Google’s AI mammography system shows real promise in NHS trials

Breast cancer is the leading cause of death for women aged 35–64 in the UK. Early screening via mammography saves lives, but the system is under strain. The NHS Breast Screening Programme relies on a double-read workflow — two human readers assess each case, with an arbitration panel stepping in when needed. It’s rigorous, but a 30% shortfall of clinical radiologists (projected to hit 40% by 2028) threatens its long-term sustainability.

Enter AI. Google Research has been working on this for a while, and they just published two companion studies in Nature Cancer, done in partnership with several NHS organizations as part of the AIMS study. The papers assess different aspects of an AI-based breast cancer detection system. One evaluates standalone AI performance and integration feasibility; the other compares the existing double-read and arbitration process to one where AI acts as a second reader.

Let’s dig into what they found.

Study 1: Can AI stand on its own?

The first study had two phases. Phase 1 was a retrospective evaluation of the AI system’s standalone performance using mammograms from 125,000 women across five NHS screening services. That’s a big dataset. The services used three different clinical workflows — varying by whether the second reader was blinded to the first and how arbitration cases were selected. AI operating points were tuned per service to account for local population differences.

The primary endpoints were sensitivity and specificity compared to the historical first reader. Ground truth was established using a 39-month follow-up window, which let them catch interval and next-round cancers long before they became symptomatic. They also assessed lesion-level localization (is the AI pointing at the right spot?) and fairness. The lesion-level analysis is key — it shows the AI isn’t just picking up spurious correlations.

Phase 2 was a prospective, non-interventional deployment study. They integrated the live AI system into real clinical workflows to see what breaks. No human readers were asked to change their behavior; the AI just ran in the background. This is the kind of practical testing that often gets skipped in AI research, so I’m glad they did it.

Study 2: AI as a second reader

The second study was an end-to-end reader study. They took a subset of cases and compared the original double-read plus arbitration process to one where AI replaced the second human reader. The AI system was used to flag suspicious cases, and a single human reader reviewed those. The arbitration step still happened when needed.

Results? The AI-assisted workflow showed non-inferior cancer detection rates, with a significant reduction in the number of cases requiring arbitration. That translates to less workload for radiologists. The AI also maintained high specificity, meaning it didn’t cause a flood of false positives.

What this means in practice

These studies are promising, but they’re not a done deal. The authors themselves say additional work is needed to prove effectiveness in prospective clinical practice. That’s the right caveat. AI in healthcare has a long history of looking great in retrospective studies and then stumbling in the real world.

Still, the scale here is impressive — 125,000 women across multiple sites with different workflows. The lesion-level analysis and fairness checks are more rigorous than what most AI studies bother with. And the prospective deployment phase, even if non-interventional, surfaces real integration challenges that lab tests miss.

The 30% radiologist shortfall isn’t going away. If AI can safely reduce the double-read burden — even by replacing one of the two readers in a subset of cases — that’s a meaningful step. The key is doing it without compromising accuracy or introducing bias.

I’ll be watching for the next phase: prospective interventional trials where the AI actually changes clinical decisions. That’s where the rubber meets the road.