AI-Assisted Diagnosis Enters Screening Workflows: Two Mammography Studies, and a Regulator List You Check Yourself

The argument about whether AI can read scans rarely goes wrong on direction; it goes wrong on magnitude. Instead of quoting "approaching senior-expert level" from a secondary review, look at two studies whose originals you can open — one real-world implementation study, one randomised controlled trial — both in the same concrete setting: mammography screening.

Real-world study: 12 screening sites in Germany#

PRAIM is an observational, multicentre, real-world, noninferiority implementation study comparing AI-supported double reading against standard double reading without AI, among women aged 50–69 at 12 sites in Germany, where radiologists voluntarily chose whether to use the AI system. From July 2021 to February 2023, 463,094 women were screened (260,739 with AI support) by 119 radiologists (Nature Medicine paper).

The paper reports a breast cancer detection rate of 6.7 per 1,000 in the AI group, 17.6% higher than and statistically superior to the 5.7 per 1,000 in the control group (95% CI +5.7% to +30.8%); a recall rate of 37.4 per 1,000, lower than and noninferior to 38.3 per 1,000 (difference −2.5%, 95% CI −6.5% to +1.7%); a positive predictive value of recall of 17.9% versus 14.9%; and a PPV of biopsy of 64.5% versus 59.2%.

Two things have to be read alongside those numbers: the design is observational, not randomised; and the study was funded by the AI vendor, Vara, with the paper's competing-interests statement disclosing that several authors are Vara employees or customers, one of whom also took part in study design and writing.

Randomised trial: higher sensitivity, unchanged specificity#

Sweden's MASAI is a randomised, controlled, noninferiority, single-blinded, population-based screening accuracy trial in which AI triaged examinations to single or double reading and provided detection support. Between April 12, 2021 and December 7, 2022, 105,934 women were randomly assigned (19 excluded from analysis); the registration number is NCT04838756 (The Lancet paper02464-X)).

The paper reports interval cancer rates of 1.55 versus 1.76 per 1,000 participants in the intervention and control groups, a noninferior proportion ratio of 0.88 (95% CI 0.65–1.18, p=0.41); higher sensitivity in the intervention group (80.5%, 95% CI 76.4–84.2, versus 73.8%, 95% CI 68.9–78.3, p=0.031); and specificity of 98.5% in both groups (p=0.88). The paper also notes this was a protocol-defined analysis, funded by public bodies including the Swedish Cancer Society — no vendor funding here, though one author still had to declare a role such as co-founding an AI company.

Regulation: trust the agency's own list#

Don't quote second-hand summaries of regulatory status. The FDA maintains a list of AI-enabled medical devices authorised for marketing in the United States and states when its contents were last current (2026-09-04 when we checked); to verify whether a given AI diagnostic device is cleared, read that list (FDA official list). This site does not use circulating figures such as "over 50 LLM-based diagnostic tools cleared," because no first-hand source could be cited.

(Revised October 12, 2026: unsourced statements — "approaching senior-expert level", "both the FDA and China's NMPA are accelerating approvals" and the pathology examples — were removed, and replaced with the PRAIM and MASAI studies whose originals can be opened, plus the FDA's own list.)