Mammography, by the numbers

Worth the Test · Written by an emergency physician · Last reviewed September 2026

A positive screening mammogram usually means "come back for more pictures," not "you have cancer." In a large study of U.S. community practice, 11.6% of screening mammograms were read as abnormal, roughly one in nine, and 4.4% of those callbacks turned out to be cancer. This page shows where those numbers come from, what a recall tells you, how the odds shift with risk and repeated screening, and how screening's long-run benefits and harms compare on one denominator.

Who the recommendation covers

Breast cancer killed an estimated 43,170 U.S. women in 2023. In April 2024 the U.S. Preventive Services Task Force (USPSTF) recommended screening mammography every two years for women aged 40 to 74 (grade B).

The recommendation covers cisgender women and all other people assigned female at birth who are 40 or older and at average risk, including those with a family history of breast cancer or dense breasts. It excludes people with a high-risk genetic marker or syndrome, such as a BRCA1 or BRCA2 variant, past high-dose chest radiation at a young age, or a previous breast cancer or high-risk lesion on biopsy. Like every Task Force recommendation, it concerns people without signs or symptoms.

For women 75 and older, and for supplemental ultrasound or MRI after an otherwise negative mammogram that shows dense breasts, the Task Force found the evidence insufficient to judge the balance of benefits and harms. That "I statement" is neither a recommendation for nor against.

What a recall means

The accuracy figures here come from the Breast Cancer Surveillance Consortium (BCSC): 1,682,504 digital screening mammograms performed from 2007 to 2013 in 792,808 women, read by 359 radiologists at 95 U.S. facilities (Lehman and colleagues). The abnormal interpretation rate was 11.6%, so about 116 of every 1,000 women screened in a round were called back.

A recall means more imaging first: extra views, perhaps an ultrasound, and less often a biopsy. Of all recalls, 4.4% ended in a cancer diagnosis; of screens that led to a biopsy recommendation, 25.6%; of biopsies actually performed, 28.6%, so roughly 7 in 10 of those biopsies were benign. Most false alarms never reach a needle: in the modeling behind the 2024 recommendation, screening every two years from 40 to 74 with tomosynthesis (3D mammography) produced 1,376 false-positive recalls but 201 benign biopsies per 1,000 women over a lifetime.

Detection rate is not prevalence

To see why only about 4 in 100 recalls are cancer, rebuild one screening round. The BCSC reported 5.1 cancers detected per 1,000 screens and 0.8 false-negative cancers per 1,000 (cancers diagnosed after a screen read as normal), with sensitivity of 86.9% and specificity of 88.9%.

The detection rate is not the prevalence. It counts only the cancers the screen found, so the sensitivity loss is already built in; applying sensitivity to it again would remove the missed cancers twice. Prevalence needs them added back: 5.1 + 0.8 = 5.9 per 1,000, or about 0.59%. These are rounded summaries, so the cohort they rebuild is approximate.

Ten thousand screens, in whole people

Start with 10,000 screening mammograms: 59 in women with breast cancer and 9,941 in women without. Apply sensitivity only to the first group and specificity only to the second.

The reconstruction recalls 1,154 of 10,000 screens, 11.5%, close to the 11.6% the BCSC measured. Follow the recalled column: 51 of 1,154 have cancer. From the unrounded rates, as the calculator computes it:

PPV = (0.869 × 0.0059) ÷ [(0.869 × 0.0059) + (0.111 × 0.9941)] = 0.00513 ÷ 0.11547 ≈ 4.44%

That matches the study's reported 4.4%. Dividing the rounded people, 51 ÷ 1,154, gives 4.42%; the difference is rounding.

For a woman who has just been called back, the chance of cancer is about 1 in 23. The recall is real information, since 4.4% is about seven and a half times the 0.59% she started with, but a raised probability is not a diagnosis. The base-rate fallacy and sensitivity vs. specificity guides explain why an accurate test yields mostly false alarms when a condition is uncommon.

The negative column reassures in a different way. Of 8,846 women with a normal screen, 8 have a cancer this round missed: from the unrounded rates, a remaining risk of about 0.087%, roughly 1 in 1,144. That is well below the starting risk, but not zero.

What moves the predictive value

The base rate does much of the work. Hold sensitivity and specificity fixed and raise prevalence to an illustrative 2%, a teaching value rather than an estimate for any age or risk group. Per 1,000 screens, 20 women now have cancer and 980 do not; the screen flags about 17 of the first group and 109 of the second:

PPV = (0.869 × 0.02) ÷ [(0.869 × 0.02) + (0.111 × 0.98)] = 0.01738 ÷ 0.12616 ≈ 13.8%

About one recall in seven is now cancer, three times the base case, from the same test: the pre-test to post-test probability lever at work.

The 2% example holds accuracy fixed to isolate one effect. Real screening does not:

• Breast density. The Task Force notes that nearly half of women have dense breasts, which are associated with lower sensitivity and specificity and with higher breast cancer risk.
• Who reads the images. Only 59.0% of BCSC radiologists had abnormal interpretation rates within recommended ranges.
• Technology and interval. In the Task Force's modeling, standard digital mammography gave benefits similar to tomosynthesis but more false-positive recalls: a median of 1,540 versus 1,376 per 1,000 women screened every two years from 40 to 74. Annual screening added benefit, and also false-positive recalls and overdiagnosis.

Ten screens, not one

An 11.1% false-positive rate describes one round, but screening is a series. Elmore and colleagues reviewed ten years of records for 2,400 women aged 40 to 69 at study entry. They estimated a 49.1% cumulative risk of at least one false-positive result after ten mammograms, and that 18.6% of women without breast cancer would undergo a biopsy over those ten screens.

Treating each round as independent at 88.9% specificity gives 1 − 0.889¹⁰ ≈ 69.2% after ten rounds. The gap between 69% and 49% does not measure how strongly false positives cluster in the same women: Elmore's cohort was a different population in an earlier era. Dependence between rounds can push cumulative risk either way, which is why the repeated-screening guide treats the independence formula as a model, not a bound.

Benefit and harm on one denominator

Whether screening is worth its harms depends on outcomes over many rounds, which need their own evidence and denominator. Two estimates dominate the debate.

The first is the CISNET modeling that informed the 2024 recommendation (Trentham-Dietz and colleagues). Six simulation models projected lifetime outcomes, assuming every woman attends every screen and is treated as in current practice. Compared with no screening, tomosynthesis every two years from 40 to 74 was associated with a 30.0% reduction in breast cancer mortality.

On this ledger, a median of 8.2 breast cancer deaths averted per 1,000 women screened sits beside more than one false-positive recall per woman on average, 201 benign biopsies and 14 overdiagnosed cancers, which would never have been diagnosed in the woman's lifetime without screening.

The second estimate is the Cochrane review by Gøtzsche and Jørgensen (2013) of randomised trials including about 600,000 women aged 39 to 74. Assuming a 15% reduction in breast cancer mortality and 30% overdiagnosis and overtreatment, the authors calculated that for every 2,000 women invited to screening over ten years, one would avoid dying of breast cancer, 10 healthy women would be treated for a cancer that would not otherwise have been diagnosed, and more than 200 would experience important psychological distress from false positives.

That estimate is contested. The three trials the review judged adequately randomised showed no statistically significant reduction in breast cancer mortality (relative risk 0.90, 95% CI 0.79–1.02); all seven combined gave 0.81 (0.74–0.87). The Task Force's 2024 statement cites an updated analysis of three Swedish trials reporting a 15% relative reduction in breast cancer mortality for women aged 40 to 74 (relative risk 0.85, 95% CI 0.73–0.98). It also cites overdiagnosis estimates of roughly 11% to 19% of the cancers diagnosed, from trials with comparable groups, adequate follow-up and no end-of-trial screening of the control group. Drawing on a newer review and this modeling, the Task Force concluded with moderate certainty that biennial screening from 40 to 74 has a moderate net benefit. The two estimates also measure different things:

• Invited versus screened. Cochrane counts women offered screening, not all of whom attend; CISNET assumes full attendance.
• Ten years versus a lifetime. Cochrane's window is ten years; CISNET follows women screened from 40 to 74 for life.
• Trial era versus modern modeling. The trials predate current treatment, and the Cochrane authors expected a smaller absolute effect today as treatment and awareness improved; CISNET models current treatment and imaging.
• Different comparators. Trials compare invited and uninvited groups; the models compare one simulated population, screened and unscreened.

Both are program-level figures, per woman invited or screened across many rounds, not per positive screen or per woman treated. Neither belongs in the calculator's treatment panel, which applies its NNT to the true positives from one round; entering 2,000 there would apply prevalence and sensitivity again and wrongly shrink the benefit. The NNT, NNH and NNS guide explains why, and screening harms covers overdiagnosis.

What generalizes

A positive screen is a prompt for more testing, and its meaning depends on the base rate as much as on accuracy. A detection rate excludes the cases the test missed, so it understates prevalence. A modest false-positive rate per round adds up over years of screening. And benefits and harms compare cleanly only on a shared denominator: per screen, per woman screened, or per woman invited. None of this settles whether screening is worthwhile for any one person; it makes the trade-off explicit.

Correction, September 2026. An earlier version of this page used the 0.51% detection rate as the prevalence, which subtracted the missed cancers twice and gave a positive predictive value of 3.9%; the corrected figure is 4.4%. The corrections log lists every change.