Introduction
Methods
Eligibility criteria
Population
Intervention
Comparison
Outcomes
Information sources
Search strategy
('breast density'/exp OR ((breast NEAR/3 densit*):ti,ab,kw OR (mammary NEAR/3 densit*):ti,ab,kw OR (mammographic NEAR/3 densit*):ti,ab,kw)) AND ('mammography'/deOR mammograph*:ti,ab,kwOR mammogram*:ti,ab,kwOR mastrography:ti,ab,kwOR ‘digital breast tomosynthesis’:ti,ab,kwOR ‘x-ray breast tomosynthesis’:ti,ab,kw)NOT('editorial'/it OR 'letter'/it OR 'note'/it) AND [english]/lim
Selection process
Data collection process
Data items
-
the report: author, publication year
-
the study: location/institution, number of cases, number of controls, study design
-
the research design and features: lapsed time from mammogram to diagnosis
-
the mammogram: machine type, mammogram view(s), breast(s) used for analysis
-
the model: how density was measured, number of texture features extracted, types of texture features extracted, whether feature extraction was machine or human, whether all features were used in the analysis, how features for analysis were chosen, non-mammogram covariates included, established confounders for density, prediction horizon, statistical methods for assessing risk association
Risk of bias
Human subjects
Registration and protocol
Results
References | Year | City/institution | Study design | Machine type | View used (CC/MLO/both) | # cases | # controls |
---|---|---|---|---|---|---|---|
Choi et al. [28] | 2016 | University of Ulsan | Retrospective cohort | General Electric Senographe DS and film | Both | 240 | 0 |
Malkov et al. [38] | 2016 | USA | Case–control studies with 2 of 5 nested in cohorts | Film | CC | 1171 | 1659 |
Tan et al. [43] | 2016 | University of Pittsburgh Medical Center | Case–control | Digital (not specified further) | Both | 159 | 176 |
Winkel et al. [49] | 2016 | Bispebjerg Hospital | Prospective cohort with case–control sampling | Film | Both | 121 | 259 |
Ali et al. [26] | 2017 | Sweden | Main analysis: case–control. Validation study: prospective cohort | Film for main analysis, GE Senographe Essential for validation study | MLO | 1170 for main analysis, 69 for validation study | 1283 for main analysis, 231 for validation study |
Eriksson et al. [29] | 2017 | Sweden | Prospective cohort with case–control sampling | Digital (not specified further) | Both | 433 for model development. An additional 137 women lacking information were included in calculating the absolute risk estimates | 1732 when comparing study participant characteristics and mammographic features, 60,237 for calculating the absolute risk estimate |
Wang et al. [45] | 2017 | National Health Service | Prospective cohort with case–control sampling | GE Senographe system | CC | 264 for training case–control study, 317 for validation case–control study | 787 for training, 931 for validation |
Winkel et al. [48] | 2017 | Copenhagen, Denmark | Prospective cohort of false positives with case–control sampling | Film | Both | 288 | 288 |
Yan et al. [50] | 2017 (August) | NR | Case–control | Hologic Selenia | CC | 83 | 85 |
Yan et al. [51] | 2017 (October) | NR | Case–control | Hologic Selenia | CC | 83 | 85 |
Gastounioti et al. [31] | 2018 | University of Pennsylvania | Case–control | Hologic Selenia Dimensions | MLO | 115 | 460 |
Heidari et al. [32] | 2018 | NR | Case–control | Digital (not specified further) | CC | 250 | 250 |
Li et al. [37] | 2018 | Fudan University Shanghai Cancer Center | Nested case–control prospective cohort | NR | CC | 84 | 987 |
Schmidt et al. [40] | 2018 | Australia and Hawaii, USA | Case–control and nested case–control in prospective cohort | Film | CC | 1236 | 2931 |
Tagliafico et al. [42]* | 2018 | Italy | Case–control | Hologic Selenia Dimensions | NR | 20 | 20 |
Ward et al. [46] | 2018 | National Health Service | Prospective cohort | Hologic Selenia | Both | 34 | 746 |
Evans et al. [30] | 2019 | Bradford (UK) Teaching Hospitals NHS Foundation Trust | Case–control | Digital (not specified further) | Mix of CC and MLO | 58 images from 35 patients with cancer. Experiment 1D included an additional 50 abnormal mammograms with visible cancerous lesions taken from 50 patients | 58 images from 35 patients without cancer. Experiment 1D included an additional 50 normal mammograms taken from 50 patients |
Hsu et al. [34] | 2019 | University of California, Los Angeles | Prospective cohort | NR | NR | 463 biopsy results | 1675 biopsy results |
Kontos et al. [35] | 2019 | University of Pennsylvania for case–control sample for evaluating associations to breast cancer, NR for screening sample for phenotype identification | Sample used in evaluating associations to breast cancer: case–control. Sample used for phenotype identification: cross-sectional | Hologic Selenia Dimensions for screening sample, GE Senographe 2000D and DS for case–control sample | Both for screening sample, NR for case–control sample | Screening sample included 18 detected cases with 12 in the training sample and 6 in the testing sample. 76 cases were in case–control sample | Screening sample included 2011 controls with 1327 in the training sample and 684 in the testing sample. 158 controls were in case–control sample |
Perez-Benito [39] | 2019 | Valencian Community, Spain | Case–control from population screening program | FUJIFILM, IMS s.r.l., Giotto IRE, HOLOGIC, SIEMENS, or unknown | Both | 808 cases with 606 in training set and 202 in test set | 755 with 566 in training set and 189 in test set |
2019 | Tampere University Hospital | Case–control | Philips MicroDose SI or General Electric Senographe Essential | CC | 114 | 114 | |
Tan et al. [44] | 2019 | Subang Jaya Medical Center | Case–control | Hologic Selenia | CC | 250 | 250 |
Abdolell et al. [25] | 2020 | NR | Case–control from population screening program | Siemens MAMMOMAT Inspiration or MAMMOMAT Novation DRimaging system | Both | 1882 | 5888 |
Ma et al. [36] | 2020 | Nanfang Hospital | Case–control | Hologic Selenia Dimensions | Both | 608 for risk model development, 203 for validation | 1261 for risk model development, 421 for validation |
Sorin et al. [41] | 2020 | NR | Retrospective cohort | GE Senographe Essential | Both | 53 | 463 |
Azam et al. [27] | 2021 | Sweden | Prospective cohort | General Electric, Philips, Spectrum Hologic, and Siemens | Both | 676 | 52,597 |
Heine et al. [33] | 2021 | Mayo Clinic for first case–control study, the San Francisco Mammography Registry for generalization study | Case–control | Hologic Selenia | CC | 514 for first study, 1474 for generalization study | 1377 for first study, 2942 for generalization study |
Warner et al. [47] | 2021 | USA | Prospective cohort with case–control sampling | Film | CC | 1900 | 3921 |
References | Time from mammogram to cancer diagnosis | Side used | Density (BIRAD categories/continuous) | Number of texture features extracted (other than density) | Types of texture features extracted (list all) | Machine or human extraction | All texture features used in the model (yes/no) | How features for analysis are chosen |
---|---|---|---|---|---|---|---|---|
Choi et al. [28] | Mean = 9.7 months (range 6–15 months) | NR | BIRADS | 8 | Normal appearing tissue, benign-appearing calcification, mass, calcification, architecture distortion, focal asymmetry | Human | Yes | N/A |
Malkov et al. [38] | Mean = 5.1 years | Avg | Cumulus and custom software comparable to Cumulus | 46 | First- and second-order features, Fourier transform, and fractal dimension analysis | Machine | No | AUCs for each feature individually given |
Tan et al. [43] | The average elapsed time between the “current” and each of “prior” #1, #2 and #3 studies was 1.16 ± 0.41, 2.30 ± 0.55 and 3.44 ± 0.72 years, respectively | Both | Computer-aided detection scheme | 158 | Mammographic density, structural similarity, and texture based image features | Machine | No | Stepwise regression analysis |
Winkel et al. [49] | Average = 26 months (range 4–45 months) | Both | BIRADS | 1 | Tabár classification of parenchymal patterns | Human | Yes | N/A |
Eriksson et al. [29] | Median = 1.74 years | Both | BIRADS and STRATUS | 2 | Calcifications and masses | Machine | Yes | N/A |
Winkel et al. [48] | Average = 82.0 months, median = 75.5 months, range = 5 to 192 months | Both | BIRADS and Cumulus-like approach | 1 | Tabár classification of parenchymal patterns | Human | Yes | N/A |
Yan et al. [50] | The interval between the prior (negative) and current (cancer detected) examinations are 410.0 ± 51.7 days for cases | Both | BIRADS | 148 | Bilateral mammographic tissue asymmetry maximum features | Machine | No | WEKA data mining and machine learning software package |
Yan et al. [51] | 12–36 months | Both | Mutual threshold | 220 | Asymmetry, mean and maximum features | Machine | No | WEKA data mining and machine learning software package |
Gastounioti et al. [31] | Average = 1.9 years ± 0.7. Cases had negative screening mammograms at least one year prior to their diagnosis | Avg | BIRADS, LIBRA, and Quantus | 34 | Anatomically oriented texture features | Machine | No | Identified pairs of features with absolute Pearson correlation greater than 0.90 and for each pair removed the feature with the lowest variability in terms of its interquartile range. Starting from the remaining features, elastic net regression with nested cross-validation was used to build a parsimonious logistic regression model with the most discriminatory subset of covariates |
Heidari et al. [32] | The time interval between the “prior” and “current” mammography screenings ranged from 12 to 18 months | Both | BIRADS | 44 | Bilateral asymmetry of mammographic tissue density distribution | Machine | No | Locally preserving projection based feature combination algorithm |
Li et al. [37] | NR | Both | AutoDensity | 1 | Breast area | Machine | Yes | N/A |
Schmidt et al. [40] | Melbourne: Cases were diagnosed, on average, 8 years after baseline interview (range, 3 months–16 years), and mammography was performed, on average, 2.8 years (standard deviation, 2.6 years; range, 0–14 years) after baseline. Australia: average 4 years for 32% of cases, and for the other affected women we used the mammogram from the opposite side to that in which the cancer was diagnosed. Hawaii: The mean time between the earliest mammogram and the breast cancer diagnosis was 6.3 years, while the earliest and the latest mammogram were, on average, 5.1 years apart for cases | Avg | Cumulus | 20 | Gray-level co-occurrence matrix textural features | Machine | Yes | N/A |
Tagliafico et al. [42] | Cancer was detected at tomosynthesis | Avg | BIRADS | 104 | Radiomics features including skewness, energy, entropy, kurtosis, 90 percentile and dissimilarity | Machine | No | Selected to reduce the risk of over-fitting and according to features previously used to associate breast parenchymal patterns with cancer risk |
Ward et al. [46] | NR | Both | Volpara | 1 | Patterns of parenchymal tissue | Human | Yes | N/A |
Evans et al. [30] | Mammograms used were acquired 3 years prior to the mammograms that had revealed visible and actionable cancer | Both | Scale similar to BIRADS used | 1 | Non-localizable global gist signal | Human | Yes | N/A |
Hsu et al. [34] | A false-positive biopsy recommendation was defined by the lack of cancer within 1 year of the screening examination | NR | BIRADS | 5 | Presence of lump, mass, calcification, architecture distortion, asymmetry | Human | No | Presence of lump included in final model. Mass, calcifications, architecture distortion, asymmetry examined individually with PPV values given |
Kontos et al. [35] | For screening sample, within 1 year. Not specified for case–control sample | Avg | BIRADS and LIBRA | 29 | Phenotypes of mammographic parenchymal complexity based on four main types of features: histogram, co-occurrence, run-length, and structural | Machine | No | Excluded features with extremely low variation and those with extreme skewness |
Abdolell et al. [25] | NR | NR | Densitas | 1 | Breast volume | Machine | Yes | N/A |
Ma et al. [36] | At least 1 year later for validation | Both | BIRADS | 1 | Normalized average glandular dose | Machine | Yes | N/A |
Sorin et al. [41] | Cancer cases were defined as all cancers detected at the time of contrast-enhanced spectral mammography imaging as well as cancers diagnosed during the follow-up period. Controls had at least 1-year follow-up | Both | BIRADS | 1 | Background parenchymal enhancement | Human | Yes | N/A |
Azam et al. [27] | The median number of years between the last negative mammogram and the date of diagnosis was 2.8 | Both | STRATUS | 1 | Microcalcification clusters | Machine | Yes | N/A |
Heine et al. [33] | At least 6 months | Avg | Volpara | 4 | Variation measures | Machine | No | The two variants of V produced similar findings so only one was discussed in the results |
Warner et al. [47] | Median = 4.1 years | Both | Cumulus and automated computer algorithm | 1 | Gray-scale variation | Machine | Yes | N/A |
References | Time from mammogram to cancer diagnosis | Side used | Density (BIRAD categories/continuous) | Number of texture features extracted (other than density) | Types of texture features extracted (list all) | Machine or human extraction | All texture features used in the model (yes/no) | How features for analysis are chosen |
---|---|---|---|---|---|---|---|---|
Ali et al. [26] | Less than 3 years before diagnosis (and at latest, at date of diagnosis) | Contralateral for cases, random side chosen for controls | Cumulus and automated measure of area PD | 13 | Spatial organization of dense vs. fatty regions of the breast | Machine | Yes for AUC given, no for further analysis | Stepwise selection procedure |
Wang et al. [45] | Training study: diagnosed at the same time as mammogram. Validation study: average = 3.0 years | Training: contralateral for cases and the left for controls. Validation: contralateral for cases and the same side for controls | Volpara | 112 | Features based on a gray-level co-occurrence matrix, neighborhood gray-tone difference matrix, form and shape of breast boundary, run-length, and gray-level size zone matrix, and statistical moments of pixel values | Machine | No | Selected from training set using least absolute shrinkage and selection operator |
Perez-Benito [39] | NR | Contralateral | DMScan | 23 | Geometrical features and a global feature based on local histograms of oriented gradients | Machine | Yes | N/A |
NR | Contralateral for cases, right for controls | Cumulus-like approach | 37 | Parenchymal features including computational features and imaging parameters related to the mammographic system (compressed breast thickness, compression force, X-ray tube voltage peak and target–filter combination) | Machine | Yes | N/A | |
Tan et al. [44] | Within a year | Contralateral | Volpara | 944 | Gray-level co-occurrence matrix features, structural/pattern measures, gray-level intensity/histogram features, run-length features, and multiresolution/spectral features | Machine | No | Stepwise regression analysis |
References | Non-mammogram covariates included (e.g., age, parity, etc.) | Established confounders for density adjusted for | Prediction horizon (< 5 year/5 year/10 year) | AUC (baseline model) | Overall AUC (with texture features added) |
---|---|---|---|---|---|
Malkov et al. [38] | Adjusted for age, body mass index, first-degree family history, percent density, study | Age and BMI | NR | 0.617 | 0.621 |
Tan et al. [43] | None | Age | NR | NR | 0.730 |
Winkel et al. [49] | Adjusted for age | Age | NR | BIRADS density = 0.63 | 0.65 |
Eriksson et al. [29] | Percentage mammographic density, age at mammography, BMI, family history of breast cancer, HRT use | Age, BMI, HRT use, and menopausal status | 2 years (for main model) and 3 years (relative risks given) | 0.64 | 0.71 with density and interaction between percentage density and masses also included |
Winkel et al. [48] | Adjusted for birth year, age at false-positive screen, invitation round at false-positive screen | Age | NR | BIRADS density = 0.65, percentage mammographic density = 0.62 | 0.63 |
Yan et al. [50] | None | None | Next sequential mammography screening | NR | 0.816 |
Yan et al. [51] | None | None | Next sequential sequencing | NR | 0.830 |
Gastounioti et al. [31] | Density, BMI, age | Age and BMI | NR | 0.62 | 0.67 |
Heidari et al. [32] | None | Age | 12 to 18 months | NR | 0.68 |
Schmidt et al. [40] | Adjusted for age, BMI | Age and BMI | NR | Percent mammographic density = 0.620 | 0.662 |
Tagliafico et al. [42] | None | Age | NR | NR | 0.567 |
Evans et al. [30] | None | None | 3 years | NR | 0.54 |
Hsu et al. [34] | BIRADS, density, age, race, BMI, age at first live birth, noticeable changes in breast, number of risk factors, 5-year Gail risk ≥ 1.67% | Age and BMI | NR | cv-AUC = 0.83 with BIRADS and density only | cv-AUC = 0.84 |
Kontos et al. [35] | Density, BMI | Age and BMI | NR for AUC model. 5-year risk from Gail and Breast Cancer Surveillance Consortium models compared by phenotype | 0.80 | 0.84 |
Abdolell et al. [25] | Age, percent mammographic density | Age | Tailored estimates of current breast cancer risk | 0.584 | 0.597 |
Ma et al. [36] | Age, age at menarche, menopausal status, age at first birth, parity, family history of breast cancer, breast density | Age and menopausal status | NR | 0.61 for training set and 0.56 for test set | 0.77 for training set and 0.75 for test set |
Heine et al. [33] | Adjusted for study, age, body mass index and also with dense volume included | Age and BMI | NR | Volumetric breast density = 0.61 for first study and 0.59 for generalization study | V = 0.61, P1 = 0.61, p1 = 0.60 for first study. V = 0.59, P1 = 0.57, p1 = 0.58 for generalization study |
References | Non-mammogram covariates included (e.g., age, parity, etc.) | Established confounders for density adjusted for | Prediction horizon (< 5 year/5 year/10 year) | AUC (baseline model) | Overall AUC (with texture features added) |
---|---|---|---|---|---|
Ali et al. [26] | Age, BMI, density, HRT status, parity, age at first birth | Age, BMI, HRT use, and menopausal status | NR | 0.687 for apparent, 0.634 for honest | 0.703 for apparent, 0.643 for honest |
Wang et al. [45] | Adjusted for age, BMI | Age, BMI, HRT use, and menopausal status | NR | mC = 0.57 | mC = 0.58 |
Perez-Benito [39] | Percent density | Age | NR | 0.560 | 0.614 |
None and with age, percent density | Age | NR | Density only = 0.609 | 0.786 | |
Tan et al. [44] | None, with age only, and with age and BMI | Age and BMI | NR | Density = 0.52 | 0.68 |
References | Non-mammogram covariates included (e.g., age, parity, etc.) | Established confounders for density adjusted for | Prediction horizon (< 5 year/5 year/10 year) | Risk other than AUC | Statistical methods for estimating the association between features and risk |
---|---|---|---|---|---|
Choi et al. [28] | N/A | None | N/A | In the minimal sign group, the most common finding was normal appearing tissue (61/78), followed by benign-appearing calcification (17/78) | Chi-square test or Fisher’s exact test |
Li et al. [37] | None | None | NR | OR = 1.018 (95% CI = 1.004–1.033) | t test, chi-square test, and binary logistic regression |
Ward et al. [46] | None | None | NR | There was a significant correlation between a diagnosis of cancer and nodular parenchymal pattern compared to not nodular parenchymal pattern (p = 0.043) | Pearson’s chi-squared test with Yate’s continuity correction |
Sorin et al. [41] | Adjusted for age, family history, breast density | Age | NR | The odds for breast cancer significantly increased with increased background parenchymal enhancement (OR = 2.24, 95% CI = 1.23–4.09) | Binary logistic model for generalized estimating equation |
Azam et al. [27] | Adjusted for BMI, baseline mammographic density, smoking status, alcohol consumption, age at menarche, age at first birth, number of children, breastfeeding duration, oral contraceptive use, menopausal hormone therapy use, family history of breast cancer | BMI and HRT use | NR | Each additional microcalcification cluster was associated with 20% increased risk of breast cancer (HR = 1.20, 95% CI = 1.13–1.28). Women with ≥ 3 microcalcification clusters had an overall twofold increased risk of breast cancer compared to women with no clusters (HR = 2.17, 95% CI = 1.57–3.0) | Cox proportional hazard regression |
Warner et al. [47] | Adjusted for age, fasting status, time of blood draw, body mass index, menopausal status, hormone therapy use, mammography read batch and also with either percent mammographic density or automated percent density | Age, BMI, HRT use, and menopausal status | NR | V was positively associated with breast cancer risk (OR = 1.15, 95% CI = 1.08–1.23 with percent mammographic density, 1.18, 95% CI = 1.07–1.31 with automated percent density) | Unconditional logistic regression |