A first-of-its-kind trial demonstrates that AI-assisted mammography can improve the outcomes of patients with breast cancer, particularly those with aggressive disease.
While many people have only recently begun to use artificial intelligence (AI) in their everyday lives, the technology’s use in medicine began about a decade ago, especially in the field of image-based diagnostics. Researchers have been training AI programs to recognize tumors and other signs of disease in various medical imagery, such as X-rays, MRIs, and tissue biopsies mounted on slides.
However, to know if an AI tool can really diagnose cancer and make a difference to patients, you need to have a “prospective” study — one in which patients who are diagnosed using the AI tool are then followed for several years to determine their health outcomes.
Now, researchers in Sweden have conducted a gold-standard trial to assess the use of AI in mammography screening. Results from the Mammography Screening with Artificial Intelligence (MASAI) trial, published Jan. 31 in the journal The Lancet, showed that mammography reading supported by AI can improve screening performance while reducing radiologists’ workload.
This is the first time AI has been shown to improve the outcomes of patients with breast cancer.
Spotting cancer earlier
The practice of regularly screening patients has significantly reduced the incidence of late-stage cancer and breast cancer deaths in much of the world. But even with regular mammograms, some cancer may go undetected.
These “interval cancers” are not detected at an initial screening but get diagnosed within the next two years, or between two screening rounds. They are often missed because they are masked during the initial screen due to breast-tissue density or the tumor disguising itself as normal tissue. Or sometimes, they can develop very quickly between screening dates.
These cancers are invasive, spreading into nearby healthy tissues, and typically aggressive, resulting in worse patient outcomes. Declines in interval cancer rates are the best way to confirm that a screening method works, meaning it drives down late-stage cancer diagnoses by spotting more cases earlier.
“If you want to improve the efficacy of screening, then the interval cancer rate is a very good surrogate measure of breast cancer mortality,” senior study author Dr. Kristina Lång, a breast radiologist and clinical researcher at Lund University in Sweden, told Live Science. “So if we can lower the interval cancers, it will likely have a positive impact on patient outcomes.”
The MASAI trial included more than 100,000 women between the ages of 40 and 80 living in Sweden. It used a commercially available AI system that was trained on more than 200,000 examinations from medical institutions all over the world.
In a comparison group, mammograms were read by two radiologists, as is the standard in Sweden. In the AI-assisted group, the AI system analyzed mammograms for suspicious findings and provided a risk score of 1 to 10. Cases with a score of 1 to 9 were subsequently read by a single radiologist, while a score of 10 would be read by two radiologists. The AI system was also able to highlight the suspicious findings within the image so the human radiologists could easily review them.
The AI-supported screening identified more clinically relevant cancers than unassisted mammography did. “Clinically relevant” cancers are those that have the potential to progress and thus require medical intervention.
It also reduced the number of interval cancer diagnoses within the two years following the screen. This shows that the AI program was more effective at identifying cancers that might normally be missed by a human radiologist, allowing medical treatments to start earlier.
Reducing false positives
While cancer screening is mostly beneficial, there are some potential downsides, such as false positives and overdiagnosis. When a patient is called back for a recheck after a screening but does not have cancer, “that can be a really stressful experience,” Lång said.
The latter situation, overdiagnosis, refers to situations where a screen detects a cancer that will ultimately cause no harm to the patient. Such cancers grow so slowly that they won’t cause symptoms within a patient’s lifetime or increase the chance of death. Overdiagnosis can subject healthy patients to unnecessary cancer treatments.
The goal of AI-assisted mammography is to improve the ability of the screening test to find cancer while mitigating these potential negative effects — and the study found that AI-assisted screening did not increase the risk of false positives and that it improved the detection of clinically relevant cancers.
Along with improving cancer detection, AI-assisted screenings could address the consistent shortage of radiologists available to provide cancer screening.
“In some places, you’re lucky to find one radiologist to read the mammograms,” said Dr. Richard Wahl, a radiation oncologist at Washington University in St. Louis who was not involved in the study. “If you don’t have the expert radiologists, women can’t benefit like they should from screening programs.”
Additionally, as the few radiologists available work more hours, their performance decreases. But AI doesn’t get tired, and its performance doesn’t decline at the end of the workday.
“The workforce issue is real, and this [study] could have an impact,” Wahl said. “I think people will gradually be interested in having AI-aided interpretation as a second set of eyes.”
Lång and her team will be starting a screening trial in Ethiopia in March, during which they will use AI to support the rapid assessment of breast cancer using bedside ultrasounds within a screening program.
“The problem in these settings where they don’t have a screening program is that many women come in with late-stage disease, and there are no radiologists there,” Lång said. With AI support, Lång hopes to improve access to accurate screening and thus enable earlier diagnosis of breast cancer in these limited resource settings.
This article is for informational purposes only and is not meant to offer medical advice.
Gommers, J., et al. (2026). Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the Masai study: A randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial. The Lancet, 407(10527), 505–514. https://doi.org/10.1016/s0140-6736(25)02464-x
