AI Mammography in Sweden's MASAI Trial: What Was Measured, and What Wasn't

What is new

By healthcarewithai.com Editorial Team

Sweden's MASAI trial tested AI-supported mammogram reading in a national screening programme rather than only on archived images. Its 2026 primary-endpoint report found that the approach met the prespecified non-inferiority criterion for interval cancer rate, using a 20% relative margin. That is a defined statistical result, not proof that cancer rates were equal or that AI screening improves survival. Source: 2026 MASAI report, The Lancet.

MASAI is a completed randomised screening-accuracy trial, not a phase 1-3 drug trial; its registry lists phase as not applicable. These findings do not themselves establish regulatory authorisation in any jurisdiction. Trial record: NCT04838756.

How it works

Standard double reading means two radiologists independently assess each screening examination. MASAI used Transpara version 1.7.0, from ScreenPoint Medical, to assign examinations a malignancy-risk score from 1 to 10. Scores 1-9 went to one radiologist; score 10 went to two. The software also highlighted suspicious regions on higher-scoring examinations. Radiologists still made clinical decisions; the intervention changed image interpretation rather than how mammograms were acquired. Source: 2023 MASAI methods.

Between April 2021 and December 2022, 105,934 women were randomised at four screening sites in southwest Sweden. After exclusions, 53,043 were analysed in the AI group and 52,872 in the standard group. Readers knew their allocation; participants and the radiographers acquiring images did not. Source: 2025 report.

Two terms matter. A recall is a request to return for further assessment after a suspicious screening result, not a cancer diagnosis. An interval cancer is diagnosed between screening rounds, or within the study's specified window after the last scheduled screen, without having been detected at screening. MASAI included diagnoses within two years after that last screen. Some interval cancers were previously present but undetected; others develop subsequently. Their rate is a screening-performance measure, not proof of an overall health benefit. Source: primary-endpoint report.

How strong is the evidence

Randomisation makes MASAI stronger evidence for comparing these workflows than studies in which clinicians choose whether to use AI. Results appeared in three reports:

• 2023 clinical safety analysis, The Lancet Oncology. The prespecified interim analysis covered 80,033 participants. Detection was 6.1 versus 5.1 cancers per 1,000 with AI-supported versus standard reading. Recall rates were 2.2% versus 2.0%, and false-positive rates were 1.5% in both groups. Screen readings were reduced by 44.3%. This was an interim assessment, not the final interval-cancer result. Read the 2023 report.
• 2025 screening performance, The Lancet Digital Health. In the full analysed population, detection was 6.4 versus 5.0 per 1,000, a ratio of 1.29. Additional invasive cancers were mainly small and lymph-node negative. In-situ cancers numbered 68 versus 45; these are non-invasive cancers confined to their site of origin. About half of the additional in-situ diagnoses were high grade. There was no statistically significant increase in recall or false-positive rates, which does not prove those rates cannot increase. Screen readings fell by 44.2%. Read the 2025 report.
• 2026 primary endpoint, The Lancet. Interval cancer rates were 1.55 versus 1.76 per 1,000. The rate ratio was 0.88, with a 95% confidence interval of 0.65-1.18. Sensitivity, the proportion of cancers detected at screening, was 80.5% versus 73.8%; specificity, the proportion without cancer correctly classified as negative, was 98.5% in both groups. Read the 2026 report.

The trial met the prespecified non-inferiority criterion for interval cancer rate, using a 20% relative margin. However, the confidence interval does not establish equal rates, demonstrate a reduction, or exclude a modest increase: its upper limit allows an 18% relative increase. Fewer invasive, larger and less favourable-subtype interval cancers in the AI group were descriptive comparisons, not proof of improved long-term outcomes. Source: primary-endpoint analysis.

Who could benefit

Screening services using double reading could potentially benefit from reducing the number of screen readings by 44.2%, alongside MASAI's interval-cancer non-inferiority result. That percentage must not be translated into equivalent reductions in working time or staffing. For participants, higher detection could matter, but its balance against additional diagnoses and treatment requires longer-term assessment. Source: screening-performance report.

Germany's PRAIM study provides observational evidence of an association with higher detection in routine practice. Its reported detection rates were 6.7 versus 5.7 per 1,000 with versus without AI support; recall rates were 37.4 versus 38.3 per 1,000. Unlike MASAI, PRAIM used AI-supported double reading, not triage to single or double reading. Radiologists voluntarily chose AI use, permitting selection bias and residual confounding despite statistical adjustment. PRAIM therefore does not independently validate MASAI's workflow or establish causation. Source: PRAIM, Nature Medicine.

Limitations and next steps

MASAI tested one software version, one mammography-equipment vendor and one regional setting, with experienced readers who were not blinded. Results may differ elsewhere or after software updates. The primary report assessed one screening round and its interval-cancer follow-up, not sustained performance over repeated rounds. Source: MASAI limitations.

The findings do not establish fewer breast cancer deaths or settle overdiagnosis, detecting cancers that would never have caused harm. Higher detection alone cannot answer those questions. Implementation decisions need local validation, ongoing monitoring and evidence about longer-term benefits, harms and cost-effectiveness. Source: primary-endpoint report and discussion.

This is not medical advice; talk to your doctor.