AI tool boosts NHS breast cancer screening accuracy

Google researchers, working with Imperial College London and the NHS, have reported results from two studies suggesting an artificial intelligence system can detect some breast cancers missed in routine screening and reduce the workload involved in reviewing mammograms.

The work, published in Nature Cancer, focused on so-called interval cancers-those diagnosed between scheduled screening rounds, often after symptoms appear. The experimental AI system identified 25% of interval cancers missed by conventional screening, the researchers reported.

The studies also explored how AI might fit into the NHS breast-screening pathway, which relies on double reading. Two specialists review each mammogram and must agree on the outcome; an arbitration panel resolves disagreements.

Pressure on screening

The team framed the work against workforce constraints in radiology. In the NHS, each specialist reviews roughly 5,000 scans a year and has about four hours of dedicated time per week for this work. The researchers also pointed to a broader shortage of radiologists.

The first study tested AI accuracy against expert reading at scale using mammograms from 125,000 women. The AI system detected 25% of previously missed interval cancers, the researchers said.

In the same analysis, the AI system found more invasive cancers and more cancers overall than expert radiologists. It also produced fewer false positives among women attending their first screening appointment, according to the findings.

Workflow impact

A second study examined operational effects of adding AI to the reading pathway, using scans from more than 50,000 women. Used as the second reader in the double-reading workflow, the AI system could cut screening workloads by an estimated 40%, the researchers reported.

This approach keeps a human reader in the process while changing how cases are allocated. When the system agrees with the first reader, fewer mammograms require a second human review. Disagreements still follow existing escalation steps.

The researchers said the model could reduce screening backlogs while maintaining the clinical benchmark associated with double reading. They also said it could free up specialist time for more complex cases and other patient-facing work.

Dr Zubir Ahmed, Health Innovation and Safety Minister, linked the findings to earlier diagnosis and service reform.

"This research gives me real hope that we can catch more cancers earlier - giving more people the time and the treatment they need. That's what building an NHS fit for the future looks like in practice," said Dr Zubir Ahmed, Health Innovation and Safety Minister.

Human-AI decisions

Beyond accuracy and workload, the studies examined how clinicians respond when AI agrees or disagrees with their assessment. The research included a large-scale simulation of how radiologists and arbitration panels react when AI challenges or confirms a diagnosis.

Arbitration is designed to improve safety and limit false positives, but the researchers said it also revealed a tension. In the simulation, arbitration panel specialists occasionally overruled cancers flagged by the AI system; in those cases, the cancers would otherwise have remained undetected.

The results point to the importance of trust and training when new decision tools enter established pathways. The researchers said the findings show a need for continued work on human-AI interaction, focusing on specialist confidence in AI detection of subtle early-stage cancers.

Real-time feasibility

The collaboration also ran an observational feasibility study across 12 NHS screening sites in London, processing more than 9,000 cases in real time. The study did not use AI results to influence patient care.

The exercise highlighted integration challenges in clinical environments. The researchers said AI systems require calibration for specific hospitals and ongoing adjustment as workflows change, equipment evolves, and patient populations differ.

The studies build on earlier research from the same group, which examined an earlier version of the AI screening system in a single-reader setting and linked it to higher cancer detection and shorter diagnostic waiting periods.

Next steps include further evaluation of how AI-supported reading affects decision-making across clinical roles and how screening services adapt operationally when AI becomes part of routine mammography review.