Back to Bigger Picture

BIGGER PICTURE Earlier diagnosis

Sweden tested AI in breast screening—and found 29% more cancers

Sweden’s MASAI trial tested whether scarce specialist attention could be allocated differently. AI selected examinations for double reading and highlighted suspicious areas; doctors still made the decisions.

AI-generated illustration: A second reading layer helps examine mammography images.
A second reading layer helps examine mammography images.AI-generated illustration of the idea, not a photograph of the actual people, equipment or event.
Trial scale
About 106,000 women
Cancers detected
338 versus 262
Reading workload
44% lower

How the idea works

  1. Score images

    Estimate risk and mark suspicious areas.

  2. Allocate readers

    Risk guides single or double reading.

  3. Check outcomes

    Compare detected and subsequently diagnosed cancers.

Two readers where they matter most

The mammography appointment itself did not change. Normally, two breast radiologists read each examination. In the AI arm, higher-risk examinations received two readers while lower-risk examinations received one. Radiologists also saw suspicious areas highlighted. The intervention reorganised work around existing images rather than requiring a different scanner.

Sources: [1][3]

A result in people, not just test images

About 106,000 women were randomly allocated to the two approaches. The 2025 analysis reported 338 detected cancers with AI support versus 262 with conventional reading. Detection rates were 6.4 versus 5.0 per thousand screened women: the origin of the 29% relative increase.

Sources: [2][3]

More findings, fewer readings

Additional detections included early invasive cancers. Image-reading workload fell by 44%, while the false-positive rate was similar. That combination matters: detecting more disease by simply sending many more women for unnecessary assessment would create a different balance of benefit and burden for patients and staff.

Sources: [2][3]

The interval between appointments also counts

The 2026 follow-up counted cancers appearing between screening rounds: 82 in the AI group and 93 in the comparison group. These counts add context but do not prove longer survival. MASAI supports a particular screening workflow; reduced breast-cancer mortality was not the demonstrated endpoint.

Sources: [1]

The original sources

  1. Lund UniversityAI support in breast cancer screening: Fewer missed cancer cases
  2. The Lancet Digital HealthScreening performance and characteristics of breast cancer detected in MASAI
  3. Lund UniversityAI-supported breast cancer screening – new results suggest even higher accuracy

What to explore next

The eye check that patients no longer had to book separatelyHow to forecast a river that has no measuring station
Explore all 99 applications