AI mammography found more breast cancers while cutting the number of image readings

Sweden’s MASAI trial improved two aspects of breast screening together. AI-supported screening found 338 cancers, compared with 262 under standard double reading, while requiring 44.2% fewer radiologist screen readings. The screening results appeared in 2025; the interval-cancer follow-up was reported in January 2026.[1][2]

The scan stays familiar; the reading route changes

Women still underwent ordinary mammography. The intervention happened after image acquisition: the system assigned examinations to one or two radiologists according to its risk assessment and marked suspicious areas to help interpretation. The control pathway sent examinations for the usual two-reader assessment. AI therefore changed both the distribution of work and the information available to the readers.[3]

The tested software was Transpara version 1.7.0. Participants were randomly allocated at four screening sites in southwest Sweden, and 105,915 were included in the analysis.[1]

The additional detections included invasive cancers, particularly small cancers without lymph-node involvement. The researchers found no statistically significant increase in false-positive results.[1]

Following the women who initially leave without a diagnosis

An interval cancer is diagnosed after screening but before the next scheduled round, or within two years of the final scheduled screen. Counting these cases tests something an initial detection figure cannot: what emerged after the screening pathway had finished its work.[2]

MASAI recorded 82 interval cancers after AI-supported screening and 93 after standard reading. The AI group also had fewer interval cancers with unfavourable characteristics in the reported counts. These were the outcomes the investigators could examine once sufficient follow-up had accumulated.[3]

The interval-cancer rate ratio was 0.88, with a 95% confidence interval of 0.65–1.18. This met the prespecified non-inferiority criterion, which allowed a maximum 20% increase. The confidence interval also includes no difference: the observed 12% lower rate does not establish a statistically significant reduction. Sensitivity, the proportion of cancers detected at screening, was significantly higher with AI; specificity was unchanged.[2]

Germany’s second-look prompt

PRAIM, published in January 2025, examined a different workflow across 12 German screening sites. Radiologists retained double reading and could choose an AI-supported viewer. Vara MG labelled likely normal examinations and activated a second-look prompt when a radiologist initially judged an AI-suspicious examination normal. The reader could accept or reject the highlighted finding.[4]

In that observational study, adjusted cancer-detection rates were 6.7 versus 5.7 per 1,000 women, while adjusted recall rates were 37.4 versus 38.3. These were model-based estimates accounting for reader selection and AI predictions: radiologists sometimes chose the viewer according to the initial AI assessment. The study’s detailed analysis included 461,818 women after exclusions. Its findings concern assisted double reading; a separate scenario removing human reads of AI-normal examinations was simulated.[4]

A more productive use of specialist attention

Together, the studies show two ways AI can earn a place in screening: route routine examinations through fewer human reads, or ask for renewed attention at a specific point of disagreement. Those are practical changes to clinical work. The useful question for a screening service becomes which allocation of expert attention finds more meaningful disease for the reading capacity it has.

Longer follow-up can connect screening gains to treatment, advanced disease and mortality. That is the next level of evidence beyond the detection and workload improvements demonstrated here.

Report an error or suggest a correction