Mammography and Breast Cancer Screening

Screening mammography looks for breast cancer in women without symptoms, so that cancers are found while they are small. Each examination is a handful of X-ray images of the breasts, nearly all of them normal, read by one or two radiologists who decide whether to recall the woman for further tests. That makes screening a large, repetitive visual task with a very low rate of positives, measured by outcomes that unfold over years, and it has become one of the most closely studied applications of medical computer vision. Computer Vision in Radiology gives the outline: the computer-aided detection (CAD) systems of the 1990s and 2000s, the observational studies that found no benefit from them in practice, and the main results of McKinney et al. (2020) and of the MASAI trial. This article goes deeper into the application itself: the screening programmes and their reading protocols, the images, the categories radiologists report, the metrics a programme is judged by, the public datasets, and how the design of each AI study decides what its result means. It gives no clinical guidance.

Problem

A screening programme invites a defined population at fixed intervals, and the reading protocol differs between countries. As of November 2026:

• England. The NHS Breast Screening Programme offers screening to women aged 50 up to their 71st birthday, and a research trial (AgeX) has been looking at offering some women one extra screen between the ages of 47 and 49 and one between 71 and 73 (NHS England, programme overview). McKinney et al. (2020) describe the programme in their data (2010–2018) as screening every 3 years with double reading: two readers each decide whether to recall, and an arbitration process settles disagreements. The readers work in sequence, so the second reader can see the first reader’s opinion.
• Germany. The national programme screens women aged 50–69; four two-dimensional images are taken, two radiologists read them independently, and if either finds the case suspicious a consensus conference of at least the two readers and a head radiologist decides on recall (Eisemann et al., 2025).
• Sweden. The MASAI trial enrolled women aged 40–80 eligible for screening at four sites, including general screening at 1.5–2-year intervals and annual screening for women at moderate hereditary risk or with a history of breast cancer; standard reading was double reading (Lång et al., 2023).
• United States. The US Preventive Services Task Force recommended in April 2024 biennial screening mammography for women aged 40 to 74 (USPSTF, 2024). McKinney et al. (2020) describe US screening as typically read by a single radiologist, with screens “annually or biannually” (their words; their test-set description gives every 1–2 years).

Programmes record what each reader decided and, through follow-up and cancer registries, what became of each woman. That record is the raw material for both the metrics below and the retrospective AI studies built on them.

Visual Data

Views and examinations

A mammogram is a projection radiograph of the compressed breast, taken with a low-energy spectrum; X-ray Radiography covers the physics and the Digital Mammography X-Ray Image object in which DICOM stores it. The standard screening examination has four images, a craniocaudal (CC) and a mediolateral oblique (MLO) view of each breast (McKinney et al., 2020; Lee et al., 2017). DICOM codes these and other views in its View for Mammography context group, alongside medio-lateral, latero-medial and other projections (NEMA, 2026, CID 4014), and besides the laterality and implant attributes that X-ray Radiography mentions, the Mammography Image Module records whether an image is a partial view, a subset of a single view of the breast (NEMA, 2026, C.8.11.7).

For vision, an examination is a small set of related images, not a single input. The two views of one breast show the same tissue from different angles, the two breasts are roughly mirror images, and readers compare with earlier examinations. The National Cancer Institute (NCI) notes that radiologists compare the most recent mammogram with past ones where possible (NCI, Mammograms), and the FDA’s 2023 final rule under the Mammography Quality Standards Act (MQSA) accepts that comparison with previous mammograms improves accuracy and reduces recalls (FDA, 2023, Response 43). Two acquisition details matter for training data. Vendors apply their own processing to the images (McKinney et al., 2020), and in both of McKinney et al.’s datasets the vast majority of images came from one manufacturer, a limitation they name. And DICOM separates images for presentation from images for further processing (X-ray Radiography describes the two classes); VinDr-Mammo contains only for-presentation images, because the hospitals did not store the others (Nguyen et al., 2023).

Digital breast tomosynthesis

Digital breast tomosynthesis (DBT), also called 3D mammography, takes multiple images of thin slices across the breast that software assembles into a three-dimensional picture, and it also produces 2D images like those of standard mammography (NCI, Mammograms). The USPSTF considers both digital mammography and DBT effective screening modalities and notes that DBT must be accompanied by traditional digital mammography or by synthetic digital mammography, a 2D image constructed from the DBT data (USPSTF, 2024). DICOM stores the reconstructed volumes as Breast Tomosynthesis Image objects, which are not meant for the source projection images (NEMA, 2026, A.55); projection images and generated 2D images are identified by the Image Type of mammography images (NEMA, 2026, C.8.11.7). For a vision system, DBT turns four images into four volumes plus synthetic views, a much larger input in different data objects. In US practice, the Breast Cancer Surveillance Consortium (BCSC) reports benchmarks from 2011–2021 (BCSC, performance table) separately for 998,135 DBT and 2,695,664 digital mammography screening examinations: recall rates of 8.7% and 9.9%, cancer detection rates of 6.2 and 5.4 per 1,000, and sensitivities of 88.0% and 87.4%. These are pooled observational figures from different women and facilities, not a comparison of the two modalities in the same population.

Breast density

Breast density describes the relative amounts of dense and fatty tissue seen on a mammogram, and mammography is more likely to miss cancer in dense breasts (NCI, Mammograms). In the United States this is now regulation. The MQSA final rule, effective 10 September 2024, requires the report to the healthcare provider to classify overall density in one of four categories: (A) the breasts are almost entirely fatty; (B) there are scattered areas of fibroglandular density; (C) the breasts are heterogeneously dense, which may obscure small masses; (D) the breasts are extremely dense, which lowers the sensitivity of mammography. It also requires the patient’s letter to state whether her tissue is dense (FDA, 2023). VinDr-Mammo uses the same four categories, which it takes from BI-RADS (Nguyen et al., 2023). The effect of density is partly an image-processing fact: dense fibroglandular tissue has the same brightness and scale as a mass, so it acts as structured noise, which the example below simulates.

BI-RADS categories

Breast Imaging Reporting and Data System (BI-RADS) is the American College of Radiology’s reporting system, published as an atlas that the BCSC definitions cite (BCSC, Standard Definitions v4); the atlas was not read for this article, so BI-RADS is described here through published sources that use it. VinDr-Mammo lists seven assessment categories: 0 (needs additional imaging or prior examinations), 1 (negative), 2 (benign), 3 (probably benign), 4 (suspicious), 5 (highly suggestive of malignancy) and 6 (known biopsy-proven malignancy), together with BI-RADS descriptors for masses, calcifications, asymmetries and architectural distortion (Nguyen et al., 2023). The NCI gives the same categories with the follow-up each implies (NCI, Mammograms).

BI-RADS scores are often used as labels or as the human comparison, and the choice of which scores count as positive changes the result. McKinney et al. (2020) treated US screening scores of 0, 4 and 5 as positive. The BCSC definitions count initial assessments of 0, 3, 4 and 5 as positive, and since 2014 treat 3 as positive to follow the fifth BI-RADS edition; for a final result they offer two definitions, one counting only 4 and 5 as positive, the other also counting 3 (BCSC, Standard Definitions v4). The MQSA rule computes the recall rate from assessments of “Incomplete: need additional imaging evaluation” and notes that the other positive assessments are clinically discouraged on screening examinations (FDA, 2023, Response 110). Lee et al. (2017) caution that many BI-RADS assessments in the DDSM were probably updated after further work-up, since 4 and 5 are unconventional on screening images, so a dataset’s BI-RADS field is not necessarily the screening reader’s decision.

Screening Metrics

A screening examination is counted as positive or negative from the reader’s assessment, and as a cancer or not from what is diagnosed within a follow-up period. In the BCSC definitions, a screening mammogram is followed for 365 days, with the follow-up truncated at the next screening mammogram (the definition’s sentence says when that occurs within 90 days, while its worked example truncates at a screen ten months later), whereas the ACR BI-RADS definition uses a strict 365 days; a true positive is a positive examination with ductal carcinoma in situ or invasive cancer diagnosed in that period, and a false negative is a negative examination with such a diagnosis (BCSC, Standard Definitions v4). From these counts over $N$ examinations:

$$\mathrm{RR} = \frac{\mathrm{TP} + \mathrm{FP}}{N}, \qquad \mathrm{CDR} = 1000\,\frac{\mathrm{TP}}{N}, \qquad \mathrm{Se} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FN}}, \qquad \mathrm{PPV}_1 = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP}}.$$

The recall rate RR is what the BCSC calls the abnormal interpretation rate, the cancer detection rate CDR is counted per 1,000 examinations, and $\mathrm{PPV}_1$ is the positive predictive value of a positive screening assessment; the BCSC also defines $\mathrm{PPV}_2$ (biopsy recommended) and $\mathrm{PPV}_3$ (biopsy performed). In the United States, the MQSA rule requires every facility’s annual medical outcomes audit to compute at least the PPV, the CDR and the recall rate (FDA, 2023). It follows from the definitions that $\mathrm{PPV}_1 = \mathrm{CDR} / (1000\,\mathrm{RR})$ when both are computed from the same assessment: once the recall rate and detection rate are known, the PPV adds no information. The BCSC computes $\mathrm{PPV}_1$ from the initial assessment but suggests the final result for sensitivity and CDR, in which case the identity holds only approximately. Precision and Recall covers these quantities in general.

The metrics behave differently from those of a classification benchmark in three ways.

• Rates per screened population. Recall rate and CDR are counted over all women screened, so they depend on prevalence and need no knowledge of the cancers a programme missed. Sensitivity needs the false negatives, which are only known through follow-up.
• Interval cancers. A cancer diagnosed after a negative screen and before the next one is an interval cancer; the BCSC codes cancers as interval-detected after a false-negative screening examination (BCSC, Standard Definitions v4). Interval cancers include cancers missed at screening and fast-growing cancers that were not visible then, so the interval cancer rate measures what screening fails to catch in a way no reading of the original images can. MASAI made it the trial’s primary endpoint (Lång et al., 2023).
• The follow-up window is part of the metric. McKinney et al. (2020) counted cancers diagnosed up to 39 months after a UK screen and 27 months after a US screen, intervals chosen to include the next screening round, and required a later non-suspicious screen before calling a woman negative. They argue that short follow-up favours readers, because a cancer the reader did not recall is less likely to be diagnosed soon, and that the longer window makes the task harder because it includes cancers with no visible sign in the original images; their sensitivities are therefore lower than figures reported for 12 months, and their Extended Data give the 12-month figures as well.

Typical values differ by protocol. MASAI’s standard double-reading arm had 1,027 recalls among 52,872 women (1.9%, our calculation) and a detection rate of 5.0 per 1,000 (Hernström et al., 2025), the German control arm in PRAIM a recall rate of 38.3 per 1,000 with 5.7 per 1,000 detected (model-based estimates, Eisemann et al., 2025), and US digital mammography in the BCSC data 9.9% with 5.4 per 1,000.

Vision Tasks Involved

The case-level task is classification: from the four views, with or without priors, give a score for the presence of cancer, which is thresholded into a recall decision. Medical Image Classification covers the general problem, including study-level classification from several images and the choice of operating point. Below the case level are three more tasks:

• Lesion detection and localization. Masses, calcifications, asymmetries and architectural distortion are marked with boxes or contours (Nguyen et al., 2023; Lee et al., 2017). Detection is a form of object detection, and localization matters for review: a reader shown a score alone cannot check what the system saw. Lesion Detection covers candidate generation, matching criteria and free-response (FROC) evaluation.
• Mass segmentation and characterization. Mass shape and margin are BI-RADS descriptors, and CBIS-DDSM provides mass outlines refined by a level-set method so that shape features can be measured (Lee et al., 2017).
• Density assessment. Density is reported per examination or per breast, and VinDr-Mammo labels it for each breast (Nguyen et al., 2023).

The system of McKinney et al. (2020) shows how these levels combine. Its accepted manuscript describes an ensemble of three models whose case-level scores are averaged. A lesion model runs a RetinaNet detector on all four images, crops the ten highest-scoring regions together with a reference region from the other breast, scores each crop, and combines the crop scores into a case score by noisy-OR, the probability that not all lesions are benign under an independence assumption, which lets the lesion classifier be trained from case-level labels. A breast model applies a ResNet to each image and concatenates the CC and MLO features of each breast, taking the maximum of the two breast scores. A case model concatenates features from all four images. Only the patient’s age was given to the system, while the original readers had history and prior examinations.

Datasets

Public mammography datasets differ in era (scanned film or digital), in what a label means, and in how they can be obtained. The table gives the facts as their papers and pages state them, as of November 2026; no data-use agreement was accepted to write this article.

Dataset	Source	Size	Images	Labels	Access
CBIS-DDSM (Lee et al., 2017)	Curated subset of the DDSM, four US institutions	1,566 participants, 10,239 images (TCIA); 753 calcification and 891 mass cases (paper)	Scanned film, decompressed to DICOM	Pathology (benign, benign without call-back, malignant), BI-RADS assessment and descriptors, density, mass outlines; standard train/test split	TCIA collection, CC BY 3.0
INbreast	Portugal, introduced 2012	115 studies, 410 images (as tabulated by Nguyen et al., 2023)	Full-field digital	Lesion contours, BI-RADS assessment, density	Terms not confirmed from an official page
VinDr-Mammo (Nguyen et al., 2023)	Two hospitals in Hanoi, 2018–2020	5,000 four-view examinations, 20,000 images	Full-field digital, for-presentation only, three vendors	Breast-level BI-RADS and density, boxes for non-benign findings, from double reading with arbitration; 4,000/1,000 split	PhysioNet restricted access: registered users who sign the data use agreement, PhysioNet Restricted Health Data License 1.5.0
OMI-DB	NHS Breast Screening Programme sites, UK	About 750,000 cases (home page)	Digital, de-identified NHSBSP images	In the data of McKinney et al. (2020): reader opinions, follow-up, rectangular regions of interest at later biopsies	Application reviewed by an OPTIMAM committee, then a project-specific sharing agreement

A few details of each matter for how results on them can be read.

• CBIS-DDSM. The DDSM is a database of 2,620 scanned film studies released in 1997, stored in an obsolete lossless JPEG format whose region annotations gave only the general position of lesions (Lee et al., 2017). CBIS-DDSM removed images in which a reviewing mammographer could not clearly see the annotated mass (339 images in the paper; the TCIA page says 254), converted the images through optical density to 16-bit DICOM, and split each of the mass and calcification sets 80/20 stratified by BI-RADS assessment. A “case” in its counts is an abnormality, not a woman, and TCIA notes that the DICOM metadata make the 1,566 participants appear as 6,671 patients, because each participant has several patient IDs; a split by DICOM patient ID is therefore not a split by woman, the leakage Medical Image Classification warns about.
• VinDr-Mammo. Biopsy results were not available, so the labels are radiologists’ assessments, and BI-RADS 6 does not occur; benign findings (BI-RADS 2) were not marked, to save time (Nguyen et al., 2023). The breasts are mostly dense: 76.5% of breasts are category C and 13.5% category D, against 0.5% A. Examinations were sampled at random from the hospitals’ archives and include diagnostic as well as screening examinations.
• OMI-DB. The UK data of McKinney et al. (2020) were compiled by OPTIMAM from three NHS screening sites, with screening and follow-up mammograms and all reader opinions including arbitration. OMI-DB’s access page lists the terms of its sharing agreements: no sharing with third parties without permission, research limited to the approved project, no attempt to obtain personal information, and a publication policy.

These datasets also differ in what their labels are measured against. CBIS-DDSM has pathology for its abnormalities, VinDr-Mammo radiologists’ reading, and screening archives such as OMI-DB have follow-up through later rounds, which is what screening metrics need. None of the public sets above reproduces the prevalence of a screening population, which is why the evaluations below rest on programme data.

Evaluation: Study Designs

Validation designs in general, from internal test splits to reader studies, prospective studies and randomized trials, are the subject of Clinical Validation and Regulation of Medical Imaging AI, and Computer Vision in Radiology describes the reader study of McKinney et al. (2020), with its enriched set of 500 US cases, and the gap between enriched test sets and screening populations. This section concentrates on how screening studies are built and on what each design can show about screening outcomes.

Retrospective studies on programme data

A retrospective study scores archived examinations and compares the system with the decisions readers made at the time. McKinney et al. (2020) built two test sets with opposite designs. The UK test set is a random 10% sample of the women screened at two sites in 2012–2015 (25,856 women), which the authors found to resemble the national programme in cancer prevalence, age and cancer type; one eligible examination per woman was chosen at random. The US test set (3,097 women, one academic centre, 2001–2018) includes every woman who had a breast biopsy and a random sample of about 5% of women never biopsied, and when choosing one examination per woman it preferred examinations before a biopsy, to enrich for difficult benign cases. To estimate screening-population performance from this enriched sample, they weighted each case by the inverse of its sampling probability: women never biopsied were upweighted by 19.04 for the 5% sampling, then by a further factor of 1.24 (estimated by Monte Carlo simulation of purely random selection) for the preferential choice of examination, a final weight of 23.61; examinations followed by a negative biopsy were downweighted by 0.64, and cancers carried a weight of 1. Without such weights, an enriched set overrepresents hard negatives and understates specificity.

Thresholds were chosen on validation data, separately for each human benchmark: in the UK, aiming at superiority in specificity and non-inferiority in sensitivity against the first reader, and at non-inferiority against the second reader and the consensus; Computer Vision in Radiology gives the resulting comparisons. The system did not outperform the consensus, the UK standard of care. A version trained only on UK data, with a threshold chosen on separate US validation data, still showed superior specificity (+3.5%) and sensitivity (+8.1%) to the US radiologists, a smaller margin than the full system’s.

Retrospective data also allow workflow simulations. McKinney et al. (2020) replaced the UK second reader by the system whenever the system agreed with the first reader, invoking the second reader and arbitration only for the 12.02% of cases with disagreement, and report performance equivalent to double reading with 88% of the second reader’s work saved. A limitation of every such study is what Eisemann et al. (2025) call the uncertainty of outcomes for examinations flagged only by the AI: if neither reader recalled a woman, a cancer the system saw is only known if it surfaces later as an interval or next-round cancer, so retrospective data can show missed cancers but cannot show what a recall prompted by the system would have found.

Prospective paired-reader studies

In a paired-reader design, every examination is read under all the strategies being compared, so the strategies are compared on the same women. ScreenTrustCAD did this at one Stockholm hospital from April 2021 to June 2022, with 55,581 women aged 40–74 included; the primary outcome was screen-detected cancer within 3 months and the primary comparison was one radiologist plus AI against standard double reading by two radiologists, with a non-inferiority margin of a 0.15 relative reduction (Dembrower et al., 2023). The authors report 261 against 250 detected cancers, a relative proportion of 1.04 (95% CI 1.00–1.09), and also found AI alone (246 against 250) and two radiologists plus AI (269 against 250) non-inferior to standard double reading. Cancers missed by every strategy stay unknown until they surface later, so the design measures screen detection, not interval cancers (our reading of the design). The setting was a single hospital.

Prospective observational implementation

PRAIM was embedded in the German programme at 12 sites from July 2021 to February 2023, with 461,818 women in the final analysis (the abstract gives 463,094 screened) read by 119 radiologists, using a CE-marked system that tagged examinations it judged highly unsuspicious as normal and raised a “safety net” alert when a reader called an examination unsuspicious that the system found highly suspicious (Eisemann et al., 2025). An examination entered the AI group if at least one of its two readers used the AI-supported viewer, a choice the radiologists made per examination. That choice was not random, and the authors observed that some radiologists preferred the AI viewer for normal-tagged examinations and a standard viewer for the others, which left more cancers in the control group. The crude rates show the bias: recall was 44.5 per 1,000 in the AI group and 38.4 in the control group, and detection 6.4 and 6.0 per 1,000. After adjusting for reader set and AI prediction by overlap weighting on propensity scores, the model-based detection rate was 6.70 against 5.70 per 1,000, a relative increase of 17.6% (95% CI 5.7–30.8%), and the recall rate 37.4 against 38.3 per 1,000 (−2.5%, 95% CI −6.5% to +1.7%). The authors count the size of the study, its five hardware vendors and its routine setting among its strengths; its estimates rest on the adjustment for confounding, which they supported with a simulation and sensitivity analyses, and interval cancer rates are left to follow-up.

Randomized trials

MASAI randomized examinations after acquisition, by the archive system, between AI-supported reading, in which examinations with the highest risk score (10 on a ten-level scale) were double read and the rest single read, with risk scores shown for all examinations and detection marks for scores 8–10, and standard double reading (Lång et al., 2023). Computer Vision in Radiology summarizes the prespecified safety analysis of the first 80,033 women. The protocol-defined analysis of the secondary outcomes for all 105,934 women randomized between April 2021 and December 2022 reports detection rates of 6.4 against 5.0 per 1,000 (ratio 1.29, 95% CI 1.09–1.51), more invasive cancers (270 against 217), mainly small and node-negative, and more in situ cancers (68 against 45), about half of the increase being high-grade; recall and false-positive rates were not significantly higher, the PPV of recall was higher (ratio 1.19), and screen readings fell by 44.2%, from 109,692 to 61,248 (Hernström et al., 2025). These are early screening outcomes. The primary endpoint, the interval cancer rate, together with sensitivity and specificity, is the subject of a later report in The Lancet (Gommers et al., 2026), whose text could not be read for this article.

Detection rate alone cannot show whether additional detected cancers would otherwise have surfaced before the next screen or would not have caused harm. By our reading, a lower interval cancer rate would speak for the first, and the cancer type, stage and grade that the trial reports alongside detection bear on the same question.

Example: Reading Strategies at Screening Prevalence

The example below simulates a screening population of synthetic image patches and compares reading strategies by the programme metrics above. Each patch is a 32 × 32 image of correlated texture (a stand-in for fibroglandular tissue) with white noise; 35% of patches are dense, with texture 1.8 times stronger, and 0.7% contain a Gaussian blob, a visible lesion. Readers score each patch with the same contrast-to-noise detector (the peak of the smoothed, background-subtracted image, divided by the patch’s own contrast) plus independent reader noise, so they disagree on borderline cases but share the masking by dense texture. The AI score uses the same detector at a coarser scale with no reader noise. All thresholds come from a separate validation population: readers A and B recall 3%, the arbitrator C 1.5%, the AI counts as a positive reader at its top 3% of scores, and its top 10% triggers double reading in the triage strategy. The strategies are single reading, double reading in which either reader recalls, double reading with arbitration (agreement decides, disagreement goes to C), arbitration with C as lenient as A and B, MASAI-like triage without detection marks, and the substitution simulated by McKinney et al. (2020), in which the AI replaces reader B when it agrees with A. It runs with NumPy 2.3.5, SciPy 1.16.3 and scikit-learn 1.7.2 over five seeds in about 35 seconds.

import numpy as np
from scipy.ndimage import gaussian_filter
from sklearn.metrics import roc_auc_score

P, PREV, P_DENSE = 32, 0.007, 0.35        # patch size, lesion prevalence, share dense
C0, C1 = slice(8, 24), slice(8, 24)       # central window searched for a lesion


def make_patches(n, rng):
    """Synthetic 'mammogram' patches: correlated texture, noise, faint blobs."""
    tex = gaussian_filter(rng.standard_normal((n, P, P), dtype=np.float32),
                          sigma=(0, 2, 2))
    tex /= tex.std(axis=(1, 2), keepdims=True)
    dense = rng.random(n) < P_DENSE
    img = tex * np.where(dense, 1.8, 1.0)[:, None, None]
    img += 0.3 * rng.standard_normal((n, P, P), dtype=np.float32)
    lesion = rng.random(n) < PREV
    yy, xx = np.mgrid[0:P, 0:P]
    for i in np.flatnonzero(lesion):
        cy, cx = rng.uniform(8, 24, 2)
        s, a = rng.uniform(1.5, 3.0), rng.uniform(2.0, 5.0)
        img[i] += a * np.exp(-((yy - cy) ** 2 + (xx - cx) ** 2) / (2 * s * s))
    return img, lesion, dense


def blob_score(img, sigma):
    """Remove the local mean, smooth at the lesion scale, take the maximum,
    divided by the image's own high-pass contrast (a contrast-to-noise score)."""
    hp = img - gaussian_filter(img, sigma=(0, 6, 6))
    peak = gaussian_filter(hp, sigma=(0, sigma, sigma))[:, C0, C1].max(axis=(1, 2))
    return peak / hp.std(axis=(1, 2))


def scores(img, rng, ref=None):
    s, s_ai = blob_score(img, 2.0), blob_score(img, 3.5)
    if ref is None:                         # z-score with validation negatives
        ref = (s.mean(), s.std(), s_ai.mean(), s_ai.std())
    z, z_ai = (s - ref[0]) / ref[1], (s_ai - ref[2]) / ref[3]
    readers = z[:, None] + 0.7 * rng.standard_normal((len(z), 3))  # A, B, C
    return readers, z_ai, ref


def metrics(recall, lesion, dense, reads):
    n, tp = len(lesion), (recall & lesion).sum()
    return dict(recall=100 * recall.mean(), cdr=1000 * tp / n,
                sens=100 * tp / lesion.sum(), ppv=100 * tp / max(recall.sum(), 1),
                missed=1000 * (lesion & ~recall).sum() / n, reads=reads / n,
                sens_d=100 * (recall & lesion & dense).sum() / (lesion & dense).sum(),
                sens_nd=100 * (recall & lesion & ~dense).sum() / (lesion & ~dense).sum())


def run(seed, n=60_000, n_val=30_000):
    # Validation population: thresholds and reference statistics only
    rv = np.random.default_rng(seed + 1000)
    img, les_v, _ = make_patches(n_val, rv)
    s_neg = blob_score(img[~les_v], 2.0); a_neg = blob_score(img[~les_v], 3.5)
    ref = (s_neg.mean(), s_neg.std(), a_neg.mean(), a_neg.std())
    rd_v, ai_v, _ = scores(img, rv, ref)
    t_read = np.quantile(rd_v[:, 0], 0.97)        # readers A, B: 3% positive
    t_arb = np.quantile(rd_v[:, 2], 0.985)        # arbitrator C: 1.5% positive
    t_ai = np.quantile(ai_v, 0.97)                # AI as a binary reader: 3%
    t_tri = np.quantile(ai_v, 0.90)               # AI triage: top 10%
    auc = (roc_auc_score(les_v, rd_v[:, 0]), roc_auc_score(les_v, ai_v))

    # Screening population
    rs = np.random.default_rng(seed)
    img, les, dense = make_patches(n, rs)
    rd, ai, _ = scores(img, rs, ref)
    A, B = rd[:, 0] > t_read, rd[:, 1] > t_read
    C = rd[:, 2] > t_arb
    agree = A == B
    arb = np.where(agree, A, C)                   # double reading + arbitration
    out = {}
    out["1 single (A)"] = metrics(A, les, dense, n)
    out["2 double, either recalls"] = metrics(A | B, les, dense, 2 * n)
    out["3 double + arbitration"] = metrics(arb, les, dense, 2 * n + (~agree).sum())
    arb3 = np.where(agree, A, rd[:, 2] > t_read)  # arbitrator as lenient as A, B
    out["3b arbitration, C at 3%"] = metrics(arb3, les, dense, 2 * n + (~agree).sum())
    hi = ai > t_tri                               # AI triage (MASAI-like)
    tri = np.where(hi, arb, A)
    out["4 AI triage, top 10% double"] = metrics(
        tri, les, dense, n + hi.sum() + (hi & ~agree).sum())
    ai_pos = ai > t_ai                            # AI replaces B when it agrees
    rep = np.where(A == ai_pos, A, arb)
    out["5 AI replaces B if agrees"] = metrics(
        rep, les, dense, n + (A != ai_pos).sum() + ((A != ai_pos) & ~agree).sum())
    return out, auc


res = []
for seed in range(5):
    out, auc = run(seed)
    res.append(out)
    print(f"seed {seed}: validation AUC reader {auc[0]:.3f}, AI {auc[1]:.3f}")
cols = ["recall", "cdr", "sens", "ppv", "missed", "reads", "sens_nd", "sens_d"]
heads = ["recall%", "CDR/1k", "sens%", "PPV%", "miss/1k", "reads", "sens ND", "sens D"]
print("\nmeans over seeds 0-4")
print(f"{'strategy':28s}" + "".join(f"{h:>8s}" for h in heads))
for name in res[0]:
    print(f"{name:28s}" + "".join(
        f"{np.mean([r[name][c] for r in res]):8.2f}" for c in cols))
print("\nmin-max over seeds 0-4")
print(f"{'strategy':28s}{'recall%':>13s}{'CDR/1k':>13s}{'sens%':>13s}")
for name in res[0]:
    print(f"{name:28s}" + "".join(
        f"{min(r[name][c] for r in res):6.2f}-{max(r[name][c] for r in res):<6.2f}"
        for c in ["recall", "cdr", "sens"]))

seed 0: validation AUC reader 0.770, AI 0.751
seed 1: validation AUC reader 0.747, AI 0.748
seed 2: validation AUC reader 0.763, AI 0.718
seed 3: validation AUC reader 0.778, AI 0.758
seed 4: validation AUC reader 0.753, AI 0.711

means over seeds 0-4
strategy                     recall%  CDR/1k   sens%    PPV% miss/1k   reads sens ND  sens D
1 single (A)                    3.02    1.99   28.79    6.60    4.92    1.00   37.64   11.61
2 double, either recalls        4.85    2.49   35.91    5.12    4.43    2.00   45.90   16.55
3 double + arbitration          1.70    1.80   25.98   10.55    5.12    2.04   34.54    9.41
3b arbitration, C at 3%         2.16    1.95   28.22    9.05    4.96    2.04   37.27   10.70
4 AI triage, top 10% double     2.34    1.92   27.71    8.18    5.00    1.12   36.48   10.79
5 AI replaces B if agrees       1.93    1.82   26.27    9.42    5.10    1.05   34.81    9.82

min-max over seeds 0-4
strategy                          recall%       CDR/1k        sens%
1 single (A)                  2.89-3.16    1.82-2.35   27.32-32.49 
2 double, either recalls      4.61-5.08    2.30-2.92   34.07-40.32 
3 double + arbitration        1.56-1.83    1.67-1.97   24.89-27.57 
3b arbitration, C at 3%       1.98-2.30    1.82-2.15   26.91-29.72 
4 AI triage, top 10% double   2.23-2.46    1.82-2.20   25.34-30.41 
5 AI replaces B if agrees     1.76-2.07    1.70-2.05   24.20-28.34 

The simulated readers are much weaker than real ones: a reader AUC near 0.76 and a mean single-read sensitivity of about 29%, against the 87% the BCSC reports for US digital mammography. The absolute values therefore say nothing about real screening; what carries over is how the metrics relate and how strategies move them on the same images.

• Detection rate is sensitivity times prevalence. With 7 lesions per 1,000 women, a CDR of about 2 per 1,000 is a sensitivity of about 29%, and the PPV column is ten times the CDR divided by the recall percentage, as the identity above requires (up to averaging over seeds). A screening programme can report CDR and recall rate without knowing its misses; the sensitivity and miss columns need the ground truth that only a simulation, or years of follow-up, provides.
• The rule for combining two readers sets the trade-off. Recalling whenever either reader is positive raises detection and sensitivity, because the second reader’s independent noise catches some lesions the first missed, but it raises recall more, so PPV falls. Arbitration does the opposite: agreement between two noisy readers filters out most false positives, and recall falls from about 3% to 1.7–2.2% while sensitivity stays near the single reader’s. In this model double reading with arbitration buys specificity rather than detection, and where it lands between the two extremes depends on how strict the arbitrator is (rows 3 and 3b). The model’s readers err independently given the image; real readers who see each other’s opinions, as in the UK protocol, do not.
• Density limits every strategy. Sensitivity in dense patches is about a third of that in non-dense ones under all strategies, and adding a reader raises it only a few points, because the masking is in the image and all readers see the same image. Here this follows from the contrast-to-noise detector; real density effects are reported per study, as in PRAIM’s subgroup analyses.
• AI triage and substitution save reads. Both AI strategies land between single reading and double reading with arbitration in recall and detection while using 1.12 and 1.05 human reads per woman instead of 2.04. The saving follows from the design: triage double reads only the top tenth, and substitution sends only disagreements to the second reader.

The missed lesions per 1,000 are not interval cancers. A simulation of one screening round has no cancers that grow between screens and no lesions invisible at screening, which is why trials need follow-up for their primary endpoints.

The first version of the script, run before any change, used 100,000 women and 50,000 validation women per seed, lesion amplitudes between 0.6 and 2.0, and an unnormalized peak score. Its readers were close to chance (validation AUC 0.55–0.57), and because the unnormalized score rises with texture strength, dense patches were recalled more often whether or not they held a lesion, so sensitivity came out higher in dense patches, an artefact of the score rather than a property of reading. Three changes followed: the lesion amplitude range was raised to 2.0–5.0, the score was divided by each patch’s high-pass contrast, and the populations were reduced to 60,000 and 30,000 with single-precision noise to bring the run time from about 65 to about 35 seconds. After that run, the difficulty was not changed again, although the readers remained much weaker than real ones. Two later changes added row 3b and split the printed table into the means and min–max tables shown; the numbers in rows 1–5 were unchanged by them.

Challenges

• Low prevalence and long outcomes. With a few cancers per 1,000 examinations, a test set needs tens of thousands of women to contain a few hundred cancers, and the reference standard takes years of follow-up. Enriched sets are cheaper but need reweighting, as McKinney et al. (2020) did, before any population metric is reported.
• Acquisition differences. Vendor processing and the dominance of one manufacturer in a dataset limit what a result shows about other equipment (McKinney et al., 2020), and DBT changes the input altogether. PRAIM reported its results by hardware vendor because its sites used five (Eisemann et al., 2025).
• Populations. Age ranges, intervals and density distributions differ between programmes; VinDr-Mammo’s breasts are mostly dense (Nguyen et al., 2023), while a third of PRAIM’s women had dense breasts (Eisemann et al., 2025), so results on one population describe its case mix; Dataset Bias covers the general problem.

Failure Modes and Risks

• Outcome blindness of retrospective data. Cancers seen only by the AI are not confirmed in archival data unless they surface later, so retrospective studies tend to undercount what a system would find and cannot measure the cost of its extra recalls (Eisemann et al., 2025).
• Reader behaviour. How radiologists use a system is part of the result. In PRAIM, readers chose per examination whether to use the AI viewer; the crude recall rate was higher in the AI group and the adjusted one lower, and the authors note that their data come from an early learning phase in which behaviour may change (Eisemann et al., 2025).
• Detection without outcome. Higher detection rates can include cancers that would not have caused harm; MASAI reports cancer type, size, node status and grade, and its primary endpoint is the interval cancer rate (Hernström et al., 2025; Lång et al., 2023).

Deployment Considerations

AI in screening is deployed inside a reading protocol, and the protocol decides what it can change: a second reader replaced, a triage that decides between one and two readers, a normal-tagging and safety-net viewer, or marks shown to a single reader. Each changes workload and metrics differently, as the studies above and the example show. In the United States, the FDA’s 2020 reclassification of medical image analyzers, including mammography CADe, is summarized in Computer Vision in Radiology, and facilities are already required by the MQSA rule to audit PPV, detection rate and recall rate every year (FDA, 2023), which gives a ready frame for monitoring a deployed system. Statistical comparisons between strategies need confidence intervals that account for the low number of cancers, as Statistical Significance in Evaluation explains, and threshold choices should be made on validation data from the target programme, as McKinney et al. (2020) did for each human benchmark; their UK-only model, though it still exceeded the US readers, did so by a smaller margin, and they suggest fine-tuning with local data. Workload saving, the main operational motive, can be read directly from screening records, while the benefit to women can only be read from outcomes that take years to accrue.