Mammography screening saves lives, but it has a stubborn blind spot: interval cancers, the ones that surface between screenings after being missed on a scan that looked clear. A major NHS study published on March 10, 2026, in two linked Nature Cancer papers suggests AI can shrink that blind spot — catching a meaningful share of the cancers conventional reading let through.
What the study tested
Researchers, working with Imperial College London and the NHS, compared the standard practice of two human radiologists reading each mammogram against one human reader paired with an AI reader using software developed by Google. The question was whether the AI could match or beat a second pair of trained human eyes.
The results
The cancer detection rate rose from 7.54 to 9.33 per 1,000 women screened, and the AI flagged 25% of the interval cancers that had previously been missed. Gains were strongest on first screens — 8.8% higher detection with 39.3% fewer unnecessary recalls — and on invasive cancers, the ones that matter most to catch early.
The workload bonus
There’s a practical dividend beyond accuracy. By taking on one of the two reads, AI could cut radiologist workload by roughly 40%, freeing scarce specialists to focus on complex and ambiguous cases. Amid a screening-staffing crunch, that’s not a side benefit — it’s part of the point.
The caveat
This is decision support, not autonomy. A human radiologist remained in the loop in the study design, and AI can still produce false positives and false negatives. Rolling it out fairly across diverse populations — and confirming it performs equally well for everyone — is exactly the work the NHS evaluation is built to test before any broad deployment.
Why it matters
An AI that catches a quarter of previously missed cancers and halves the reading burden hits the rare sweet spot of better outcomes and lower cost. For a national screening program stretched thin, this is one of the most concrete real-world demonstrations yet that AI can quietly make standard care better — measured not in benchmark scores, but in tumors found in time.
