Pith. sign in

REVIEW 3 major objections 4 minor 43 references

Adding cancer-enriched external mammograms to a screening cohort's training data consistently lowered screening performance in every configuration tested.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 23:25 UTC pith:UU6GWAEC

load-bearing objection A credible negative result for pooling abnormal-enriched mammography datasets with a screening cohort, but the attribution to dataset-origin shortcuts is undercut by an uncontrolled dilution effect in the training sampler. the 3 major comments →

arxiv 2607.15416 v1 pith:UU6GWAEC submitted 2026-07-16 cs.CV

Dataset-Origin Signatures and Shortcut Learning in Screening Mammography AI: A Cross-Dataset Case Study

Parham Hajishafiezahramini , Matthew Hamilton , Oscar Meruvia-Pastor , Edward Kendall This is my paper
classification cs.CV
keywords screening mammographydataset mixingdomain shiftshortcut learningdataset-origin classificationabnormal-enriched datasetslinear probeMammo-CLIP
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests a common assumption in medical AI: that supplementing a real screening dataset with biopsy-confirmed abnormal cases from public mammography datasets will improve detection. It finds the opposite. Using a frozen mammography-pretrained encoder with a linear probe on the Newfoundland and Labrador Breast Screening Dataset, the screening-only model reached 0.737 AUC, while every configuration that added external positives fell to 0.620–0.644, with the drop growing as more external sources were added. A three-class diagnostic that predicts each image's dataset of origin separated the datasets almost perfectly (macro AUC 0.9998) despite identical preprocessing, showing that dataset-specific signatures survive normalization. The authors conclude that naïve pooling introduces domain shift and dataset-origin shortcuts that outweigh the benefit of more positive examples, motivating domain-aware integration strategies.

Core claim

Central claim: holding architecture, preprocessing, and training procedure constant, adding abnormal-enriched external positives to a real screening cohort does not improve screening performance and consistently degrades it. Evidence is a monotonic AUC drop from 0.737 (NLBSD-only) to 0.620 (both external sources added), significant after Holm correction in every mixed configuration. A three-class dataset-origin diagnostic separates NLBSD, CBIS-DDSM, and CMMD almost perfectly (macro AUC 0.9998) despite identical preprocessing, showing that dataset-specific signatures survive normalization and can act as shortcuts. The authors conclude that naïve pooling introduces domain shift that outweighs

What carries the argument

The central mechanism is the dataset-origin classification diagnostic: a three-class experiment that relabels each image by its source dataset instead of its clinical label, run under the same frozen EfficientNet-B5 encoder, preprocessing, augmentation, and balanced batch sampling. It reveals whether dataset-specific structure remains after normalization, and here it separates the three datasets almost perfectly. The binary experiments use a frozen linear probe, so differences across training configurations cannot come from adapting the encoder—isolating the effect of dataset composition.

Load-bearing premise

The conclusion that domain shift rather than a training-batch dilution effect causes the decline assumes that fixing the total number of positives per batch at 20 while letting external datasets contribute most of those positives is neutral; if reduced exposure to NLBSD positives alone explains the drop, the dataset-origin shortcut may not be the primary cause.

What would settle it

Retrain the three mixed configurations while holding the number of NLBSD positives per batch fixed at the level of the NLBSD-only baseline (for example, 20 NLBSD positives per batch, with external positives added outside that quota). If the AUC on the fixed NLBSD test set recovers to roughly 0.737, dilution is the driver; if the AUC still drops to the 0.620–0.644 range, the dataset-origin shortcut explanation is supported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Training a screening-triage model on the screening cohort alone yields higher AUC than mixing in abnormal-enriched public datasets; adding more external sources makes the degradation larger.
  • Dataset-origin signatures survive identical DICOM handling, per-image normalization, cropping, and laterality alignment, so standard preprocessing does not neutralize domain shift.
  • A model can learn to identify an image's source almost perfectly, making dataset-origin shortcuts a plausible mechanism for negative transfer in medical imaging.
  • At a fixed threshold, mixed training drops sensitivity to 0.53–0.55, moving away from the high-sensitivity regime that screening triage requires.
  • Effective multi-dataset mammography training will likely need domain-aware strategies such as intensity harmonization, site-conditioned models, or domain adaptation rather than naïve pooling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The mixed-training batches always contain 20 positives, but external datasets supply far more positive images than NLBSD's 149 suspicious exams; the observed drop could partly reflect reduced exposure to NLBSD positives per batch rather than domain shift alone. A variant that fixes the NLBSD-positive quota per batch would separate dilution from shortcut learning.
  • The near-perfect origin separability suggests a cheap pre-merge diagnostic: train a small source classifier on any planned multi-dataset combination; near-perfect separability predicts negative transfer unless domain alignment is applied.
  • Because the encoder was frozen, fine-tuning may change the size or direction of the effect; testing the same mixing protocol with fine-tuning would reveal whether the shortcut phenomenon is robust beyond linear probes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper tests whether supplementing a real screening cohort (NLBSD) with biopsy-confirmed abnormal cases from CBIS-DDSM and CMMD improves a frozen EfficientNet-B5 linear-probe classifier for screening mammography. On a fixed NLBSD test set, the NLBSD-only model achieves AUC-ROC 0.737; adding external positives lowers AUC to 0.620–0.644, with Holm-corrected DeLong p < 0.05 and monotonic degradation as more sources are added. A three-class dataset-origin diagnostic separates the three datasets almost perfectly (macro-average AUC 0.9998). The authors conclude that naive pooling introduces domain shift and dataset-origin shortcuts that outweigh the benefit of additional positives. The main concern is that the mixed-training batches may contain very few NLBSD positives, so the observed decline could reflect dilution rather than domain shift, and the origin diagnostic does not directly show that the binary head uses origin features.

Significance. If the core finding survives a properly controlled exposure experiment, it is a useful and credible negative result: it documents that simply adding abnormal-enriched external data can harm screening-cohort performance, contrary to a common assumption. The study has clear strengths: fixed patient-level splits, identical preprocessing across datasets, a frozen mammography-pretrained encoder, bootstrap confidence intervals, Holm-corrected DeLong tests, and a common decision threshold. The origin-separability result is a clean demonstration that dataset-specific structure persists after identical preprocessing, which is relevant to dataset-pooling claims. There is no circularity: the origin diagnostic is an independent observation, not a quantity used to fit the binary model. However, the paper's central mechanism—shortcut learning—requires a controlled exposure test and direct evidence in the trained head.

major comments (3)
  1. [§2.4 and §2.5 (Phase A)] The balanced batch sampler fixes only the total number of positives per batch (20), not the number drawn from NLBSD. In the mixed phases the positive pool contains roughly 149 NLBSD suspicious exams (after the 70/10/20 split, about 104 patients) plus thousands of CBIS-DDSM/CMMD abnormal images (1,644 cases; 3,712 images), while negatives are always NLBSD. If the 20 positives are sampled from this combined pool, the expected number of NLBSD positives per batch is very small, so the linear head may see few NLBSD positives over 30 epochs. The monotonic AUC drop (0.737→0.644→0.640→0.620) is then also consistent with a dilution/curriculum effect rather than the model learning dataset-origin shortcuts. Please (i) report the actual origin composition of positives per batch in each configuration, and (ii) rerun Phase A with a stratified sampler that holds NLBSD-positive exposure fixed, or subsam
  2. [§3.4 and §4.1] The three-class origin diagnostic does not establish the mechanism named in the title. It shows that the frozen Mammo-CLIP features separate datasets almost perfectly, which is a property of the fixed encoder, not of the binary head trained in Phase A. A linear probe could also fail on NLBSD positives simply because they are under-represented in its batches (§2.4); near-perfect origin separability in feature space does not prove that the head uses those features. To support 'shortcut learning,' test the trained binary model directly: correlate binary errors with predicted origin, remove origin-discriminative directions, or compare against a matched-exposure model. Without such evidence, the diagnostic is an interesting observation but does not carry the causal load assigned to it.
  3. [Table 2, bottom block / §3.3] The Phase B comparison is not controlled because the test set changes across rows. The NLBSD-only row is evaluated on the NLBSD test set, whereas mixed rows are evaluated on NLBSD∪CMMD, NLBSD∪CBIS-DDSM, or the union. Lower AUC in the mixed rows (0.653, 0.651, 0.648 vs 0.737) therefore does not support the abstract's statement that 'no configuration surpassed the NLBSD-only model' under domain-matched evaluation; the NLBSD-only model would need to be scored on the same mixed test sets. This issue does not affect the primary Phase A conclusion, but the abstract overstates a supplementary result.
minor comments (4)
  1. [Table 2] Dataset abbreviations are inconsistent: the table uses 'DDSM+' while the text uses 'CBIS-DDSM+'. Please standardize.
  2. [§2.7] Case-level results are described as 'reported as a supplementary analysis,' but no case-level table or appendix appears in the manuscript. Either include the numbers or remove the statement.
  3. [§3.2 and §2.7] The image-level DeLong p-values are computed on four correlated views per exam. The authors acknowledge this, but a patient-level clustered bootstrap or case-level DeLong would make the statistical claim more robust; please provide the case-level values.
  4. [§2.2 / §2.3] The per-image min-max normalization is applied independently to each image. This is a defensible choice, but a sentence explaining why a global or dataset-level normalization was avoided would help readers assess the 'identical preprocessing' claim.

Circularity Check

0 steps flagged

No significant circularity: the reported comparisons are direct empirical measurements, and the dataset-origin diagnostic is an independent observation rather than a fitted input to the binary screening model.

full rationale

The paper's central claims are direct empirical measurements, not derivations. The NLBSD-only AUC (0.737) and the mixed-training AUCs (0.620–0.644) are observed values produced by the described training and evaluation procedures. The three-class dataset-origin experiment is an independent diagnostic: the labels used are dataset source, which is not an input to the binary screening model, and the diagnostic is not used to fit, select, or tune the binary classifier. No fitted parameter is renamed as a prediction, and no equation reduces one claimed result to another by construction. Self-citations to the NLBSD dataset papers [11,24] and prior NLBSD-related analyses [15,16] provide the data and background context, not the load-bearing result; the dataset is public, and the measurements are reproducible outside the present paper. The dilution/curriculum concern raised in review is a limitation on causal attribution (whether the observed drop is due to domain shift or to reduced exposure to NLBSD positives), not a circularity: the reported AUC values would remain exactly the same under either interpretation. Because no circular step can be exhibited with specific quotes and reductions, the appropriate score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central empirical comparison is largely self-contained, but its causal interpretation leans on unverified label equivalence across datasets, a single frozen encoder, and a training-batch design that lets external positives dominate the positive pool. No new entities are introduced; the 'dataset-origin signature' is an observed property, not a postulated mechanism.

free parameters (4)
  • Positive-batch origin mix = 20 positives per batch, no NLBSD quota
    Manual choice in §2.4; because external positives vastly outnumber NLBSD positives, mixed-training batches contain mostly external positives, so the design cannot separate 'more positives' from 'different positives'.
  • Per-image min-max normalization = [0,1] per image
    Choice of normalization; paper argues it doesn't remove domain differences, but the comparison assumes this is the right harmonization baseline.
  • Decision threshold = selected on baseline, applied to all models
    §2.7/§3.2; threshold choice affects sensitivity/specificity comparisons.
  • Frozen linear probe hyperparameters = lr=3e-5, wd=1e-4, batch=80, dropout=0.3, patience=10, epochs=30
    Hand-chosen; single architecture/optimizer, so results may not generalize to other training regimes.
axioms (5)
  • domain assumption NLBSD normal/suspicious labels are correct and clinically meaningful for screening
    Section 2.1.1; the entire evaluation depends on these labels being accurate and representing the target screening task.
  • domain assumption Biopsy-confirmed labels in CBIS-DDSM/CMMD correspond to the same abnormality concept as NLBSD recall-suspicious labels
    Sections 2.1.2–2.1.3; if the label semantics differ, the experiment tests label mismatch rather than domain shift alone.
  • domain assumption Mammo-CLIP frozen features provide a valid representation for detecting screening abnormalities
    Section 2.4; a single frozen encoder is used, so results may be encoder-specific.
  • domain assumption The DICOM LUT, segmentation, and normalization pipeline is applied correctly and identically across datasets
    Section 2.2; the conclusion that dataset signatures survive preprocessing depends on this implementation being correct.
  • standard math DeLong test and stratified bootstrap assumptions hold
    Section 2.7; authors note the correlated-view caveat, which weakens the p-values but not necessarily the ordering.

pith-pipeline@v1.3.0-alltime-deepseek · 13463 in / 13308 out tokens · 127982 ms · 2026-08-01T23:25:18.369056+00:00 · methodology

0 comments
read the original abstract

Reliable AI for screening mammography requires training data representative of the low cancer prevalence and subtle abnormalities found in screening populations. We examined whether supplementing such data with biopsy-confirmed cases from abnormal-enriched external datasets improves performance. Using the Newfoundland and Labrador Breast Screening Dataset (NLBSD) alongside CBIS-DDSM and CMMD, we evaluated an EfficientNet-B5 encoder initialized with Mammo-CLIP weights as a frozen linear probe under consistent preprocessing and patient-level splits. The NLBSD-only model achieved an AUC-ROC of 0.737 (95% CI [0.686, 0.785]). Adding external positive cases reduced performance in every configuration (AUC-ROC = 0.620--0.644; DeLong test, Holm-corrected $p < 0.05$), with degradation increasing as additional sources were introduced. Domain-matched evaluation produced modest gains only when the training and test domains coincided, and no configuration surpassed the NLBSD-only model. As a diagnostic, we reframed the task as predicting each examination's dataset of origin. The datasets were separated almost perfectly despite identical preprocessing, indicating that dataset-specific characteristics strongly influence the learned representation. These findings show that na\"ively pooling abnormal-enriched mammography datasets can introduce domain shift that outweighs the benefit of additional positive cases. Differences in acquisition, intensity mapping, and dataset construction persist after normalization, motivating domain-aware strategies for combining heterogeneous mammography datasets.

Figures

Figures reproduced from arXiv: 2607.15416 by Edward Kendall, Matthew Hamilton, Oscar Meruvia-Pastor, Parham Hajishafiezahramini.

Figure 1
Figure 1. Figure 1: Overview of the experimental pipeline, including data preprocessing, patient [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Image-level AUC with 95% bootstrap confidence intervals for each training con [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Normalized confusion matrix for the three-class dataset-origin experiment. Rows [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 1 canonical work pages

  1. [1]

    H. Sung, J. Ferlay, R. L. Siegel, M. Laversanne, I. Soerjomataram, A. Je- mal, F. Bray, Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA Cancer J. Clin. 71 (3) (2021) 209–249

    2020

  2. [2]

    Arnold, E

    M. Arnold, E. Morgan, H. Rumgay, A. Mafra, D. Singh, M. Laversanne, J. Vignat, J. R. Gralow, F. Cardoso, S. Siesling, I. Soerjomataram, Current and future burden of breast cancer: Global statistics for 2020 and 2040, Breast 66 (2022) 15–23

    2020

  3. [3]

    Plunkett, T

    A. Plunkett, T. L. Scott, E. Tracy, Regional anesthesia for breast cancer surgery: which block is best? a review of the current literature, Pain Manag. 12 (8) (2022) 943–950

    2022

  4. [4]

    Parada, Jr, X

    H. Parada, Jr, X. Sun, C.-K. Tse, A. F. Olshan, M. A. Troester, Lifestyle patterns and survival following breast cancer in the carolina breast can- cer study, Epidemiology 30 (1) (2019) 83–92

    2019

  5. [5]

    Canadian Cancer Society, Survival statistics for breast cancer, https://cancer.ca/en/cancer-information/cancer-types/ breast/prognosis-and-survival/survival-statistics, accessed: 2023-01-27 (2023)

    2023

  6. [6]

    Tabár, P

    L. Tabár, P. B. Dean, F. Lee Tucker, A. M.-F. Yen, S. L.-S. Chen, G. H. H. Jen, J. W.-C. Wang, R. A. Smith, S. W. Duffy, T. H.-H. Chen, A new approach to breast cancer terminology based on the anatomic site of tumour origin: The importance of radiologic imaging biomarkers, Eur. J. Radiol. 149 (110189) (2022) 110189

    2022

  7. [7]

    Tabár, P

    L. Tabár, P. B. Dean, T. H.-H. Chen, A. M.-F. Yen, S. L.-S. Chen, J. C.-Y. Fann, S. Y.-H. Chiu, M. M.-S. Ku, W. Y.-Y. Wu, C.-Y. Hsu, Y.-C. Chen, K. Beckmann, R. A. Smith, S. W. Duffy, The incidence of fatal breast cancer measures the increased effectiveness of therapy in women participating in mammography screening, Cancer 125 (4) (2019) 515–523

    2019

  8. [8]

    S. W. Duffy, L. Tabár, A. M.-F. Yen, P. B. Dean, R. A. Smith, H. Jon- sson, S. Törnberg, S. L.-S. Chen, S. Y.-H. Chiu, J. C.-Y. Fann, M. M.- S. Ku, W. Y.-Y. Wu, C.-Y. Hsu, Y.-C. Chen, G. Svane, E. Azavedo, 20 H. Grundström, P. Sundén, K. Leifland, E. Frodis, J. Ramos, B. Epstein, A. Åkerlund, A. Sundbom, P. Bordás, H. Wallin, L. Starck, A. Björk- gren, S....

    2020

  9. [9]

    Sardanelli, H

    F. Sardanelli, H. S. Aase, M. Álvarez, E. Azavedo, H. J. Baarslag, C. Balleyguier, P. A. Baltzer, V. Beslagic, U. Bick, D. Bogdanovic- Stojanovic, R. Briediene, B. Brkljacic, J. Camps Herrero, C. Colin, E. Cornford, J. Danes, G. de Geer, G. Esen, A. Evans, M. H. Fuch- sjaeger, F. J. Gilbert, O. Graf, G. Hargaden, T. H. Helbich, S. H. Heywang-Köbrunner, V....

    2017

  10. [10]

    R.-J. I. Sweeney, S. J. Lewis, P. Hogg, M. F. McEntee, A review of mammographic positioning image quality criteria for the craniocaudal projection, Br. J. Radiol. 91 (1082) (2018) 20170611

    2018

  11. [11]

    Kendall, P

    E. Kendall, P. Hajishafiezahramini, M. Hamilton, G. Doyle, N. Wad- den, O. Meruvia-Pastor, Full field digital mammography dataset from a population screening program, Sci. Data 12 (1) (2025) 1479

    2025

  12. [12]

    A. N. Giaquinto, H. Sung, L. A. Newman, R. A. Freedman, R. A. Smith, J. Star, A. Jemal, R. L. Siegel, Breast cancer statistics 2024, CA Cancer J. Clin. 74 (6) (2024) 477–495

    2024

  13. [13]

    21 URLhttps://www.cancer.org/cancer/types/breast-cancer/ screening-tests-and-early-detection/mammograms/ limitations-of-mammograms.html

    American Cancer Society, Limitations of mammograms, accessed: 2024-10-18 (2023). 21 URLhttps://www.cancer.org/cancer/types/breast-cancer/ screening-tests-and-early-detection/mammograms/ limitations-of-mammograms.html

    2024

  14. [14]

    Jafari, E

    Z. Jafari, E. Karami, Breast cancer detection in mammography im- ages: ACNN-basedapproachwithfeatureselection, Information(Basel) 14 (7) (2023) 410

    2023

  15. [15]

    E. J. Kendall, M. G. Barnett, K. Chytyk-Praznik, Automatic detection of anomalies in screening mammograms, BMC Med. Imaging 13 (1) (2013) 43

    2013

  16. [16]

    E. J. Kendall, M. T. Flynn, Automated breast image classification us- ing features from its discrete cosine transform, PLoS One 9 (3) (2014) e91015

    2014

  17. [17]

    Abdikenov, T

    B. Abdikenov, T. Zhaksylyk, A. Imasheva, Y. Orazayev, T. Karibekov, Innovative multi-view strategies for AI-assisted breast cancer detection in mammography, J. Imaging 11 (8) (2025) 247

    2025

  18. [18]

    R. S. Lee, F. Gimenez, A. Hoogi, K. K. Miyake, M. Gorovoy, D. L. Rubin, A curated mammography data set for use in computer-aided detection and diagnosis research, Sci. Data 4 (1) (2017) 170177

    2017

  19. [19]

    H. Cai, J. Wang, T. Dan, J. Li, Z. Fan, W. Yi, C. Cui, X. Jiang, L. Li, An online mammography database with biopsy confirmed types, Sci. Data 10 (1) (2023) 123

    2023

  20. [20]

    S. F. Taskiran, B. Turkoglu, E. Kaya, T. Asuroglu, A comprehensive evaluation of oversampling techniques for enhancing text classification performance, Sci. Rep. 15 (1) (2025) 21631

    2025

  21. [21]

    O. M. Velarde, C. Lin, S. Eskreis-Winkler, L. C. Parra, Robustness of deep networks for mammography: Replication across public datasets, J. Imaging Inform. Med. 37 (2) (2024) 536–546

    2024

  22. [22]

    X. Wang, G. Liang, Y. Zhang, H. Blanton, Z. Bessinger, N. Jacobs, Inconsistent performance of deep learning models on mammogram clas- sification, J. Am. Coll. Radiol. 17 (6) (2020) 796–803. 22

    2020

  23. [23]

    Ghosh, C

    S. Ghosh, C. B. Poynton, S. Visweswaran, K. Batmanghelich, Mammo- CLIP: A vision language foundation model to enhance data efficiency and robustness in mammography, in: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, Springer Nature Switzerland, Cham, 2024, pp. 632–642

    2024

  24. [24]

    Kendall, P

    E. Kendall, P. Hajishafiezahramini, M. Hamilton, G. Doyle, N. Wad- den, O. Meruvia-Pastor, Newfoundland and labrador breast screening (NLBS) dataset (NL-Breast-Screen) (2025).doi:10.20383/103.01526. URLhttps://doi.org/10.20383/103.01526

    work page doi:10.20383/103.01526 2025

  25. [25]

    H. Guan, M. Liu, Domain adaptation for medical image analysis: A survey, IEEE Trans. Biomed. Eng. 69 (3) (2022) 1173–1185

    2022

  26. [26]

    Suckling, J

    J. Suckling, J. Parker, D. Dance, S. Astley, I. Hutt, C. Boggis, I. Rick- etts, E. Stamatakis, N. Cerneaz, S. Kok, P. Taylor, D. Betal, J. Savage, Mammographic image analysis society (MIAS) database v1.21 (2015)

    2015

  27. [27]

    Heath, et al., Current status of the digital database for screening mammography, in: M

    M. Heath, et al., Current status of the digital database for screening mammography, in: M. J. Yaffe (Ed.), Proceedings of the Fifth Interna- tional Workshop on Digital Mammography, Medical Physics Publishing, 2001, pp. 212–218

    2001

  28. [28]

    M. D. Halling-Brown, L. M. Warren, D. Ward, E. Lewis, A. Mackenzie, M. G. Wallis, L. S. Wilkinson, R. M. Given-Wilson, R. McAvinchey, K. C. Young, OPTIMAM mammography image database: A large-scale resource of mammography images and clinical data, Radiol. Artif. Intell. 3 (1) (2021) e200103

    2021

  29. [29]

    Dembrower, P

    K. Dembrower, P. Lindholm, F. Strand, A multi-million mammogra- phy image dataset and population-based screening cohort for the train- ing and evaluation of deep neural networks-the cohort of Screen-Aged women (CSAW), J. Digit. Imaging 33 (2) (2020) 408–413

    2020

  30. [30]

    H. M. L. Frazer, J. S. N. Tang, M. S. Elliott, K. M. Kunicki, B. Hill, R. Karthik, C. F. Kwok, C. A. Peña-Solorzano, Y. Chen, C. Wang, O. Al-Qershi, S. K. Fox, S. Li, E. Makalic, T. L. Nguyen, D. F. Schmidt, P. Basnayake Ralalage, J. F. Lippey, P. Brotchie, J. L. Hop- per, G. Carneiro, D. J. McCarthy, ADMANI: Annotated digital mam- mograms and associated ...

    2023

  31. [31]

    I. C. Moreira, I. Amaral, I. Domingues, A. Cardoso, M. J. Cardoso, J. S. Cardoso, INbreast: toward a full-field digital mammographic database, Acad. Radiol. 19 (2) (2012) 236–248

    2012

  32. [32]

    H. T. Nguyen, H. Q. Nguyen, H. H. Pham, K. Lam, L. T. Le, M. Dao, V. Vu, VinDr-Mammo: A large-scale benchmark dataset for computer- aided diagnosis in full-field digital mammography, Sci. Data 10 (1) (2023) 277

    2023

  33. [33]

    Carr, et al., Rsna screening mammography breast cancer detection,https://kaggle.com/competitions/ rsna-breast-cancer-detection, accessed: 2023-01-10 (2022)

    C. Carr, et al., Rsna screening mammography breast cancer detection,https://kaggle.com/competitions/ rsna-breast-cancer-detection, accessed: 2023-01-10 (2022)

    2023

  34. [34]

    J. J. Jeong, B. L. Vey, A. Bhimireddy, T. Kim, T. Santos, R. Cor- rea, R. Dutt, M. Mosunjac, G. Oprea-Ilies, G. Smith, M. Woo, C. R. McAdams, M. S. Newell, I. Banerjee, J. Gichoya, H. Trivedi, The EMory BrEast imaging dataset (EMBED): A racially diverse, granular dataset of 3.4 million screening and diagnostic mammographic images, Radiol. Artif. Intell. 5...

    2023

  35. [35]

    A. Sahu, P. K. Das, S. Meher, High accuracy hybrid CNN classifiers for breast cancer detection using mammogram and ultrasound datasets, Biomed. Signal Process. Control 80 (104292) (2023) 104292

    2023

  36. [36]

    R. S. Raaj, Breast cancer detection and diagnosis using hybrid deep learning architecture, Biomed. Signal Process. Control 82 (104558) (2023) 104558

    2023

  37. [37]

    M. Tan, Q. Le, EfficientNet: Rethinking model scaling for convolutional neural networks, in: K. Chaudhuri, R. Salakhutdinov (Eds.), Proceed- ings of the 36th International Conference on Machine Learning, Vol. 97 of Proceedings of Machine Learning Research, PMLR, 2019, pp. 6105– 6114. URLhttps://proceedings.mlr.press/v97/tan19a.html

    2019

  38. [38]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization (Nov. 2017).arXiv:1711.05101

    Pith/arXiv arXiv 2017

  39. [39]

    Z. Liu, K. He, A decade’s battle on dataset bias: Are we there yet?, in: International Conference on Learning Representations (ICLR), 2025, arXiv:2403.08632. 24

    Pith/arXiv arXiv 2025

  40. [40]

    E. R. DeLong, D. M. DeLong, D. L. Clarke-Pearson, Comparing the ar- easundertwoormorecorrelatedreceiveroperatingcharacteristiccurves: a nonparametric approach, Biometrics 44 (3) (1988) 837–845

    1988

  41. [41]

    Holm, A simple sequentially rejective multiple test procedure, Scan- dinavian Journal of Statistics 6 (2) (1979) 65–70

    S. Holm, A simple sequentially rejective multiple test procedure, Scan- dinavian Journal of Statistics 6 (2) (1979) 65–70. URLhttp://www.jstor.org/stable/4615733

    arXiv 1979

  42. [42]

    J. R. Zech, M. A. Badgeley, M. Liu, A. B. Costa, J. J. Titano, E. K. Oermann, Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study, PLoS Med. 15 (11) (2018) e1002683

    2018

  43. [43]

    Zhang, X

    L. Zhang, X. Wang, D. Yang, T. Sanford, S. Harmon, B. Turkbey, B. J. Wood, H.Roth, A.Myronenko, D.Xu, Z.Xu, Generalizingdeeplearning for medical image segmentation to unseen domains via deep stacked transformation, IEEE Trans. Med. Imaging 39 (7) (2020) 2531–2540. 25

    2020