European mammography trials are beginning to show what artificial intelligence changes in routine screening. The next decisions concern missed cancers, reading capacity and the patients who need further investigation.
The most consequential result from an imaging algorithm may arrive months after the scan. A programme can report more cancers at screening and faster image interpretation, yet still need to establish what happens to people whose examinations were considered normal. Do fewer cancers emerge before their next appointment? Does the service find disease that matters clinically, without creating unnecessary investigations?
Europe now has unusually substantial evidence with which to address those questions. Swedish randomised research and a large German implementation study have moved artificial intelligence in mammography beyond competitions on stored images. They also reveal why the phrase AI screening describes several different interventions. Software can direct examinations to one or two readers, mark suspicious regions or provide a safety check after an initial judgement.
For hospitals and technology developers, selecting the right intervention is becoming as important as selecting the algorithm. Each arrangement changes who examines a scan, when disagreements are resolved and how much work reaches assessment clinics. A benefit in one part of that pathway can create new demands elsewhere.
Following the cancers that screening misses
The Swedish MASAI trial offers a particularly important advance. Its 2026 report followed interval cancers: cancers diagnosed after a negative screening examination and before the next scheduled round. Among approximately 106,000 randomised participants, there were 82 interval cancers in the AI supported group and 93 with standard double reading. The corresponding rates were 1.55 and 1.76 per 1,000 women. The trial met its criterion for noninferiority; the numerically lower rate should not be presented as proof of superiority.
Sensitivity was 80.5% with AI support and 73.8% under standard screening, while specificity was 98.5% in both groups. Sensitivity describes the proportion of cancers detected; specificity concerns correctly classifying women without cancer. These measures answer different questions from a laboratory estimate of how accurately software separates selected images.
The findings strengthen the case for the tested workflow. They do not establish a reduction in breast cancer mortality. Longer term outcomes, including the consequences of detecting additional disease, remain relevant when screening programmes decide how much confidence to place in earlier diagnosis.
Releasing reading capacity without removing judgement
An earlier report from the same MASAI population found cancer detection rates of 6.4 per 1,000 women with AI supported screening and 5.0 with standard double reading, a relative increase of 29%. These are successive analyses of one trial, rather than independent replications.
The intervention combined risk based allocation to single or double reading with detection support for radiologists. It reduced screen reading workload by approximately 44%. That figure concerns reading examinations, not all labour across a breast screening service. It should not be converted directly into a percentage reduction in staffing or total costs.
The operational attraction is nevertheless clear. A service can consider directing specialist attention towards examinations that most need it. The corresponding management question is where released capacity goes: faster reporting, additional screening activity, complex assessment or resilience during staff absence. Those choices will shape the practical value of deployment.
They also require measurements outside the image viewer. A purchaser would want to follow time to assessment and biopsy alongside reader activity. Otherwise, a faster initial interpretation could leave the patient waiting at the next stage, while the apparent productivity gain remains confined to one departmental metric.
Germany tests implementation at scale
The German PRAIM study adds evidence from a different setting. Published in Nature Medicine in 2025, it evaluated 461,818 women at 12 screening sites, with 119 radiologists. AI supported examinations had a cancer detection rate of 6.7 per 1,000 women, compared with 5.7 without AI support: a relative difference of 17.6%. Recall rates were 37.4 and 38.3 per 1,000 respectively.
This was a prospective observational study. Radiologists could choose whether to use the AI supported viewer, so the comparison was not randomised. Statistical adjustment helps address measured differences between groups but cannot guarantee that every relevant difference has been removed.
Its value lies partly in demonstrating that integration across multiple sites is feasible. At the same time, the study should not be treated as a direct replication of MASAI’s allocation of examinations to single or double reading. Technologies embedded in different reading arrangements may produce different effects.
For regional screening authorities, the combined evidence supports a more specific procurement conversation: which clinical task was tested, in which workflow, and with what downstream consequences? A detection headline alone cannot answer those questions.
One reader and AI is a particular clinical design
ScreenTrustCAD examined another arrangement in Stockholm. In a prospective study of 55,581 women, examinations were assessed by two radiologists and an AI system, allowing different reader combinations to be compared. One radiologist plus AI detected 261 cancers, against 250 with two radiologists, meeting the study’s noninferiority criterion. ScreenTrustCAD prospective study
This paired design is important. It compared combinations evaluating the same screening material, rather than separately randomising women to autonomous AI and usual care. The primary cancer detection assessment concerned cancers identified within three months. It therefore answers a different question from later interval cancer follow up.
The result makes a practical alternative to conventional double reading credible enough to investigate and implement under appropriate safeguards. It does not establish that removing all human assessment would produce equivalent outcomes.
Hospitals considering a change need an explicit route for disagreement. An AI flag may require arbitration, further imaging or a decision that no action is needed. The staffing and expertise required for that route belong in the deployment plan, because they determine how a promising reader combination functions as a service.
Prostate imaging shows why the comparator matters
A different lesson emerges from PI-CAI, an international study of prostate MRI published in 2024. On a 400 case reader study subset, the AI system achieved an area under the receiver operating characteristic curve of 0.91, compared with 0.86 for 62 radiologists. This measure summarises discrimination across possible decision thresholds.
However, a separate comparison against historical multidisciplinary routine practice across 1,000 test cases did not establish the prespecified noninferiority claim. The observed specificity difference was small, but the statistical conclusion still matters. Routine decisions incorporated clinical context and consultation that a constrained reader exercise does not fully reproduce.
The study therefore supports the promise of AI interpretation without settling the design of an improved prostate diagnostic pathway. Fewer false positive image classifications in an experimental comparison are not automatically fewer unnecessary biopsies in clinical practice.
For European services, a useful next evaluation would follow the combined decisions of radiologists, urologists and patients. Relevant outcomes include clinically significant cancers detected, biopsies avoided, cancers missed and the consequences of delayed diagnosis. That moves the question from an algorithm’s ranking ability to the choices it helps clinicians make.
The human response is part of the technology
AI assistance does not have an identical effect on every reader. A Nature Medicine study involving 140 radiologists and 15 chest X ray diagnostic tasks found substantial variation in its effects. Experience, subspecialty and previous AI use did not reliably predict who would benefit. Incorrect AI predictions could adversely affect performance.
This was an experimental reader study using retrospectively collected cases, not a demonstration of patient outcomes. Its practical implication is still significant: a hospital cannot assume that installing a high performing system makes the combined clinical process uniformly better.
Implementation should make it possible to examine disagreements and learn from errors without encouraging automatic acceptance of the software. Local review can distinguish a useful second opinion from a distracting flag or an overlooked failure. Such feedback is especially valuable when practice changes after the initial evaluation.
Cancer screening is becoming a leading test of whether AI can improve an organised health service. Europe’s strongest studies now connect image analysis with reader workload and subsequent cancer diagnoses. The next gains will depend on preserving that connection as programmes expand, so that capacity, clinical decisions and patient outcomes are assessed together.
