Can AI replace the second reader in mammography double reading? European evidence says it can safely share the job. Sweden’s MASAI trial found 29% more cancers with 44% less reading work, and Germany’s PRAIM study found 17.6% higher detection without more recalls. The unresolved question is interval cancers, and who catches what the AI misses.
In most European breast screening programmes, every mammogram is read twice. Two radiologists, working independently, look at the same four images, and if either one is worried the case goes to discussion or recall. It is expensive, slow and deeply ingrained. It is also one of the reasons Europe’s screening programmes have a reputation for finding cancers without drowning women in false alarms.
That second pair of eyes is now the most obvious job in medicine for AI to take. The trials are large, the results are consistent, and radiology workforces are stretched. By the time you read this, several national programmes will be deciding not whether to use AI in double reading, but how.
The debate usually runs as “AI versus the second radiologist.” That framing misses the point. Double reading was never only a detection system. It was also an error-catching system with a particular property: the two readers made mistakes more or less independently. Replace one of them with software, and the errors that remain will be different in kind. The harder question is not whether AI finds more cancers. It is who, in the new workflow, is positioned to catch the cancers the AI itself will miss.
Can AI replace a second radiologist in breast screening?
The evidence says that AI plus one radiologist can match or beat two radiologists on the measures we can already see.
The strongest data come from Sweden. The MASAI trial, a randomised study run inside the national screening programme, used AI to triage exams: low-risk mammograms went to a single radiologist, high-risk ones to two, and the AI highlighted suspicious areas for the readers. The first safety analysis, published in The Lancet Oncology in 2023, showed the approach was safe to continue. The fuller screening-performance results, published in The Lancet Digital Health in February 2025 and covering more than 105,000 women, reported a 29% increase in cancer detection and a 44% reduction in screen-reading workload, without an increase in false positives.
A second Swedish study, ScreenTrustCAD, tested a different design at a Stockholm hospital. It compared standard double reading with combinations including one radiologist plus AI. Published in The Lancet Digital Health in 2023, it found that one radiologist and AI together detected 261 cancers against 250 for two radiologists, a noninferior 4% increase. Cases the AI flagged went to a consensus discussion before any recall, which kept recall rates in check.
Then came Germany. The PRAIM study, published in Nature Medicine in January 2025, was not a randomised trial but a real-world implementation across 12 screening sites. Between July 2021 and February 2023, 463,094 women were screened by 119 radiologists, 260,739 of them with AI support. Detection was 6.7 per 1,000 with AI against 5.7 without, 17.6% higher, while the recall rate was slightly lower at 37.4 per 1,000 against 38.3.
The UK is running the next test at scale. In February 2025 the government and the National Institute for Health and Care Research launched EDITH, an £11 million trial involving close to 700,000 women across 30 sites, comparing AI tools as a replacement for one of the two human readers.
Taken together, this is unusually solid evidence for clinical AI. It is also the narrow, well-defined kind of task where AI performs best, a pattern we have described in why AI diagnostic evidence is strongest where the question is narrowest.
What does double reading actually catch?
To see what might be lost, it helps to be precise about what double reading does.
The second reader catches some cancers the first reader missed. That is the obvious part. Less obvious is why it works. Two humans miss cancers for partly different reasons: fatigue at the end of a session, a lesion hidden behind dense tissue at an angle one reader happened not to study, a subtle architectural distortion one of them has seen many times before and the other has not. Because those lapses are not perfectly correlated, adding a second reader shrinks the error in a fairly predictable way.
Double reading also does work that never shows up in a detection rate. Arbitration meetings, where two readers disagree, are where a lot of informal teaching happens. Junior readers see what seniors flagged. Everyone builds volume, and volume is how mammographers get good.
AI changes each of these. It does not get tired, and it catches things humans miss, as the trials show. But its errors are systematic. If a model struggles with a certain breast density, a particular type of subtle lesion, or images from an older machine, it will struggle every time, in every screening centre running that version. Human misses scatter. Machine misses cluster.
What cancers does AI miss in mammography?
This is where the published evidence gets thinner. The trials so far report cancers detected at screening. The test that decides whether a screening strategy works is interval cancers: tumours that surface between rounds after a screen was called normal. Those take years of follow-up to count properly, and MASAI’s primary interval-cancer endpoint had not yet been reported when the screening-performance results appeared.
Two uncertainties sit inside that gap.
The first is what the extra cancers are. A higher detection rate is good news if it means catching aggressive tumours earlier. It is less good if a meaningful share are slow-growing lesions that would never have caused harm, which is the familiar problem of overdiagnosis in screening. Detection rate alone cannot separate the two.
The second is what the AI missed and the single remaining radiologist also missed. In a two-human system, the second reader is a genuinely independent check. In an AI-plus-one system, the human reads knowing the AI’s view, or reads exams the AI has already triaged as low risk. That creates a risk of automation bias: a lesion the model scored as low risk is also less likely to get a hard look from the person reading it. The trials suggest the net effect is positive. But net effects can hide a specific group of cancers that falls between both readers.
A worked example: one screening programme, two workflows
Hypothetical, with round numbers for illustration. A regional programme screens 100,000 women a year. Under double reading, that is about 200,000 human reads. Suppose it detects 5 cancers per 1,000 women, so around 500 cancers.
Switch to an AI-supported design with MASAI-like performance. Reading workload falls by roughly 44%, to around 112,000 human reads, freeing the equivalent of tens of thousands of reads for diagnostic work, assessment clinics or shorter waiting lists. Detection rises by a similar proportion to the trial, perhaps to around 640 cancers.
Now the uncomfortable part. Suppose, for illustration, that the AI systematically under-scores one subtle presentation that accounts for 2% of cancers, and that the single reader on low-risk exams, influenced by the triage, catches only half of those. That is five cancers a year that a two-human system might have found and this one misses, sitting inside a system that finds far more cancers overall. Nobody would see them in the headline numbers. They would appear years later as interval cancers, attributed to “the limits of screening.”
The example is invented, but the mechanism is not. It is the reason programmes need to audit interval cancers by the AI score the original exam received, not just in aggregate.
Isn’t this just protecting radiologists’ jobs?
The strongest counterargument goes like this. Two-reader systems also miss cancers, and nobody demanded perfection from them. The trials consistently show more cancers found and no more false alarms. Meanwhile radiology workforces in many European countries cannot keep up, and screening delays harm women too. Holding AI to a standard of “prove it never misses anything a human would catch” is a standard no human reader meets, and it risks preserving an expensive workflow out of habit.
Most of that is right. The evidence is good enough that refusing AI in double reading is becoming hard to defend, and the workload savings are real capacity that can go to assessment clinics, which is where screening programmes often lose women, as we argued in the cancer screening gap is about the second visit.
But the counterargument answers a question that is not really being asked. The concern is not that AI is worse than a human. It is that correlated errors behave differently from independent ones, and a screening programme that runs one model version across a whole country concentrates risk in a way that 200 individual radiologists never did. That is a design problem with design answers, not a reason to stop.
What this means for screening programmes, radiologists and vendors
For screening programme leaders, the practical steps are audit and independence. Record the AI score on every exam, then link interval cancers back to it. Keep some form of human double reading on a sample of AI-cleared exams, so there is a running estimate of what the machine and the single reader are jointly missing. Treat each model update as a new device that needs local checking, because systematic errors change when versions change.
For radiologists, the role shifts from primary detector toward arbiter and auditor. That is more demanding, not less. It also raises a training problem. If fewer exams are read by humans, fewer junior readers build the volume that makes them good. Programmes should protect reading time for trainees on exams the AI has already cleared, the same logic we set out in teach residents to work without the AI first.
For vendors, the bar is interval-cancer data and subgroup reporting: performance by breast density, age, machine type and ethnicity, published rather than kept in the sales deck. Programmes will increasingly write that into procurement.
For women being screened, little changes in practice. The invitation, the mammogram and any recall work the same way. The sensible habit is unchanged too: report a new lump or change to a clinician promptly, even after a normal screen.
Frequently asked questions
What is double reading in mammography?
Double reading means two radiologists independently review each screening mammogram, with disagreements resolved by discussion, a third reader or recall. It is standard in many European programmes, including in the UK, Sweden and Germany, while US practice more commonly relies on a single reader. Its aim is to catch cancers one reader misses while keeping false-positive recalls low.
Is AI mammography screening more accurate than radiologists?
In large European studies, AI combined with radiologists has matched or beaten two radiologists alone. MASAI reported 29% more cancers detected with 44% less reading workload, and PRAIM found 17.6% higher detection without more recalls. Evidence on interval cancers, the cancers found between screens, is still accumulating and is the key remaining test.
What is an interval cancer?
An interval cancer is a breast cancer diagnosed after a screening mammogram was reported as normal and before the next scheduled screen. Interval cancer rates are a key measure of how many cancers screening misses. Because they take one to three years to appear, trials of AI in screening report them later than detection rates.
Will AI mean women get fewer false-positive recalls?
Not necessarily fewer, but the studies so far show no increase. PRAIM’s recall rate was slightly lower with AI support, 37.4 versus 38.3 per 1,000, and MASAI reported more cancers found without more false positives. Recall rates depend heavily on how each programme sets thresholds and runs consensus discussion, so results can vary locally.
What is the EDITH trial?
EDITH is a UK trial launched in February 2025, funded with £11 million through the National Institute for Health and Care Research. It involves close to 700,000 women across 30 sites and tests whether AI tools can safely take the place of one of the two human readers in NHS breast screening.
The second reader is not worth defending as a job. It is worth defending as an independent check. Programmes that replace it with AI and then build a new independent check, through audit, sampling and interval-cancer linkage, will get the gains without the blind spot.
Related on the GHC blog: Why Universal Health Coverage Stalls After the Pilot Phase and Primary Care Is the Cheapest Hospital Bed, But Only If You Close One.
If your screening programme adopted AI tomorrow, who would be responsible for finding the cancers it misses?
