Researchers at Imperial College London published a study in Nature Cancer in 2026 evaluating version 1.2 of a Google mammography AI model across one of the larger breast-screening datasets assembled for this kind of analysis: 115,973 mammograms drawn from five UK National Health Service screening services, followed for 39 months to see which flagged cases turned into confirmed cancers.
A two-part study design
The Imperial team didn’t stop at retrospective analysis. Alongside the large historical dataset, they ran a prospective feasibility deployment of the same AI model across 12 screening sites, covering 9,266 live cases as they came through routine NHS screening. Combining a large retrospective look-back with a real-time prospective rollout let the researchers check whether patterns found in historical data actually held up when the AI was reading scans as they arrived, under normal clinic conditions rather than in a controlled research environment.
What the numbers showed
On sensitivity — the ability to correctly flag cancers that were actually present — the AI model scored 0.541, compared with 0.437 for the first human reader in the standard double-reading process the NHS uses. That’s a meaningful gap in a screening context where every missed cancer can mean a delayed diagnosis. Crucially, the AI achieved that higher sensitivity without a meaningful trade-off in specificity, meaning it wasn’t simply flagging everything to catch more true cases. The practical effect showed up in the cancer detection rate, which rose from 7.54 to 9.33 cancers detected per 1,000 women screened when AI was part of the reading process. That gap carries extra weight given that breast cancer can take years to become clinically apparent, which is why the study’s 39-month follow-up window on the retrospective cohort matters as much as the raw sensitivity figures, since a shorter window risks crediting the AI with catches that would have surfaced on the next routine screening anyway.
Why comparing AI to a single reader matters
NHS breast screening normally relies on two radiologists independently reading each mammogram, with disagreements resolved by a third reader or arbitration. By benchmarking the AI against the first human reader specifically, the Imperial study speaks to a scenario screening programs are actively considering: using AI to replace or support one of the two human reads, potentially freeing up radiologist time in a system that, like many national health services, faces persistent workforce shortages in radiology.
Context from a busy year for AI mammography research
This Imperial/Google result lands amid a wave of 2026 publications testing AI-assisted mammography in different countries and health systems, including the UK’s MASAI randomized trial and a large German real-world deployment. Each uses a different AI system and study design, but the through-line is consistent: multiple independent groups, using different tools and different populations, are converging on the conclusion that AI can meet or exceed a human radiologist’s detection performance on mammograms.
What the study doesn’t settle
Higher sensitivity in a screening AI is not an unqualified win. Screening programs still need to confirm that the additional cancers detected are clinically meaningful — not indolent findings that would never have caused harm — and the prospective arm’s smaller sample of 9,266 cases limits how confidently its real-world findings generalize before wider rollout. There is also the practical question of integrating any new AI tool into radiologists’ existing workflow and liability structure, which large retrospective and prospective studies can inform but not fully resolve on their own.
What’s next
The Imperial team’s next step will likely involve tracking prospective outcomes over a longer follow-up window, mirroring the interval-cancer analysis used in randomized trials like MASAI. If the sensitivity gains hold up at scale and with longer follow-up, this study adds significant weight to the case for the NHS and other health systems to formally integrate AI into routine double-reading rather than treating it as an experimental add-on. Researchers are also likely to examine how the model performs across different demographic groups and breast density categories, since mammography AI systems have historically shown performance gaps tied to exactly those factors. How the NHS and other national screening programs choose to weigh this study alongside similar findings from Germany’s PRAIM program and the UK-based MASAI trial will likely determine how quickly AI-assisted reading becomes standard practice rather than a research pilot confined to a handful of sites.
