Artificial intelligence has shown promise in improving breast cancer screening by helping detect cancers earlier and reducing the number of dangerous cases that emerge between routine tests, according to a major Swedish clinical trial involving more than 100,000 women.
The study found that women whose mammograms were assessed with the support of AI were less likely to be diagnosed with so-called “interval cancers” — cancers that appear after a woman receives a clear screening result but before her next scheduled test. Such cancers are often more aggressive and harder to treat because they are detected at a later stage.
“Interval cancers are a very good measure of how effective the screening method is,” said lead author Kristina Lang, a breast radiologist and clinical researcher at Lund University. “Since it takes a long time to evaluate breast-cancer mortality, interval cancer rate has been used as a surrogate measure.”
The findings, published in The Lancet, come from the first randomised controlled trial to assess the impact of AI on breast cancer screening outcomes and represent the largest test of AI in cancer screening to date. Earlier analyses from the same project had already shown that AI could increase cancer detection while reducing the workload for radiologists. What remained unclear was whether those benefits translated into fewer cancers being missed and diagnosed later.
The trial was embedded in Sweden’s national breast cancer screening programme and included nearly 106,000 women. Participants were randomly assigned to standard screening — in which mammograms are read independently by two radiologists — or to AI-supported screening. In the AI group, software analysed images alongside radiologists, flagging higher-risk scans for extra attention while allowing lower-risk scans to be read once instead of twice.
Over two years of follow-up, the rate of interval cancers was about 12 per cent lower among women screened with AI support. These women were also less likely to be diagnosed later with invasive tumours or aggressive cancer subtypes.
Importantly, the use of AI did not increase false alarms. False-positive rates were similar in both groups, suggesting that screening accuracy was not compromised.
“Our study does not support replacing health-care professionals with AI,” said first author Jessie Gommers, a PhD candidate at Radboud University Medical Centre in the Netherlands. “However, our results potentially justify using AI to ease the substantial pressure on radiologists’ workloads.”
Whether AI-supported screening is cost-effective remains uncertain. While earlier modelling from Norway suggested even small reductions in missed cancers could justify the expense, a full economic analysis of the Swedish data is still under way.
Researchers cautioned that the study was conducted within a single national system, using one AI tool and one type of mammography machine. Further studies in other countries will be needed to confirm the findings and assess long-term impacts, including whether earlier detection reduces the need for aggressive treatment.
