Google’s AI takes on breast cancer screening: two new studies show real promise

Breast cancer is the leading cause of death for women aged 35–64 in the UK, and the NHS Breast Screening Programme has been a critical line of defense. It relies on a double-read workflow: two human readers assess each mammogram, and if they disagree, an arbitration panel steps in. It’s rigorous, but it’s also under strain. A 30% shortfall of clinical radiologists is projected to hit 40% by 2028. That’s not sustainable.

Google Research has been working on AI for mammography for years, and they just published two companion studies in Nature Cancer that evaluate their system at scale across multiple NHS screening services. The results are solid, but let’s be clear: this is still research, not a product ready for rollout.

Study 1: Standalone performance and integration feasibility

The first study had two phases. Phase 1 was a retrospective evaluation of the AI’s standalone performance using mammograms from 125,000 women across five NHS screening services. That’s a big dataset, and they applied a 39-month follow-up window to catch interval cancers and next-round cancers that might have been missed initially. The AI’s operating points were tuned per site to account for local differences in screening populations and workflows — a smart touch that acknowledges real-world variation.

The primary endpoints were sensitivity and specificity compared to the historical first reader. They also looked at lesion-level localization (did the AI correctly identify the abnormality in the breast?) and did fairness analyses. The lesion-level analysis is particularly important because it addresses whether the AI is picking up real patterns or just spurious correlations.

Phase 2 was a prospective, non-interventional deployment study. They integrated the live AI system into real clinical workflows without interfering with the actual screening process. This is the kind of practical testing that often gets skipped in AI research, so it’s good to see them doing it. The goal was to identify integration challenges — things like latency, data pipeline issues, and how the AI’s output fits into existing PACS systems.

Study 2: AI as a second reader in double-reading workflows

The second study was an end-to-end reader study. They compared the original double-read and arbitration process to one where the AI system acted as the second reader. This is a more realistic deployment scenario: instead of replacing humans, the AI works alongside them.

The results showed that using AI as a second reader maintained or improved cancer detection accuracy while reducing the workload on human readers. That’s the kind of win-win that makes AI adoption more palatable in a resource-constrained system. But — and this is a big but — the study was retrospective, not a prospective clinical trial. The authors are upfront about this: “additional work is needed to prove the effectiveness of this system in prospective clinical practice.”

What I think

I’ve seen a lot of AI-in-medicine papers that look great in a lab but fall apart in the real world. These studies are more grounded than most. The multi-site retrospective evaluation is solid, and the prospective feasibility study addresses real operational concerns. The focus on lesion-level localization and fairness is also a step in the right direction — too many AI papers skip these details.

That said, the 39-month follow-up window, while rigorous, introduces survivorship bias: cancers that appear after 39 months are excluded, and those patients might have different characteristics. Also, the AI operating points were tuned per site, which is good for performance but raises questions about generalizability. What happens when you deploy this system at a site that wasn’t part of the study?

The real test will be a prospective, interventional trial where the AI actually influences clinical decisions. That’s the gold standard, and it’s still ahead. Google knows this, and they’re likely already planning it. But for now, these studies are a solid step forward. They show that AI can reduce radiologist workload without sacrificing accuracy, which is exactly what the NHS needs.

One thing that bugged me: the press release language is a bit too polished. “Demonstrates its potential to enhance cancer detection accuracy” — yeah, we get it. I’d rather see more raw numbers and error bars. The actual papers in Nature Cancer probably have those details, so I’ll be digging into them.

The bottom line

Google’s mammography AI is not ready for prime time, but it’s getting closer. The evidence is building, and the approach — integrating AI as a second reader rather than a replacement — is pragmatic. The NHS’s radiologist shortage isn’t going away, and AI won’t solve it overnight. But if these studies lead to a real clinical trial, we might see something that actually helps.

For now, I’m cautiously optimistic. The methodology is sound, the scale is impressive, and the focus on real-world workflows is refreshing. But I’ve been burned before by AI promises in healthcare. Let’s see the prospective data before we pop the champagne.