AI to Help Read Mammograms: What Trials Show — and What It Doesn't Replace

What the headlines say, and what was tested

Coverage of artificial intelligence in breast screening tends to arrive in one of two flavors: machines are about to replace radiologists, or machines cannot be trusted with something this important.

The largest randomized evidence so far says neither. It comes from a Swedish trial called MASAI, and it tested something narrower and more useful than either headline: whether AI can help radiologists read screening mammograms.

How the trial worked

MASAI recruited women eligible for screening at four sites in Sweden and randomly assigned them, one to one, either to AI-supported screening or to standard double reading without AI.

Standard double reading means two radiologists independently read every mammogram. It is the routine in much of Europe and it is thorough and expensive.

In the AI arm, the software did two jobs. It scored each examination for risk, which was used to decide whether it needed one reader or two. And it marked suspicious areas on the images as a prompt for the human reader.

Between April 2021 and December 2022, 105,934 women were randomly assigned. Nineteen were excluded from analysis. Median age was about 53.

What it found

The published results cover three questions.

Cancer detection. In the AI-supported group, 338 cancers were found among 53,043 participants. In the standard group, 262 were found among 52,872. That works out to 6.4 cancers per 1,000 screened against 5.0 per 1,000 — a ratio of 1.29, with a 95% confidence interval of 1.09 to 1.51 and p=0.0021.

The extra cancers were mostly the kind you want to find early. There were 58 more small T1 tumors, 46 more with no lymph node involvement, and an increase in high-grade in situ disease with no increase in the lowest-grade in situ disease.

False alarms. Recalls rose slightly, but not significantly: a ratio of 1.08 with p=0.084. The false-positive rate was essentially unchanged, at a ratio of 1.01 with p=0.92. The positive predictive value of a recall — how often being called back actually turned out to be cancer — improved, with a ratio of 1.19 and p=0.012.

Workload. There were 61,248 screen readings in the AI arm and 109,692 in the control arm. That is a 44.2% reduction in reading workload.

The result that mattered most

Finding more cancers at screening is not automatically good. It could simply mean finding more harmless things. The test of a screening change is what happens to interval cancers — cancers that appear between screening rounds, having been missed or having grown fast.

The MASAI interval cancer analysis reported rates of 1.55 per 1,000 in the AI arm and 1.76 per 1,000 in the control arm. The ratio, 0.88 with a 95% confidence interval of 0.65 to 1.18, met the trial's non-inferiority standard.

Sensitivity — the share of cancers the program caught — was 80.5% with AI support against 73.8% without, with p=0.031. Specificity was identical at 98.5% in both arms.

So AI support caught more cancer without calling back more women who did not have it. That combination is unusual and is the reason this trial is taken seriously.

What AI did not do

At no point in MASAI did software read a mammogram alone. A radiologist read every examination. The AI decided how many radiologists were needed and pointed at things it considered suspicious.

That is a workflow change inside a screening program, not a replacement of clinical judgment. And it was tested in one national screening system, with one software version, on one population. Whether it transfers is a question each health system has to answer for itself, ideally with its own data.

What this means if you are being screened

For the person having the mammogram, nothing about the appointment changes. NCI's fact sheet describes what to expect: the breast is placed between two plates and compressed, several images are taken from different angles, and results normally arrive within about two weeks.

The practical advice is unchanged too, and it is duller than the technology: go when you are invited. A screening program cannot find a cancer in someone who does not attend, whatever software is reading the images. Our overview of mammograms covers what a callback means and what happens next.

When to get checked

Screening is for people without symptoms. If you have any of these, do not wait for your next invitation:

• A new lump or thickened area in a breast or armpit.
• A change in the size or shape of one breast.
• Skin dimpling, puckering, or thickening like orange peel.
• A nipple newly turned inward, or a rash or crusting on it.
• Discharge from one nipple, particularly if bloody and without squeezing.
• A breast that is red, hot, and swollen and does not settle with antibiotics.

Our page on breast cancer explains how these are assessed.

What this does not mean

• The trial measured cancer detection, recalls, and interval cancers. It did not show that fewer women died of breast cancer, and it was not designed to. That would take far longer follow-up.
• One AI product was tested, at one version. Results do not transfer automatically to other software or other settings.
• The comparison arm was double reading by two radiologists. In systems where a single radiologist reads each mammogram, the math of both benefit and workload is different.
• Screening still finds some cancers that would never have caused harm, and still misses some that do. AI support narrowed those gaps in this trial. It did not close them.

Sources

• MASAI clinical safety analysis, Lancet Oncol 2023, PMID 37541274 — record retrieved via NCBI eutils
• MASAI screening performance, Lancet Digit Health 2025, PMID 39904652 — record retrieved via NCBI eutils
• MASAI interval cancer and sensitivity, Lancet 2026, PMID 41620232 — record retrieved via NCBI eutils
• NCI: Mammograms Fact Sheet

How this article was prepared

An AI-assisted editorial system helped prepare this page. No named medical reviewer has reviewed it unless one is listed.

Cancer Explained is published by the National Cancer Information Foundation. It is not medical advice and does not suggest a test or treatment.

See an error, old source, or unclear wording? Tell us.