Google’s AI Takes on Breast Cancer Screening: What the New NHS Studies Really Show

Google Research just dropped two studies in Nature Cancer about using AI for breast cancer screening in the UK’s NHS. The headlines are predictable — “AI improves detection,” “reduces workload” — but let’s dig into what actually happened, because the devil is in the workflow details.

The Problem: Double Reading Is Dying

The NHS Breast Screening Programme relies on a double-read system: two human readers look at every mammogram, and if they disagree, a third person arbitrates. It’s thorough, but it’s also labor-intensive. There’s already a 30% shortfall of clinical radiologists, projected to hit 40% by 2028. Something has to give.

Study 1: Can AI Stand Alone?

The first study had two phases. Phase 1 was retrospective: they fed 115,973 mammograms from five NHS screening services into Google’s AI system. The ground truth wasn’t just what the original readers said — they used a 39-month follow-up window to catch interval cancers and next-round cancers that the original readers missed. That’s a rigorous standard, and it’s higher than I expected for a retrospective study.

The AI’s performance was compared against the first human reader’s sensitivity and specificity. They also checked lesion-level localization — making sure the AI was pointing at the right spot in the breast, not just correlating with some spurious pixel pattern. This matters because earlier AI systems have been caught cheating by looking at things like patient ID tags or imaging artifacts.

Phase 2 was prospective but non-interventional: they deployed the live AI system into real clinical workflows without actually using it for decisions. The goal was to see if the technical integration worked — could the AI process images fast enough, could it handle the NHS’s varied IT setups, did it break anything? This is the boring but essential stuff that most AI papers gloss over.

Study 2: AI as a Second Reader

The second study was a reader study — the kind where you take a bunch of cases, have radiologists read them with and without AI, and compare. They simulated the full double-read workflow but replaced one of the human readers with the AI system. The comparison was against the original human double-read plus arbitration.

The results showed that AI as a second reader maintained or improved cancer detection while reducing the number of cases that needed arbitration. That’s the key metric: arbitration is where the workflow bogs down. If AI can reduce the disagreement rate, you free up senior radiologists for actual complex cases.

The Catch

Both studies have caveats. The retrospective phase used historical data, so there’s always the risk that the AI was optimized for that specific dataset. The prospective phase only tested technical feasibility, not clinical impact. The reader study was controlled, not a real-world deployment where radiologists might behave differently when they know the AI is watching.

Also, the AI operating points — the thresholds for flagging a case — had to be adjusted separately for each screening service. That’s not a bug, it’s a feature, but it means deployment isn’t plug-and-play. Each site needs calibration, which adds complexity.

What This Means

This isn’t the first time Google has published on mammography AI, and it won’t be the last. The real test will be a prospective interventional study — actually letting the AI influence real patient decisions and tracking outcomes. That’s still pending.

But the direction is clear: double-read workflows are unsustainable, and AI is the most viable replacement for one of those readers. The question isn’t whether AI will be used, but how quickly and carefully we can validate it for real-world use.

For now, these studies add solid evidence that the approach is technically sound. The next step is proving it saves lives in practice. I’ll believe it when I see the prospective data, but this is further along than most medical AI projects I’ve tracked.