At a glance
- Source
- Hernström, Josefsson, Sartor, Schmidt, Larsson, Hofvind, Andersson, Rosso, Hagberg and Lång (2025). Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study. The Lancet Digital Health, 7(3), e175–e183. doi:10.1016/S2589-7500(24)00267-X
- How it was done
- Women were assigned by chance to AI-supported screening or to the standard, in which two radiologists read every scan. In the AI group, the AI sent each scan to one or two readers by risk and marked suspicious areas. Randomised experiment. People were assigned by chance to work with or without the tool, and the two groups were compared.
- Who took part
- 105,934 women eligible for breast screening, assigned between 12 April 2021 and 7 December 2022.
- Where
- Four screening sites in southwest Sweden (Malmö, Lund, Landskrona and Trelleborg), within the national screening programme.
“The findings suggest that AI contributes to the early detection of clinically relevant breast cancer and reduces screen-reading workload without increasing false positives.”
What it found
In the AI group, an AI system (Transpara, made by ScreenPoint Medical) scored each scan. Low-risk scans went to one radiologist and high-risk scans to two. The AI also marked suspicious areas for the readers.
- More cancers found. Screening with AI found 338 cancers among 53,043 women, or 6.4 per 1,000. Standard screening found 262 among 52,872, or 5.0 per 1,000. AI-supported screening found 1.29 times as many (likely range 1.09 to 1.51).
- Which cancers. The extra cancers found were mainly small invasive cancers that had not spread to the lymph nodes, and early cancers, about half of them of a fast-growing type.
- No more false alarms. The share of women called back and the false-positive rate did not rise beyond chance.
- Less reading. Radiologists did 61,248 readings in the AI group against 109,692 in the standard group, 44.2% fewer.
When the result holds
The gain came from deciding who checks what. The AI decided how many people looked at each scan, instead of a person checking every AI result. It suits work where every item is checked now and where the risk of each item can be scored.
Limits
- This paper reports the trial’s secondary measures. The main measure, cancers found between screening rounds, is reported separately.
- Breast screening is a tightly regulated process with a fixed protocol. Office work is less standard.
- The trial was publicly funded, by the Swedish Cancer Society and the Swedish government among others. One author has advised Siemens Healthineers.
Practices that use this study
This study supports:
- Keep a person checking anything that goes to a customer or carries legal weight Before AI output leaves the company or is relied on in a decision, a named person checks it against real sources and signs it off. Make the check a step in the process, with time allowed for it. Do first · Effort: medium (weeks, a few people)
- Redesign the workflow instead of adding a tool to it Giving people an AI tool inside an unchanged process changes little. Decide again who does each step, in what order, and where a person checks. Do next · Effort: high (months, several teams)