AI as Second Reader in Breast Screening: Better Detection, Fewer Reads, Unproven Long-Term Benefit

The use case

Between July 2021 and February 2023, radiologists at 12 German screening sites could choose to read mammograms with an AI system as the second reader, and 260,739 of 463,094 screens were read that way³. That is the use case: in population breast screening, two radiologists normally read each mammogram and human arbitration settles disagreements⁶, and software takes the place of one of the two readers. In the Swedish MASAI trial the software did two jobs: it sent each scan to a single or a double human reading, and it highlighted suspicious findings for the radiologist¹. At least one human radiologist still reads every scan¹.

Scope note. Counted as AI here is machine-learning image-analysis software trained on large sets of mammograms, such as Transpara in MASAI² and Lunit INSIGHT MMG in a Korean accuracy study⁵. Not counted are earlier computer-aided detection tools that mark areas using fixed image-analysis rules, and the radiologists' own double reading, which is the comparison baseline throughout. AI as a stand-alone reader is out of scope: a 2025 systematic review judges the evidence not yet sufficient for that step². Breast-density measurement and risk prediction are also outside this paper.

Where it is running

Denmark and Spain are among the countries covered by the 2025 review, but the scale and dates of any programmes there could not be established from the sources opened².

What the evidence shows

Randomised trial, Sweden. The strongest evidence is MASAI, reported in February 2026 through a news summary of the Lancet paper; the paper itself could not be opened¹. Interval cancers, which are cancers diagnosed between screening rounds and so a measure of what a round missed, occurred at 1.55 per 1,000 women in the AI-supported arm and 1.76 per 1,000 in the control arm (measured in a randomised trial; 95% CI 1.23 to 1.92 and 1.42 to 2.15)¹. The reported proportion ratio was 0.88 (95% CI 0.65 to 1.18)¹. With an interval that includes 1, the result does not show that AI-supported screening reduces interval cancers. Sensitivity, the share of cancers that screening found, was higher with AI support at 80.5% against 73.8% (measured in a randomised trial, P = .031), while specificity was reported as the same in both arms at 98.5%¹. That pattern is consistent with the AI helping to find cancers without adding false alarms. Sensitivity was higher across age and density groups but not for in situ cancer¹. A review published in December 2025 quotes MASAI cancer detection of 6.4 per 1,000 with AI support against 5.0 per 1,000 (p = 0.0021, randomised trial), with stable or lower false-positive and recall rates².

Real-world implementation, Germany. In the PRAIM study, cancer detection was 6.7 per 1,000 with AI support against 5.7 per 1,000 for standard double reading, a relative difference of 17.6% (95% CI +5.7% to +30.8%)³. Recalls were 37.4 against 38.3 per 1,000 (-2.5%, 95% CI -6.5% to +1.7%), so the recall rate did not rise³. The positive predictive value, the share of recalled women found to have cancer, was 17.9% against 14.9%, and for biopsies 64.5% against 59.2%³. All of these were measured without randomisation between July 2021 and February 2023 and are taken from the abstract³.

Workload. Reading time is saved because scans the software rates as low risk need one reader instead of two, so the saving depends on where the threshold is set. An earlier MASAI interim analysis is reported to have cut radiologist workload by 44% (randomised trial, date not shown)¹. A Norwegian study modelling 2024 BreastScreen Norway data estimated that AI as one of two readers would cut screen-reading from 6.5 to 3.3 man-years, a 50% reduction, or from 9% to 4.5% of radiologists' total workload; it assumed reading time per scan would be unchanged with AI and used 180 working days per year, so it is a model and not a measurement⁴. The study describes the effect on overall radiologist workload as moderate⁴. The 2025 review reports triage set-ups reducing reading volumes by 40% to 90% while keeping non-inferior detection when thresholds were set conservatively (mixed study designs, published December 2025)².

Overall assessment. The review covered 31 studies and more than two million screening examinations, searched to June 2025, and concluded that AI performance is broadly comparable to radiologists and can cut reading workload as a second reader or triage tool, with certainty limited by varied designs, enriched datasets and incomplete interval-cancer follow-up².

What it takes: cost, integration, governance

Cost. No public price per examination was found in the sources opened, and none of them reports a cost-effectiveness analysis. The cost question is open.

Integration. In the EDITH trial the AI reads mammograms in a cloud environment hosted by CIMAR (UK) Ltd, which holds the data for a maximum of 30 days⁷.

People and process. In PRAIM, radiologists voluntarily chose whether to use the AI³. In EDITH, final decisions about cancer detection are made by human experts⁶⁷. Whether time is actually released depends on routing rules and on radiologists' reading habits; the Norwegian estimate of a moderate overall effect on workload shows the limit⁴.

Governance and rules. EDITH lets women opt out of AI use, although if they opt out after screening, AI may already have been used⁷. Its results are to be shared with the UK Government and screening services to inform decisions on adopting AI in national programmes⁷. The 2025 review calls for transparent reporting, standardised evaluation and long-term population studies before AI is considered as a stand-alone reader². The official texts of medical-device and AI regulation were not read in this run, so no legal requirement is described here.

Where it falls short

Design of the main trial. MASAI ran in one country, was single-blinded, and tested one product, Transpara v1.7¹². Public funding (Swedish Cancer Society, Confederation of Regional Cancer Centres and Swedish governmental funding) supports its independence, but vendor involvement could not be checked¹. The news summary lists no limitations¹.

Interval cancers and outcomes. The central safety measure, interval cancers, favours AI only as a point estimate¹. None of the sources opened reports deaths averted or overdiagnosis, so the extra cancers found cannot yet be read as lives saved. The review adds that incomplete interval-cancer follow-up affects several major studies².

Selection in the German data. AI use in PRAIM was voluntary and not randomised³. Radiologists who adopt a new tool may differ from those who do not, so the 17.6% higher detection is an association in routine practice and not a measured effect of the AI alone. The abstract does not report interval cancers³.

Misses and uneven performance. In a retrospective study of 1,097 invasive cancers diagnosed at two Korean university hospitals between January 2014 and December 2020, Lunit INSIGHT MMG version 1.1.7.3, run alone, missed 14% (154 of 1,097)⁵. The miss rate was 9% (36 of 398) in HER2-enriched, 17.2% (106 of 616) in luminal and 14.5% (12 of 83) in triple-negative cancers; 56 of the 154 misses were in dense breasts, and 61.7% (95 of 154) were judged actionable in blinded review⁵. The cases are older than three years and one software version was tested, so the figures are a warning about uneven accuracy and not an estimate for current products. In MASAI, sensitivity was not higher for in situ cancer¹. These patterns are why a human reader stays in the loop.

Dependence on people and process. The workload savings in the review applied when thresholds were set conservatively², which means part of the result comes from how the service is run and not from the software.

Failures and withdrawals. Dedicated searches for pauses, withdrawals, lawsuits and regulator action on AI breast screening found accuracy critiques such as the Korean study, but no programme withdrawn or paused and no enforcement action. Few official regulator pages could be opened, so this is not proof that none exists.

Answer to the challenge question. The use case looks less successful once three things are separated: detection gains seen in one trial and one observational study, interval-cancer evidence that does not yet show a reduction, and no data on deaths or overdiagnosis.

Phase: Pilot with results — why

The hypothesis is kept. A randomised trial and a large real-world study show results, but they cover one country each, EDITH has not reported, and national programme adoption was not confirmed in the sources opened. Where the scale of confirmed machine-learning use cannot be established, the lower plausible phase applies.

What to watch next

• EDITH results: no reporting date is given on the trial pages, but the trial will publish detection and recall outcomes in a journal and share them with the UK Government and screening services⁶⁷.
• Interval-cancer follow-up from implementation studies and long-term MASAI follow-up: the observable trigger is publication of interval-cancer or later-round data, the gap named in the 2025 review².
• Any national screening decision on adopting AI that cites EDITH, which the trial is designed to inform⁷.

Questions to ask before adopting

• Which AI product and version would be used, and has it been tested on women and equipment like ours, including dense breasts and each cancer subtype?
• Will the AI replace a second reader or add a third look, and who sets and reviews the threshold that sends scans to a single reading?
• What do our own audits of interval cancers, recalls and biopsy outcomes show after the first 12 months, and who acts on them?
• What is the full cost per examination, including integration, data hosting, training and monitoring, against the radiologist time actually released?
• How will women be informed and offered an opt-out, given that the evidence is on detection and not yet on deaths?