NHS AI breast cancer screening works but trust is the catch

TL;DR:

• Google’s AI completed mammogram reads in an average of 17.7 minutes compared to 2.08 days for human radiologists, but when paired with a human in the NHS double-reading system, the AI-human team performed roughly equally to two humans.
• Doctors overruled the AI’s correct cancer identifications in 93 cases during arbitration, and the number of arbitration panels needed rose by up to 142% at some sites.
• The NHS faces a 30% shortfall in clinical radiologists, projected to reach 40% by 2028, making AI assistance attractive despite the implementation challenges.

The Google-NHS breast cancer AI can read mammograms roughly 170 times faster than a human radiologist. It catches interval cancers that specialists miss. By almost every metric, it performs at least as well as an expert. So why is the NHS not adopting it tomorrow?

Because the humans working alongside it do not yet trust it, and the NHS is not set up to accommodate it. That is the picture emerging from Sky News’s analysis of the studies published in Nature Cancer this week.

Speed without simplicity

The AI averaged 17.7 minutes per scan read. The first human radiologist took 2.08 days. The potential time saving is substantial, particularly given the NHS’s 30% radiologist shortfall, forecast to hit 40% by 2028.

But speed gains at one stage shifted workload to another. The NHS’s double-reading system requires two specialists to agree on every mammogram, with a third expert arbitrating disagreements. When AI replaced one human reader, arbitration demand surged: up 142% at one site and 22% at another. Doctors found it harder to interpret the AI’s reasoning than a colleague’s.

In 93 cases, arbitration panels overruled the AI when it had correctly identified cancer. The system caught real cancers that were then discounted by human reviewers unfamiliar with how the AI reached its conclusions.

Infrastructure gaps

There is a practical barrier too: most NHS radiologists still use paper scans that the AI cannot process. At sites where digital scans were available, the AI proved sensitive to equipment changes. When radiologists switched scanning machines at one site, patient recall rates doubled as the AI flooded the system with false alarms.

The researchers recommended a “phased, iterative approach” with careful calibration to each local environment. In plain terms: this technology needs specialist supervision during rollout and cannot be deployed at scale overnight.

The research makes a strong case that AI breast cancer screening can work within the NHS. The harder question is whether the NHS can reorganise itself, its equipment, its training, and its clinical culture, quickly enough to benefit from it. With the radiologist shortage worsening each year, the cost of slow adoption is measured in missed diagnoses.