Artificial intelligence developed by Google can detect breast cancer on screening mammograms with accuracy comparable to, and in some cases exceeding, that of human radiologists, according to a large UK study involving more than 175,000 women.
The research, published in Nature Cancer, is the largest study to date within the UK’s National Health Service (NHS) evaluating the role of AI in breast cancer screening. The project was conducted by a consortium including Imperial College London, University of Cambridge, University of Surrey and several NHS Trusts. The findings suggest that AI could help detect more cancers while reducing the workload for radiologists.
Addressing workforce pressure
Breast cancer is the most common cancer in the UK, with approximately one woman diagnosed every ten minutes. At the same time, the country faces a shortage of radiologists: current estimates indicate a 29% shortfall in clinical radiologists, nearly 2,000 specialists, with the gap expected to grow to 39% by 2029.
Screening mammograms in the UK are typically assessed by two readers, usually specialist radiologists. Each reads the scan independently, and if the assessments differ, a third reader reviews the case to make a final decision. The new research examined how this standard approach compares with a workflow combining one human reader and an AI system.
Fewer false positives
In the largest part of the study, researchers retrospectively analysed screening data from 125,000 women aged 50 to 70 who underwent mammography in five NHS screening programmes between 2015 and 2016. The final analysis included 115,973 scans.
When used as a second reader alongside a human radiologist, the AI system achieved a higher cancer detection rate than the human first reader alone. The detection rate increased from 7.54 cancers per 1,000 women for the human reader to 9.33 per 1,000 when AI was used.
The AI system also identified more invasive cancers and significantly reduced the number of false-positive findings. In addition, it detected approximately 25% of interval cancers. These are tumours that appear between scheduled screening exams. Performance improvements were particularly notable in women undergoing their first screening. In this group, AI reduced recall rates by 39.3% and increased cancer detection by 8.8%.
Reduced workload and faster analysis
The introduction of AI also reduced the number of mammogram reads required. In the retrospective study, scan reading workload decreased by 32.1%, representing a substantial reduction in radiologists’ workload.
A second study component examined 9,266 current screening cases at two screening services across 12 sites in London. While the AI system initially produced a higher recall rate than human readers, the researchers adjusted the recall thresholds during the study. Despite this, the system demonstrated significant efficiency gains. On average, AI completed a scan assessment in 17.7 minutes, compared with 2.08 days for the first human reader.
The third part of the study explored the use of AI during arbitration, the process in which a third reviewer evaluates cases where two readers disagree. This analysis included data from 50,000 women.
According to the researchers, this is the first time AI has been evaluated in such a role. The results showed that the AI system performed comparably to human arbitrators. Although the system produced a slightly higher arbitration rate, the overall screening workload still declined.
Supporting radiologists
Researchers emphasise that the technology is intended to support clinicians rather than replace them. “Early detection is our most powerful tool in the fight against breast cancer, and these findings mark a genuine turning point,” said Susan Thomas, Clinical Director at Google and co-author of the study. “This is the first time that we've been able to rigorously test doctors and AI working alongside each other in a clinical setting.”
Deborah Cunningham, consultant radiologist at Imperial College Healthcare NHS Trust, noted that AI could help address workforce shortages while improving screening efficiency. “The time saved will free up radiologists to perform more hands-on tasks such as needle biopsy, an essential part of the cancer diagnostic pathway,” she said.
The findings may also inform future research initiatives, including the upcoming EDITH trial, an international study evaluating different AI tools for mammography screening across 30 sites. According to the researchers, further development and validation of AI systems could ultimately enable earlier cancer detection while helping health systems manage growing screening demands.
Earlier study
Last year, a study by the UCLA Health Jonsson Comprehensive Cancer Centre showed that AI can help detect so-called interval breast cancers earlier. Interval cancer develops between regular screenings and is sometimes missed because the signs on a mammogram are too subtle for the human eye to detect
In the study researchers analysed nearly 185,000 mammograms from the period 2010–2019. They examined 148 cases of interval cancer. Using the AI software Transpara, subtle abnormalities in previous screening images could be identified or flagged as suspicious. According to the researchers, integrating AI into screening programmes could potentially reduce the number of interval cancers by about 30 per cent, leading to earlier diagnosis, less aggressive treatment and better outcomes for patients.
