to single or double reading and marking suspicious findings — or to standard double reading by two radiologists. (Lund University, Jan 2026)
  • At the interim analysis the AI arm detected 338 cancers against 262 under standard reading, a rate of 6.4 versus 5.0 per 1,000, with only seven additional false positives. (Healio, on The Lancet Digital Health, Feb 2025)
  • Final results in The Lancet showed 82 interval cancers in the AI arm against 93 with standard screening — 12% fewer — including 27% fewer aggressive subtypes, alongside a 44% reduction in radiologists’ screen-reading workload. (Lund University, Jan 2026)

The workload problem behind the trial

European breast screening typically relies on double reading: two radiologists examine every mammogram, a safeguard that consumes scarce specialist time even though the great majority of screens are normal. The MASAI trial, led by Lund University within Sweden’s public screening programme, asked whether AI could direct that expertise — sending lower-risk exams to one reader and higher-risk exams to two — without losing safety (Lund University, Jan 2026).

The design kept humans in the clinical decisions. The AI scored each exam for risk and marked suspicious findings as detection support; radiologists still read every scan. And unlike most corporate AI claims, the question was settled by randomisation: about 106,000 women aged 40 to 74 were assigned to one pathway or the other between 2021 and 2022, and all completed two years of follow-up by December 2025.

More found, fewer missed, less reading

The interim analysis, published in The Lancet Digital Health, found 338 cancers in the AI arm against 262 under standard reading — a detection rate of 6.4 versus 5.0 per 1,000 — with only seven more false positives (Healio, Feb 2025). Higher detection alone could have meant overdiagnosis of harmless lesions, which is why the trial’s decisive measure came later: cancers that surfaced between screening rounds.

The final results, published in The Lancet in January 2026, answered that. Interval cancers numbered 82 in the AI arm against 93, 12% fewer, with 16% fewer invasive and 27% fewer aggressive subtypes, while radiologists’ screen-reading workload fell by 44% (Lund University, Jan 2026). “Interval cancers are typically more aggressive and should be as low as possible,” said the study’s lead, Kristina Lång.

Evidence first, rollout second

What separates MASAI from most AI deployments is sequencing. A public system measured before it scaled, against a safety endpoint that could have embarrassed the technology. The integration itself is mundane — “in principle it is merely to integrate an AI software into existing IT systems,” Lång said (Lund University, Jan 2026). The discipline lies in the evaluation, and in being prepared to publish a result either way.

The limits are equally clear. The trial ran in one country, with one AI system, one type of mammography machine and radiologists of moderate to high experience; it did not record ethnicity and did not assess long-term outcomes or cost-effectiveness, as an independent expert noted (Science Media Centre Spain, Jan 2026). Fewer interval cancers is a strong signal, not yet proof of lives saved.

The lesson

It worked because AI was used to allocate scarce expert attention rather than to replace the expert, and because the claim was tested in a randomised trial with a safety endpoint before anyone scaled it. Public institutions can demand that standard of evidence; most private deployments never attempt it.

How we’re reading this

This is the strongest evidence in the section, but it is one trial in one Swedish screening programme, using one vendor’s software and one mammography platform, read by experienced radiologists. Two-year interval cancers are a proxy: mortality and cost-effectiveness were not measured, and ethnicity was not recorded. Some coverage of the results is vendor-promoted. Results may not transfer to systems that already use single reading or have fewer specialists.

Sources

  1. 01
  2. 02
  3. 03
The pattern, anonymizedGood automation, bad automation

Compiled from public filings, court records, company statements and reputable press. Figures are attributed to their sources; allegations are labeled as such. Not legal or investment advice.