Part VII · Detection, diagnosis, and biomarker assessment · Chapter 37
Screening and early detection
A screening programme trades deaths prevented against cancers that never needed finding. Both numbers exist, and both should be said out loud.
1 · Population screening strategies and the trial evidence
Every randomised trial of breast screening randomised an invitation, not a mammogram. That single design fact governs how the results should be read.
An independent panel reviewed the randomised evidence and the observational literature and reached a conclusion that has held up. In a meta-analysis of 11 randomised trials, the relative risk of breast cancer mortality for women invited to screening against controls was 0.80, with a 95% confidence interval of 0.73 to 0.891. That is a 20% relative reduction. The panel considered the internal biases and the age of the trials and judged 20% still reasonable.
The absolute version is the one a patient can use. For every 10,000 women in the United Kingdom aged 50 and invited to screening for the next 20 years, about 43 deaths from breast cancer would be prevented1.
Three properties of that estimate should be held together. It is an effect of a programme, diluted by everyone who was invited and did not attend. It comes from trials conducted before modern adjuvant therapy, and a better treated population has less to gain from earlier detection. And more recent observational studies give larger estimates, which the same panel judged likely to be biased.
The disagreement in this field is not mostly about whether screening works. It is about how large the benefit is, and what it costs in overdiagnosis. Quantifying overdiagnosis and communicating it honestly takes the second question.
2 · Risk-adapted and hereditary-risk screening
A uniform screening interval is an administrative decision rather than a biological claim. Risk-adapted screening asks whether the interval and the modality should follow individual risk, and it is being tested prospectively.
The WISDOM study is the first population-scale attempt. In its risk-based arm, 21,631 participants received a polygenic risk score. Adding that score to the Breast Cancer Surveillance Consortium clinical model changed the screening recommendation for 14% of women aged 40 to 49, and for 10% of women aged 50 to 742. Projected net screening encounters at population level were similar with and without the score.
Read that carefully, because it is more modest than the framing usually suggests. The polygenic score moved roughly one woman in seven in the younger group, and it did not reduce the total amount of screening. The trial has not reported whether risk-based screening is as safe as annual screening.
Hereditary risk is a different problem with a different answer. Carriers of high-penetrance pathogenic variants are managed with intensified surveillance from a younger age, usually including magnetic resonance imaging. The models that assign risk, and the variants that drive it, belong to Gail, Tyrer-Cuzick, BOADICEA and CanRisk compared and Inherited susceptibility and familial breast cancer.
A risk model is calibrated in the population it was built in. Polygenic scores were derived predominantly in cohorts of European ancestry, and WISDOM had to adapt its score for use in diverse populations2. Among the 21,631 participants who received a score, differences between major racial and ethnic groups were significant at P below 0.001. The report does not quantify how large those differences were, so what is established is that they exist. A score that is miscalibrated in a group will systematically assign the wrong interval to that group, which converts a precision tool into a disparity mechanism. Ancestry-associated tumor biology, what replicates and what is confounded treats the general problem.
3 · Mammography, tomosynthesis, ultrasound, and MRI
The four modalities fail in different ways, and the differences are physical rather than philosophical.
Mammography projects a three-dimensional breast onto a plane, so overlapping fibroglandular tissue can hide a lesion and can also create one. Tomosynthesis acquires a limited-angle arc and reconstructs planes, which reduces superimposition. It produces many more images per examination and takes longer to read.
Ultrasound adds no ionising radiation and is highly operator-dependent, with a false positive rate that limits its use as a primary screening test. Magnetic resonance imaging detects lesions by contrast kinetics rather than by density, which is why it is the most sensitive modality and also the least specific. The American College of Radiology recommends supplemental magnetic resonance imaging for women with a genetic risk, a lifetime risk of 20% or more, or chest irradiation at a young age, and risk assessment for every woman by age 303.
The endpoint question is where this section earns its place. Cancer detection rate is the outcome every modality comparison reports, and it is not the outcome that matters. A modality that finds more cancers without lowering the rate of advanced disease has increased diagnosis and not benefit. The Tomosynthesis Mammographic Imaging Screening Trial was designed with advanced cancer as its primary endpoint for exactly this reason, and it has not yet reported that endpoint.
The general rule follows. Detection is a property of the assay. Benefit is a property of what the detected disease would have done, and only the second can justify a change in policy.
4 · Dense breasts, supplemental imaging, and notification policy
Breast density is two separate things that share one word. It masks lesions on mammography, which is a property of the assay. It is also an independent risk factor for developing breast cancer, which is a property of the tissue. Conflating them produces most of the confusion in this area, and Breast density and other risk modifiers develops the risk side.
The masking side has been tested. The DENSE trial randomised 40,373 Dutch women aged 50 to 75 with extremely dense breast tissue and a normal screening mammogram. It allocated 8,061 to an invitation for supplemental magnetic resonance imaging and 32,312 to mammography alone4. The interval cancer rate was 2.5 per 1,000 screenings in the invitation group and 5.0 per 1,000 in the mammography-only group. The difference was 2.5 per 1,000, with a 95% confidence interval of 1.0 to 3.7 and a P value below 0.001.
The uptake figure is the one that makes this a policy result rather than an assay result. Only 59% of invited women accepted. Among those who actually underwent magnetic resonance imaging the interval cancer rate was 0.8 per 1,000, against 4.9 per 1,000 in those who declined4.
The costs are quantified in the same trial. The cancer detection rate on magnetic resonance imaging was 16.5 per 1,000 screenings, with a 95% confidence interval of 13.3 to 20.5. The false positive rate was 79.8 per 1,000. Positive predictive value was 17.4% for recall and 26.3% for biopsy4.
The DENSE endpoint was interval cancer, not mortality. Fewer interval cancers means the disease was found at screening instead of between screens. It does not establish that any woman lived longer, and some of the additional cancers found will be overdiagnosed Quantifying overdiagnosis and communicating it honestly.
Density notification legislation creates an obligation to inform without creating an entitlement to act. A woman told her breasts are dense, in a system that does not fund supplemental imaging, has been handed a problem rather than a service. Access to pathology, molecular testing, and imaging covers who can actually obtain the follow-up test.
5 · Interval cancers and what they reveal about biology
An interval cancer is one diagnosed between scheduled screens. Two entirely different events produce that label, and separating them is the whole content of this section.
The first is a cancer that was present at the last screen and was not seen. That is a failure of the assay or the reader, and masking by dense tissue is its commonest cause. The second is a cancer that was not present, or not detectable, at the last screen and grew to clinical attention within the interval. That is a statement about growth rate.
Genetic architecture separates the two. One unselected Swedish cohort followed 5,099 women with breast cancer. Carriers of protein-truncating variants in breast cancer predisposition genes were younger, and had more aggressive tumours, including oestrogen receptor-negative, high-grade, highly proliferative, luminal B and basal-like disease5. Ten-year breast cancer-specific survival was worse, with a hazard ratio of 1.65 and a 95% confidence interval of 1.16 to 2.36. Excluding 92 BRCA carriers, the association with high grade and worse survival persisted, with a hazard ratio of 1.76 and a 95% confidence interval of 1.21 to 2.56.
Common variants pointed the other way. Higher polygenic risk score was associated with less aggressive tumour characteristics, including oestrogen receptor-positive, low-grade, low-proliferation luminal A disease5. Among women with low mammographic density, below 25%, a higher polygenic score was associated with a lower risk of interval compared with screen-detected cancer. The odds ratio was 0.77, with a 95% confidence interval of 0.64 to 0.93.
That last finding is the argument. In women in whom masking is minimal, genetic risk architecture still predicts mode of detection. Interval cancers are therefore enriched for fast-growing biology and not only for dense tissue.
The consequence for policy is specific. Shortening the interval addresses the growth-rate cause. Changing the modality addresses the masking cause. They are different interventions for different problems, and a programme that applies one to the other will underperform.
Mode of detection is a heterogeneity readout. A screening programme is a filter with a fixed temporal resolution, and what passes through it is selected for growth rate. Screen-detected disease is therefore systematically enriched for the slow, receptor-positive end of the spectrum, and interval disease for the fast, receptor-negative end. That is the same axis that separates early from late recurrence in Early and late recurrence as different biological problems, and the same axis that determines how much circulating tumour DNA a deposit sheds Biology of circulating tumor DNA, shedding, half-life, and detectability. One biological variable, growth rate, is being measured three times by three instruments that appear unrelated.
6 · Quantifying overdiagnosis and communicating it honestly
Overdiagnosis is the detection of a cancer that would never have become clinically apparent in that woman's lifetime. It is a counterfactual, so it can never be observed in an individual patient. It can only be estimated in a population, and only where a control group was never offered screening.
The best estimates come from three randomised trials in which control-group women were not invited to screen at the end of the active trial period. A meta-analysis of those trials gave two figures. Excess incidence was 11% as a proportion of the cancers diagnosed in the invited group over the long term, with a 95% confidence interval of 9% to 12%. It was 19% as a proportion of the cancers diagnosed during the active screening period, with a confidence interval of 15% to 23%1.
Those two numbers describe the same phenomenon. They differ by a factor of nearly two because they have different denominators, and both are correct. Quoting one without its denominator is the single commonest source of argument about overdiagnosis, and a claim that screening overdiagnoses 11% or 19% of cancers means nothing until the denominator is stated1.
The absolute frame is better for a conversation. For every 10,000 women aged 50 invited to screening for 20 years, about 43 breast cancer deaths would be prevented and about 129 cancers, invasive and in situ, would be overdiagnosed1. That is roughly one death prevented for every three overdiagnosed cases identified and treated. Just over 1% of women invited to begin screening would have an overdiagnosed cancer over the following 20 years.
Give absolute numbers with a shared denominator, not percentages. "For every 10,000 women like you invited over 20 years, about 43 deaths prevented and about 129 cancers found that would never have caused trouble" is a sentence a person can weigh1. "A 20% mortality reduction and 19% overdiagnosis" is not, because the two percentages are of different things.
Say that overdiagnosis cannot be identified in an individual. The woman in front of you cannot be told whether her cancer is one of the 129, and that is the honest and unwelcome part1.
Do not present the estimates as precise. They come from old trials with real limitations, and the panel that produced them said they should be regarded as an approximate guide1.
The active surveillance trials for low-risk ductal carcinoma in situ are the direct clinical response to this problem, and COMET, LORD and LORIS are covered in Active surveillance trials and what they will and will not answer.
7 · Artificial intelligence in screening, performance and equity effects
Artificial intelligence in mammography screening has now been tested in a randomised trial, which distinguishes it from most claims in this area.
The MASAI trial randomised women in the Swedish national programme to artificial intelligence-supported screen reading or to standard double reading. Its prespecified clinical safety analysis covered 80,033 women. The cancer detection rate was 6.1 per 1,000 screened in the intervention group and 5.1 per 1,000 in the control group. The ratio was 1.2, with a 95% confidence interval of 1.0 to 1.5 and a P value of 0.0526. Screen-reading workload fell by 44.3%.
The larger protocol-defined analysis of 105,934 women sharpened the result. The cancer detection rate was 6.4 per 1,000 against 5.0 per 1,000, a ratio of 1.29 with a 95% confidence interval of 1.09 to 1.51 and a P value of 0.00217. Invasive cancers rose from 217 to 270, and in situ cancers from 45 to 68. The recall ratio was 1.08 and the false positive ratio was 1.01, neither statistically significant. Positive predictive value of recall rose, with a ratio of 1.19 and a P value of 0.012. Screen-reading workload fell by 44.2%.
More cancers found, no more false positives, and half the reading time is a genuine result. It is also not yet an outcome result. The primary endpoint of MASAI is interval cancer rate, and it has not reported. Some of the extra detection was in situ disease, which is where overdiagnosis concentrates Quantifying overdiagnosis and communicating it honestly.
The equity question is the live controversy, and the most useful fact about it is a gap in the evidence. MASAI did not collect race and ethnicity data6. The strongest trial supporting artificial intelligence in screening therefore cannot say whether its benefit is distributed equally.
Two mechanisms pull in opposite directions and neither has been measured in a trial. Artificial intelligence can substitute for a second reader, which raises the floor in programmes that cannot staff double reading, and those are disproportionately the programmes serving underserved populations. A model trained on one population can also perform differently on another, and breast density, age distribution and tumour subtype distribution all differ across populations Subtype distribution and young-onset disease across populations.
Do not read the absence of a demonstrated equity harm as evidence of its absence. Subgroup performance was not measured in the trial that produced the strongest evidence6. A model can improve average performance and worsen performance in a subgroup at the same time, and an average is exactly the statistic that hides it. Data governance, privacy, and algorithmic equity treats algorithmic equity as a general problem.
8 · Emerging detection technologies
Three families of technology are in development, and each faces a different constraint.
Contrast-enhanced mammography and abbreviated magnetic resonance imaging address the same problem from opposite directions. The first adds vascular contrast to an existing widely available platform. The second shortens an expensive protocol to increase throughput. Both aim at the dense-breast masking problem in Dense breasts, supplemental imaging, and notification policy, and both are limited by access and cost rather than by physics.
Blood-based multicancer early detection tests face an arithmetic problem that is not solved by improving the assay. Positive predictive value in a screening population is governed by prevalence, and breast cancer prevalence in a screening round is low. A test with excellent specificity still generates mostly false positives when prevalence is low enough, and each one initiates a diagnostic workup across multiple organ systems.
There is also a biological ceiling specific to circulating tumour DNA. Shedding scales with cell turnover and vascular access, so a small early tumour contributes very little DNA to plasma Biology of circulating tumor DNA, shedding, half-life, and detectability. The tumours these tests are least able to detect are precisely the tumours a screening test exists to find. Improving analytic sensitivity narrows that gap without closing it, because the signal is bounded by the biology rather than by the instrument.
None of these technologies has a randomised trial showing a reduction in advanced-stage disease or in mortality. The evidence bar set in Population screening strategies and the trial evidence applies to them as it applied to mammography.
9 · Early detection in resource-constrained settings
Where mammographic screening is not feasible, the question becomes whether a cheaper test can downstage disease enough to matter.
A cluster-randomised trial in Mumbai answered it more precisely than anything else in this literature. Twenty geographically distinct clusters and 151,538 women aged 35 to 64 were allocated to one of two arms. The screening arm received four rounds of clinical breast examination by trained female primary health workers every two years. The control arm received cancer awareness and active surveillance8.
Downstaging was clear. The proportion of women diagnosed with stage III or IV disease was 37% in the screening arm and 47% in the control arm, with a P value of 0.0018.
Mortality was not. The overall breast cancer mortality rate ratio was 0.85, with a 95% confidence interval of 0.71 to 1.01 and a P value of 0.07, which is a non-significant 15% reduction. A post hoc subset analysis found a rate ratio of 0.71 in women aged 50 and older, with a 95% confidence interval of 0.54 to 0.94 and a P value of 0.02. In women under 50 the rate ratio was 0.93, with a confidence interval of 0.79 to 1.09. All-cause mortality showed a rate ratio of 0.95, with a confidence interval of 0.81 to 1.108.
The honest reading is that the trial met its downstaging endpoint and missed its mortality endpoint, and that the positive subgroup was defined after the fact.
The structural point is larger than the trial. Early detection converts a late diagnosis into an early diagnosis. It converts an early diagnosis into a life saved only where diagnostic pathology, surgery, radiotherapy and systemic therapy are available and reachable. Detection programmes built ahead of treatment capacity generate diagnoses that the system cannot act on, and Pathology and diagnostic infrastructure as the rate-limiting step makes the case that pathology infrastructure is the rate-limiting step. The resource-stratified framework for deciding what to build first is in Resource-stratified guidelines and tiered recommendations.
References
All 8 references
- Independent UK Panel on Breast Cancer Screening. The benefits and harms of breast cancer screening: an independent review. Lancet 2012 380:1778-1786. PMID 23117178
- Fergus KB, Heise RS, Madlensky L, et al. Integrating breast cancer polygenic risk scores at scale in the WISDOM Study: a national randomized personalized screening trial. Genome Med 2025 17:97. PMID 40877879
- Monticciolo DL, Newell MS, Moy L, Niell B, Monsees B, Sickles EA. Breast Cancer Screening in Women at Higher-Than-Average Risk: Recommendations From the ACR. J Am Coll Radiol 2018 15:408-414. PMID 29371086
- Bakker MF, de Lange SV, Pijnappel RM, et al. Supplemental MRI screening for women with extremely dense breast tissue. N Engl J Med 2019 381:2091-2102. PMID 31774954
- Li J, Ugalde-Morales E, Wen WX, et al. Differential burden of rare and common variants on tumor characteristics, survival, and mode of detection in breast cancer. Cancer Res 2018 78:6329-6338. PMID 30385609
- Lang K, Josefsson V, Larsson AM, et al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI). Lancet Oncol 2023 24:936-944. PMID 37541274
- Hernstrom V, Josefsson V, Sartor H, et al. Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI). Lancet Digit Health 2025 7:e175-e183. PMID 39904652
- Mittra I, Mishra GA, Dikshit RP, et al. Effect of screening by clinical breast examination on breast cancer incidence and mortality after 20 years: prospective, cluster randomised controlled trial in Mumbai. BMJ 2021 372:n256. PMID 33627312