Pith. sign in

REVIEW 4 major objections 6 minor 60 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

Style-transferred patches boost cross-site mammogram calcification AUC by 4-5 points

2026-07-08 01:58 UTC pith:WMZXWMOV

load-bearing objection Style transfer for calcification classification across mammography sites: real clinical question, modest gains, missing control the 4 major comments →

arxiv 2607.06549 v1 pith:WMZXWMOV submitted 2026-07-07 cs.CV

Unsupervised Domain Adaptation for Calcification Classification in Mammography Across Multi-Site Datasets

Xuan Liu , Derek L. Nguyen , Emily C. Barre , Jennifer Thomas , Thomas Lynch , Jeffrey R. Marks , E. Shelley Hwang , Marc D. Ryser
show 2 more authors
Joseph Y. Lo Lars J. Grimm
This is my paper
classification cs.CV
keywords domain adaptationmammographycalcification classificationstyle transferCycleGANAdaINSwin Transformer V2multi-site generalization
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.

This paper claims that unsupervised domain adaptation via CycleGAN-generated vendor-specific image patches can meaningfully improve the generalization of a deep learning calcification classifier (benign vs. malignant) when trained on one mammography dataset and tested on others from different institutions and scanner vendors. The authors train a Swin Transformer V2 classifier on the OPTIMAM dataset (UK, Hologic FFDM) and then augment training with patches whose visual style has been transferred to match target-domain vendors and imaging techniques (GE FFDM, Hologic synthetic 2D) using unpaired image-to-image translation. On two independent external test sets—EMBED (Emory) and the Duke Calcification Dataset—the augmented model raises AUC from 0.68 to 0.72 and 0.68 to 0.73, respectively. The central mechanism is that style transfer models learn to map the texture, brightness, and sharpness statistics of source-domain calcification patches into target-domain appearances without requiring any labeled data from the target domain, and these stylized patches then teach the downstream classifier features that are more robust to vendor and technique shifts. CycleGAN outperforms AdaIN in this role, apparently because it better preserves calcification boundary structure during translation.

Core claim

The paper's central discovery is that CycleGAN-based unpaired style transfer, applied to generate vendor- and technique-specific mammography patches from unlabeled images, produces training augmentations that yield consistent AUC improvements of 4-5 points on external datasets from different institutions and scanner vendors, without any annotated target-domain data. The improvement holds across vendor subgroups within the Duke dataset (GE: 0.70 to 0.75; Hologic: 0.67 to 0.71), suggesting the benefit is not specific to one vendor shift. AdaIN, a simpler statistical style alignment method, produced smaller gains and introduced visual artifacts (blurred calcification boundaries, spurious lesion

What carries the argument

Two style transfer models serve as the domain adaptation engine: AdaIN (adaptive instance normalization), which aligns channel-wise mean and variance of content and style features in a single pass, and CycleGAN, which uses two generators and two discriminators with a cycle-consistency loss to perform unpaired image-to-image translation. The downstream classifier is Swin Transformer V2, a hierarchical vision transformer with shifted-window self-attention. The style transfer models are trained on unlabeled patches from mass cases (not calcification cases) to learn vendor-specific appearance, then applied to annotated calcification patches from the source domain to create stylized training data

Load-bearing premise

The paper assumes that visual style transfer—matching brightness, sharpness, and texture statistics between vendor domains—captures the clinically relevant portion of the domain shift. If the shift is substantially driven by factors beyond image style, such as patient demographics, annotation protocols, or vendor-specific rendering of calcification morphology, then style transfer alone cannot fully address it.

What would settle it

If the same AUC improvements could be achieved by adding non-stylized augmented patches (e.g., random contrast or noise perturbations) of equal quantity to the training set, the benefit would be attributable to generic data augmentation rather than vendor-specific style transfer.

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

If this is right

  • If style-transfer augmentation generalizes to other lesion types and modalities, institutions could deploy CAD systems trained on public datasets to local patient populations with reduced annotation burden.
  • The finding that CycleGAN preserves calcification morphology better than AdaIN suggests that structure-preserving translation is more important than pure statistical style matching for clinically meaningful domain adaptation.
  • The vendor-subgroup improvements (GE and Hologic both improved) indicate the mechanism is not overfitting to a single vendor's texture but is learning something closer to vendor-invariant features.
  • The approach could be extended to other imaging domains where domain shift is driven primarily by acquisition hardware differences—CT scanners, MRI vendors—though the paper does not test this.

Where Pith is reading between the lines

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

  • The paper sources style-transfer training patches from mass cases rather than calcification cases, meaning the generated calcification content may not reflect true vendor-specific calcification appearance. If the domain shift in calcification classification is partly driven by how different vendors render calcification morphology (not just brightness and texture), this approach may leave residual
  • The modest absolute AUC values (0.72-0.73 even after improvement) suggest that vendor/technique style is only one component of the domain shift; patient cohort differences, annotation protocols, and class imbalance likely contribute substantially and are not addressed by image-level style transfer alone.
  • A direct test of whether the improvement comes from vendor-style learning versus simple data augmentation would compare CycleGAN-stylized patches against randomly perturbed patches (noise, contrast jitter) of equal quantity; the paper does not include this control.

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

4 major / 6 minor

Summary. This paper proposes a multi-stage framework for benign vs. malignant calcification classification across multi-site mammography datasets. The framework combines an unsupervised domain adaptation module (using AdaIN or CycleGAN style transfer to generate vendor- and technique-specific patches from unlabeled mammograms) with a Swin Transformer V2 classifier. The model is trained on OPTIMAM (n=2994) and externally evaluated on EMBED (n=125) and the Duke Calcification Dataset v1 (n=788). The authors report AUC improvements from 0.68 to 0.72 on EMBED and 0.68 to 0.73 on Duke when incorporating CycleGAN-generated patches into training.

Significance. The paper addresses a clinically relevant problem—cross-site generalization of calcification classification—and is one of the first to systematically study domain adaptation for this specific lesion type. Strengths include the use of two independent external test sets, a 5-fold cross-validation protocol, a backbone comparison across eight architectures, and transparent reporting of per-vendor AUC breakdowns on the Duke dataset. The overall framework design is reasonable and the problem is well-motivated.

major comments (4)
  1. [Table II and Section IV.C] The central claim—that style-transfer-based domain adaptation improves cross-site classification—cannot be distinguished from a generic data augmentation effect. The training set expands from original OPTIMAM patches to original + CycleGAN-generated patches, but no control experiment adds an equal number of non-style-transferred augmented patches (e.g., random brightness/contrast/noise jitter or simple geometric augmentation). Without this ablation, the observed AUC gains (0.68→0.72 on EMBED, 0.68→0.73 on Duke) could be entirely attributable to the model seeing more diverse training data rather than to vendor-specific style alignment. This is load-bearing for the paper's attribution of improvement to domain adaptation.
  2. [Table IA and Section III.D] The EMBED test set is 100% Hologic FFDM screening—the same vendor and technique as the OPTIMAM training set. The domain adaptation module generates GE FFDM and Hologic synthetic style patches (Table IB), neither of which matches the EMBED domain. Any improvement on EMBED must therefore come from general augmentation effects rather than targeted vendor/technique adaptation, which directly undermines the paper's attribution of the EMBED gain to domain adaptation. The authors should either provide the augmentation control ablation mentioned above or reframe the EMBED result as not being attributable to domain adaptation.
  3. [Table II] No statistical significance testing is reported for the AUC improvements. The standard deviations (±0.02–0.03) are across 5 cross-validation folds, not bootstrap confidence intervals of the test-set AUC. Given EMBED's small size (n=125, 35 malignant), a 0.04 AUC improvement could easily be non-significant. The Duke improvement (0.05, n=788) is more likely real but also remains untested. Bootstrap or DeLong test comparisons between baseline and domain-adapted models on the external test sets are needed to support the claimed improvements.
  4. [Section III.D] The domain adaptation patches are sourced from mass cases rather than calcification cases. The authors state this choice was made to introduce 'more diverse and complex tissue patterns,' but it means the generated calcification content (transferred from OPTIMAM calcification patches) is combined with style statistics derived from mass-case backgrounds. If vendor-specific calcification appearance differs from vendor-specific mass-case background appearance, the style transfer may not capture the clinically relevant domain shift for calcifications. The authors should discuss this limitation and ideally provide evidence that the style statistics transfer appropriately.
minor comments (6)
  1. [Abstract] The abstract states 'AUC 0.68 to 0.72' for EMBED without noting that this is a relative gain of ~5.9%, not exceeding 5%. The text in Section IV.C says 'relative gains exceeding 5%,' which is correct for Duke (7.4%) but borderline for EMBED (5.9%). Consider clarifying.
  2. [Figure 7] The figure caption could better indicate which columns correspond to AdaIN vs. CycleGAN outputs. The current description in the caption text helps, but direct column labels would improve clarity.
  3. [Eq. (1)] The values of λ and α are not specified in the main text. These are important hyperparameters for reproducibility and should be reported.
  4. [Section III.C] The malignant:benign batch ratio of 1:2 is mentioned, but it is unclear whether this ratio applies to the combined original + generated patches or only to the original patches. Clarification would help reproducibility.
  5. [Table IB] The number of GE FFDM cases is listed as 857 in Table IB but 431 in Section III.D. Please reconcile.
  6. [Section V] The discussion mentions that random oversampling and undersampling were explored but does not report the corresponding AUC values. Including these results (even briefly) would strengthen the claim that the chosen strategy was optimal.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a thorough and constructive review. The referee raises four major points: (1) the absence of a non-style-transfer augmentation control, (2) the mismatch between generated target domains and the EMBED test set, (3) the lack of statistical significance testing, and (4) the use of mass-case backgrounds for calcification style transfer. We agree that points 1, 2, and 3 require revisions to strengthen the paper's claims, and we describe below how we will address each. On point 4, we provide a substantive defense of our design choice while acknowledging the limitation the referee identifies.

read point-by-point responses

  1. Referee: The central claim—that style-transfer-based domain adaptation improves cross-site classification—cannot be distinguished from a generic data augmentation effect. No control experiment adds an equal number of non-style-transferred augmented patches. Without this ablation, the observed AUC gains could be entirely attributable to the model seeing more diverse training data rather than to vendor-specific style alignment.

    Authors: The referee is correct that without a matched augmentation control, we cannot definitively attribute the observed AUC gains to vendor-specific style alignment rather than to a generic data augmentation effect. This is a fair and important criticism. We will add a control experiment in which an equal number of patches augmented with standard geometric and photometric transformations (random rotation, flipping, brightness/contrast jitter, and Gaussian noise) are added to the original OPTIMAM training set. This will allow a direct comparison between generic augmentation and style-transfer-based domain adaptation. We will revise the manuscript to include this ablation in Table II and to temper our attributional language accordingly, noting that the gains reflect the contribution of style-transferred patches specifically relative to both a no-augmentation baseline and a generic-augmentation control. revision: yes

  2. Referee: The EMBED test set is 100% Hologic FFDM screening—the same vendor and technique as the OPTIMAM training set. The domain adaptation module generates GE FFDM and Hologic synthetic style patches, neither of which matches the EMBED domain. Any improvement on EMBED must therefore come from general augmentation effects rather than targeted vendor/technique adaptation, which directly undermines the paper's attribution of the EMBED gain to domain adaptation.

    Authors: The referee makes an accurate observation about the vendor and technique overlap between OPTIMAM and EMBED. We agree that the EMBED improvement cannot be attributed to targeted vendor or technique adaptation, since EMBED is Hologic FFDM screening—the same domain as the OPTIMAM training data. In the revised manuscript, we will reframe the EMBED results explicitly: the gains on EMBED should be interpreted as reflecting general robustness improvements from exposure to diverse image appearances (including cross-vendor and cross-technique style-transferred patches), not as evidence of targeted domain adaptation to the EMBED domain. We will retain the EMBED results as an external validation but will separate the attribution narrative: the Duke dataset, which includes GE FFDM and Hologic synthetic images that match the generated target domains, is the appropriate test set for evaluating targeted domain adaptation. We will make this distinction clear in the abstract, results, and discussion. revision: yes

  3. Referee: No statistical significance testing is reported for the AUC improvements. The standard deviations are across 5 cross-validation folds, not bootstrap confidence intervals of the test-set AUC. Given EMBED's small size (n=125, 35 malignant), a 0.04 AUC improvement could easily be non-significant. Bootstrap or DeLong test comparisons between baseline and domain-adapted models on the external test sets are needed.

    Authors: We agree that statistical significance testing is necessary and currently absent. We will add DeLong tests comparing the baseline and domain-adapted models on both external test sets, and we will also report bootstrap 95% confidence intervals for the test-set AUCs. We acknowledge that the EMBED improvement (n=125, 35 malignant) may not reach statistical significance, and we will report this transparently. The Duke improvement (n=788) is more likely to be significant, but we will let the test results speak for themselves. We will update Table II to include confidence intervals and p-values, and we will revise the text to qualify claims of improvement where significance is not established. revision: yes

  4. Referee: The domain adaptation patches are sourced from mass cases rather than calcification cases. The style statistics derived from mass-case backgrounds may not capture the clinically relevant domain shift for calcifications. The authors should discuss this limitation and ideally provide evidence that the style statistics transfer appropriately.

    Authors: We appreciate the referee's attention to this design choice. Our rationale for using mass-case patches as style sources was pragmatic: we needed unlabeled images from the target vendor/technique domains (GE FFDM and Hologic synthetic), and calcification cases in OPTIMAM are predominantly Hologic FFDM, leaving no GE FFDM or Hologic synthetic calcification patches available for style extraction. Mass-case patches provided the necessary vendor- and technique-specific image statistics (noise texture, contrast response, acquisition artifacts) that constitute the domain shift we aim to address. We note that AdaIN-style transfer operates on global feature statistics (channel-wise mean and variance), and CycleGAN learns domain-level appearance mappings—both of which capture vendor-level image characteristics that are not lesion-type-specific. Nevertheless, the referee raises a valid concern about whether mass-case background statistics fully represent the calcification-relevant domain shift. We will add a discussion of this limitation, including the constraint that motivated the choice, and we will note that future work with calcification-specific target-domain patches could provide more targeted style alignment. We respectfully note that providing quantitative evidence of appropriate style transfer for calcifications would require target-domain calcification patches with known vendor labels, which are not available in our current datasets. revision: partial

Circularity Check

0 steps flagged

No significant circularity: the central claim is evaluated on external held-out datasets and the one self-citation is a minor auxiliary component, not load-bearing for the main result.

full rationale

The paper's central claim—that incorporating CycleGAN-generated vendor-specific patches into training improves cross-site calcification classification (AUC 0.68→0.72 on EMBED, 0.68→0.73 on Duke)—is evaluated on external datasets (EMBED, Duke) that are never used during training, style transfer, or validation. The domain adaptation module is trained on unlabeled OPTIMAM patches from different vendors, and the classification improvement is measured against an external benchmark, not a fitted parameter renamed as a prediction. The one self-citation (reference [42], Hou et al. 2019) provides a U-Net model used to generate calcification segmentation masks for the auxiliary segmentation loss in the AdaIN pipeline. This is a minor auxiliary component: (a) it only affects the AdaIN variant, not the CycleGAN variant that achieves the best results; (b) the segmentation mask is used as a regularization signal, not as the target output; (c) the cited model is a standard U-Net trained on a different task (DCIS upstaging prediction), not a result that is re-derived or assumed in the present paper. The CycleGAN results—the paper's strongest results—do not depend on this self-citation at all. No step in the derivation chain reduces to its inputs by construction. The absence of a data-augmentation control ablation and significance testing is a correctness risk, not a circularity issue: the measured AUC gains are empirical observations on external data, not quantities forced by definition or fit. The derivation is self-contained against external benchmarks, so the circularity score is 1 (one minor self-citation that is not load-bearing). The reader's skepticism about whether the gains are attributable to domain adaptation versus generic augmentation is a valid experimental-design concern, but it does not make the paper's claim circular—the claim is falsified or supported by external test performance, not by a definitional identity.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The paper introduces no new entities, particles, forces, or postulated objects. It combines existing architectures (Swin Transformer V2, AdaIN, CycleGAN, U-Net) and existing datasets. The free parameters are standard hyperparameters for the methods used; their values are either stated or unfortunately omitted.

free parameters (4)
  • lambda (style loss weight in AdaIN) = not specified
    Weight for style loss term L_sty in Eq. 1; value not stated in the paper.
  • alpha (segmentation loss weight in AdaIN) = not specified
    Weight for auxiliary segmentation loss L_seg in Eq. 1; value not stated in the paper.
  • Malignant:benign batch ratio = 1:2
    Ratio of malignant to benign patches in each training batch, chosen to address class imbalance.
  • Learning rate = 0.00005
    Adam optimizer learning rate for classification backbone.
axioms (3)
  • domain assumption Visual style transfer (matching brightness, sharpness, texture statistics) is sufficient to capture the clinically relevant domain shift between mammography vendors and imaging techniques.
    Invoked in Section III.B and III.C; the entire framework rests on the premise that style transfer can reduce domain discrepancies relevant to calcification classification.
  • domain assumption Mass-case patches provide appropriate tissue background variations for learning domain-invariant features applicable to calcification classification.
    Section III.D states that unlabeled mass-case patches were used for domain adaptation rather than normal or calcification patches, justified as providing 'more realistic and challenging tissue background variations.'
  • domain assumption The U-Net segmentation model from prior work [42] produces sufficiently accurate calcification masks for the AdaIN auxiliary segmentation loss.
    Section III.B uses masks generated by a previously trained U-Net for the segmentation branch without validating mask quality on the current dataset.

pith-pipeline@v1.1.0-glm · 17053 in / 2406 out tokens · 237344 ms · 2026-07-08T01:58:46.733821+00:00 · methodology

0 comments
read the original abstract

Deep learning-based computer-aided diagnosis (CAD) systems have shown strong performance in breast cancer diagnosis, particularly for classification tasks in mammography. However, domain shifts across multi-site datasets remain a challenge, especially when models are applied to unseen domains. In this work, we proposed a calcification classification framework to improve malignant versus benign breast disease classification across multi-site mammography datasets. The framework consisted of two components: (1) an unsupervised domain adaptation module based on style transfer models (AdaIN and CycleGAN) to generate vendor-specific and technique-specific training samples without additional annotations, and (2) a supervised classification module using Swin Transformer V2 as the backbone. We evaluated the proposed method on three datasets: cross-validation on OPTIMAM (National Health Service, United Kingdom; n=2994), followed by external validation on EMBED (Emory University; n=125), and Duke Calcification Dataset v1 (n=788). These datasets cover multiple vendors and include both full-field digital mammography and synthetic 2D images derived from digital breast tomosynthesis. The proposed framework improved cross-site performance for both EMBED (AUC 0.68 to 0.72) and the Duke Calcification Dataset (AUC 0.68 to 0.73). These findings indicate that domain adaptation can reduce domain shifts and improve the generalization for calcification classification across multi-site datasets.

Figures

Figures reproduced from arXiv: 2607.06549 by Derek L. Nguyen, Emily C. Barre, E. Shelley Hwang, Jeffrey R. Marks, Jennifer Thomas, Joseph Y. Lo, Lars J. Grimm, Marc D. Ryser, Thomas Lynch, Xuan Liu.

Figure 1
Figure 1. Figure 1: Illustration of single-site and multi-site dataset settings and domain adaptation across institutions. widely applied over the past decade. However, many existing studies rely on older, small scale-datasets, such as CBIS￾DDSM [30] and INbreast [31], or only use limited subsets of larger datasets such as OPTIMAM [18] and EMBED [19]. This is largely due to restricted access to the complete datasets and the l… view at source ↗
Figure 2
Figure 2. Figure 2: Representative examples of mammograms from different scan￾ner vendors (GE and Hologic) using different techniques (full field digital mammogram and synthetic mammogram; screening and diagnostic), where (a) and (b) are from the same patient, while (c) and (d) are from different patients. also been applied to mitigate domain discrepancies arising from different vendors and clinical centers [37]. In breast im… view at source ↗
Figure 4
Figure 4. Figure 4: Overall framework of the proposed domain adaptation method for calcification classification. The framework included two pipelines: (a) an inference pipeline, which used only the classification backbone network, and (b) a training/validation pipeline, which included an unsupervised domain adaptation module and a supervised malignant/benign classification module. For the single-site classification task, only… view at source ↗
Figure 5
Figure 5. Figure 5: Performance of seven CNN-based backbone networks and a Swin Transformer-based backbone. Although the pre-trained backbones differed greatly on ImageNet-1k, fine-tuned calcification classification performance was similar (AUC 0.78-0.81). Swin Transformer V2 was chosen as the baseline because it achieved the highest AUC (0.81) with moderate model complexity [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of OPTIMAM-trained baseline without domain adaptation versus domain adaptation-based framework in external val￾idations with EMBED and Duke datasets (Duke Calcification Dataset v1). The full framework with CycleGAN achieves consistently higher sensitivity at corresponding specificity operating points on both external datasets. patches, both AdaIN and CycleGAN generate outputs with vendor-specifi… view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison of style transfer results. GE FFDM and Hologic synthetic patches were used as style patches in the first and second rows, respectively, while Hologic FFDM patches are used as content patches in both rows. The first and second columns show results from AdaIN and CycleGAN, respectively. C. Comparison Across Multi-Site Datasets Next, we evaluated the full multi-site calcification classifi￾ca… 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

60 extracted references · 60 canonical work pages · 3 internal anchors

  1. [1]

    and Jemal, A.,

    Siegel, R.L., Kratzer, T.B., Giaquinto, A.N., Sung, H. and Jemal, A.,

    work page

  2. [2]

    Ca, 75(1), p.10

    Cancer statistics, 2025. Ca, 75(1), p.10

    work page 2025

  3. [3]

    and Wilcox, M., 2013

    Marmot, M.G., Altman, D.G., Cameron, D.A., Dewar, J.A., Thompson, S.G. and Wilcox, M., 2013. The benefits and harms of breast cancer screening: an independent review. British journal of cancer, 108(11), pp.2205-2240

    work page 2013

  4. [4]

    and Sieh, W., 2019

    Shen, L., Margolies, L.R., Rothstein, J.H., Fluder, E., McBride, R. and Sieh, W., 2019. Deep learning to improve breast cancer detection on screening mammography. Scientific reports, 9(1), p.12495

    work page 2019

  5. [5]

    and Tolba, M.F., 2020, March

    Hamed, G., Marey, M.A.E.R., Amin, S.E.S. and Tolba, M.F., 2020, March. Deep learning in breast cancer detection and classification. In The International Conference on Artificial Intelligence and Computer Vision (pp. 322-333). Cham: Springer International Publishing

    work page 2020

  6. [6]

    and Choudhury, T., 2018, December

    Sharma, S., Aggarwal, A. and Choudhury, T., 2018, December. Breast cancer detection using machine learning algorithms. In 2018 Interna- tional conference on computational techniques, electronics and mechan- ical systems (CTEMS) (pp. 114-118). IEEE

    work page 2018

  7. [7]

    and Aly, M.H., 2021

    Salama, W.M. and Aly, M.H., 2021. Deep learning in mammography images segmentation and classification: Automated CNN approach. Alexandria Engineering Journal, 60(5), pp.4701-4709

    work page 2021

  8. [8]

    and Elmaghraby, A.S., 2021

    Baccouche, A., Garcia-Zapirain, B., Castillo Olea, C. and Elmaghraby, A.S., 2021. Connected-UNets: a deep learning architecture for breast mass segmentation. NPJ Breast Cancer, 7(1), p.151

    work page 2021

  9. [9]

    and Ma, Y ., 2019

    Li, H., Zhuang, S., Li, D.A., Zhao, J. and Ma, Y ., 2019. Benign and malignant classification of mammogram images based on deep learning. Biomedical Signal Processing and Control, 51, pp.347-354

    work page 2019

  10. [10]

    and Hashmi, A., 2021

    Malebary, S.J. and Hashmi, A., 2021. Automated breast mass classi- fication system using deep learning and ensemble learning in digital mammogram. IEEE Access, 9, pp.55312-55328

    work page 2021

  11. [11]

    and Cox, D., 2017, September

    Lotter, W., Sorensen, G. and Cox, D., 2017, September. A multi-scale CNN and curriculum learning strategy for mammogram classification. In International Workshop on Deep Learning in Medical Image Analysis (pp. 169-177). Cham: Springer International Publishing

    work page 2017

  12. [12]

    and Lekadir, K., 2022

    Garrucho, L., Kushibar, K., Jouide, S., Diaz, O., Igual, L. and Lekadir, K., 2022. Domain generalization in deep learning based mass detection in mammography: A large-scale multi-center study. Artificial Intelli- gence in Medicine, 132, p.102386. LIUet al.: UNSUPERVISED DOMAIN ADAPTATION FOR CALCIFICATION CLASSIFICATION IN MAMMOGRAPHY ACROSS MULTI-SITE DATASETS 9

    work page 2022

  13. [13]

    and Parra, L.C., 2024

    Velarde, O.M., Lin, C., Eskreis-Winkler, S. and Parra, L.C., 2024. Robustness of deep networks for mammography: Replication across public datasets. Journal of Imaging Informatics in Medicine, 37(2), pp.536-546

    work page 2024

  14. [14]

    and Shenton- Taylor, C., 2025

    Hickman, A.J., Gomes, S., Warren, L.M., Smith, N.A. and Shenton- Taylor, C., 2025. Assessing the generalisation of artificial intelli- gence across mammography manufacturers. PLOS Digital Health, 4(8), p.e0000973

    work page 2025

  15. [15]

    and Ryser, M.D., 2019

    Grimm, L.J., Miller, M.M., Thomas, S.M., Liu, Y ., Lo, J.Y ., Hwang, E.S., Hyslop, T. and Ryser, M.D., 2019. Growth dynamics of mammographic calcifications: differentiating ductal carcinoma in situ from benign breast disease. Radiology, 292(1), pp.77-83

    work page 2019

  16. [16]

    and iCAIRD Radiology Collaboration, 2023

    de Vries, C.F., Colosimo, S.J., Staff, R.T., Dymiter, J.A., Yearsley, J., Dinneen, D., Boyle, M., Harrison, D.J., Anderson, L.A., Lip, G. and iCAIRD Radiology Collaboration, 2023. Impact of different mammog- raphy systems on artificial intelligence performance in breast cancer screening. Radiology: Artificial Intelligence, 5(3), p.e220146

    work page 2023

  17. [17]

    and Belongie, S., 2017

    Huang, X. and Belongie, S., 2017. Arbitrary style transfer in real- time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision (pp. 1501-1510)

    work page 2017

  18. [18]

    and Efros, A.A., 2017

    Zhu, J.Y ., Park, T., Isola, P. and Efros, A.A., 2017. Unpaired image- to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision (pp. 2223-2232)

    work page 2017

  19. [19]

    and Young, K.C., 2020

    Halling-Brown, M.D., Warren, L.M., Ward, D., Lewis, E., Mackenzie, A., Wallis, M.G., Wilkinson, L.S., Given-Wilson, R.M., McAvinchey, R. and Young, K.C., 2020. OPTIMAM mammography image database: a large-scale resource of mammography images and clinical data. Radiology: Artificial Intelligence, 3(1), p.e200103

    work page 2020

  20. [20]

    and Woo, M.,

    Jeong, J.J., Vey, B.L., Bhimireddy, A., Kim, T., Santos, T., Correa, R., Dutt, R., Mosunjac, M., Oprea-Ilies, G., Smith, G. and Woo, M.,

    work page

  21. [21]

    Radiology: Artificial Intelligence, 5(1), p.e220047

    The EMory BrEast imaging Dataset (EMBED): A racially diverse, granular dataset of 3.4 million screening and diagnostic mammographic images. Radiology: Artificial Intelligence, 5(1), p.e220047

    work page

  22. [22]

    and Nagwanshi, K.K., 2021

    Heenaye-Mamode Khan, M., Boodoo-Jahangeer, N., Dullull, W., Nathire, S., Gao, X., Sinha, G.R. and Nagwanshi, K.K., 2021. Multi- class classification of breast cancer abnormalities using Deep Convolu- tional Neural Network (CNN). Plos one, 16(8), p.e0256500

    work page 2021

  23. [23]

    and Li, L., 2016

    Wang, J., Yang, X., Cai, H., Tan, W., Jin, C. and Li, L., 2016. Discrim- ination of breast cancer with microcalcifications on mammography by deep learning. Scientific reports, 6(1), p.27327

    work page 2016

  24. [24]

    and Geras, K.J., 2021

    Shen, Y ., Wu, N., Phang, J., Park, J., Liu, K., Tyagi, S., Heacock, L., Kim, S.G., Moy, L., Cho, K. and Geras, K.J., 2021. An interpretable classifier for high-resolution breast cancer screening images utilizing weakly supervised localization. Medical image analysis, 68, p.101908

    work page 2021

  25. [25]

    and Choe, S.W., 2023

    Ayana, G., Dese, K., Dereje, Y ., Kebede, Y ., Barki, H., Amdissa, D., Husen, N., Mulugeta, F., Habtamu, B. and Choe, S.W., 2023. Vision- transformer-based transfer learning for mammogram classification. Di- agnostics, 13(2), p.178

    work page 2023

  26. [26]

    and Al-Antari, M.A., 2022

    Al-Tam, R.M., Al-Hejri, A.M., Narangale, S.M., Samee, N.A., Mah- moud, N.F., Al-Masni, M.A. and Al-Antari, M.A., 2022. A hybrid workflow of residual convolutional transformer encoder for breast cancer classification using digital X-ray mammograms. Biomedicines, 10(11), p.2971

    work page 2022

  27. [27]

    and Xu, H., 2022, September

    Sun, Z., Jiang, H., Ma, L., Yu, Z. and Xu, H., 2022, September. Transformer based multi-view network for mammographic image clas- sification. In International conference on medical image computing and computer-assisted intervention (pp. 46-54). Cham: Springer Nature Switzerland

    work page 2022

  28. [28]

    and Igarashi, T.,

    Uematsu, T., Nakashima, K., Harada, T.L., Nasu, H. and Igarashi, T.,

    work page

  29. [29]

    Breast Cancer, 30(1), pp.46-55

    Artificial intelligence computer-aided detection enhances synthe- sized mammograms: comparison with original digital mammograms alone and in combination with tomosynthesis images in an experimental setting. Breast Cancer, 30(1), pp.46-55

    work page

  30. [30]

    and Igarashi, T., 2023

    Uematsu, T., Nakashima, K., Harada, T.L., Nasu, H. and Igarashi, T., 2023. Comparisons between artificial intelligence computer-aided detection synthesized mammograms and digital mammograms when used alone and in combination with tomosynthesis images in a virtual screening setting. Japanese Journal of Radiology, 41(1), pp.63-70

    work page 2023

  31. [31]

    SCREENet: A Multi-view Deep Convolutional Neural Network for Classification of High-resolution Synthetic Mammographic Screening Scans

    Seyyedi, S., Wong, M.J., Ikeda, D.M. and Langlotz, C.P., 2020. SCREENet: A multi-view deep convolutional neural network for clas- sification of high-resolution synthetic mammographic screening scans. arXiv preprint arXiv:2009.08563

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  32. [32]

    and Diaz, O., 2018, July

    Trovini, G., Napoli, C., Marti, R., Martin, A., Bria, A., Marrocco, C., Molinara, M., Tortorella, F. and Diaz, O., 2018, July. A deep learning framework for micro-calcification detection in 2D mammography and C- view. In 14th International Workshop on Breast Imaging (IWBI 2018) (V ol. 10718, pp. 267-275). SPIE

    work page 2018

  33. [33]

    and Rubin, D.L., 2017

    Lee, R.S., Gimenez, F., Hoogi, A., Miyake, K.K., Gorovoy, M. and Rubin, D.L., 2017. A curated mammography data set for use in computer-aided detection and diagnosis research. Scientific data, 4(1), pp.1-9

    work page 2017

  34. [34]

    and Cardoso, J.S., 2012

    Moreira, I.C., Amaral, I., Domingues, I., Cardoso, A., Cardoso, M.J. and Cardoso, J.S., 2012. Inbreast: toward a full-field digital mammographic database. Academic radiology, 19(2), pp.236-248

    work page 2012

  35. [35]

    A Neural Algorithm of Artistic Style

    Gatys, L.A., Ecker, A.S. and Bethge, M., 2015. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  36. [36]

    and Yang, M.H., 2017

    Li, Y ., Fang, C., Yang, J., Wang, Z., Lu, X. and Yang, M.H., 2017. Universal style transfer via feature transforms. Advances in neural information processing systems, 30

    work page 2017

  37. [37]

    and Xu, C.,

    Deng, Y ., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L. and Xu, C.,

    work page

  38. [38]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp

    Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 11326-11336)

    work page

  39. [39]

    and Fried, O., 2022

    Avrahami, O., Lischinski, D. and Fried, O., 2022. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 18208- 18218)

    work page 2022

  40. [40]

    and Heng, P.A., 2019, July

    Chen, C., Dou, Q., Chen, H., Qin, J. and Heng, P.A., 2019, July. Syn- ergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation. In Proceedings of the AAAI conference on artificial intelligence (V ol. 33, No. 01, pp. 865-872)

    work page 2019

  41. [41]

    and Tao, Q., 2019, October

    Yan, W., Wang, Y ., Gu, S., Huang, L., Yan, F., Xia, L. and Tao, Q., 2019, October. The domain shift problem of medical image segmentation and vendor-adaptation by Unet-GAN. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 623-631). Cham: Springer International Publishing

    work page 2019

  42. [42]

    and Zhou, K., 2020

    Shen, R., Yao, J., Yan, K., Tian, K., Jiang, C. and Zhou, K., 2020. Unsupervised domain adaptation with adversarial learning for mass detection in mammogram. Neurocomputing, 393, pp.27-37

    work page 2020

  43. [43]

    and Gu, Y ., 2025

    Li, Z., Cui, Z., Zhang, L., Wang, S., Lei, C., Ouyang, X., Chen, D., Zhao, X., Liu, C., Liu, Z. and Gu, Y ., 2025. Domain generalization for mammographic image analysis with contrastive learning. Computers in Biology and Medicine, 185, p.109455

    work page 2025

  44. [44]

    and Fei-Fei, L., 2009, June

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K. and Fei-Fei, L., 2009, June. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (pp. 248-255)

    work page 2009

  45. [45]

    and Grimm, L.J., 2024

    Coffey, K., Dodelzon, K., Dialani, V ., Joe, B.N., Omofoye, T.S., Thomas, C. and Grimm, L.J., 2024. Survey on current utilization and perception of synthesized mammography. Journal of Breast Imaging, 6(6), pp.636- 645

    work page 2024

  46. [46]

    and Lo, J.Y ., 2019

    Hou, R., Mazurowski, M.A., Grimm, L.J., Marks, J.R., King, L.M., Maley, C.C., Hwang, E.S.S. and Lo, J.Y ., 2019. Prediction of upstaged ductal carcinoma in situ using forced labeling and domain adaptation. IEEE Transactions on Biomedical Engineering, 67(6), pp.1565-1572

    work page 2019

  47. [47]

    and Bengio, S., 2019

    Raghu, M., Zhang, C., Kleinberg, J. and Bengio, S., 2019. Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems, 32

    work page 2019

  48. [48]

    and Zwiggelaar, R., 2019

    George, M., Chen, Z. and Zwiggelaar, R., 2019. Multiscale connected chain topological modelling for microcalcification classification. Com- puters in biology and medicine, 114, p.103422

    work page 2019

  49. [49]

    and Lehotsk ´a, V ., 2024

    Mracko, A., Cimr ´ak, I., Vanovcanov´a, L. and Lehotsk ´a, V ., 2024. Deep Learning in Breast Calcifications Classification: Analysis of Cross- Database Knowledge Transferability. In BIOSTEC (1) (pp. 527-535)

    work page 2024

  50. [50]

    and Tortorella, F., 2023, September

    Kassahun, R.K., Molinara, M., Bria, A., Marrocco, C. and Tortorella, F., 2023, September. Breast mass detection and classification using transfer learning on OPTIMAM dataset through RadImageNet weights. In International Conference on Image Analysis and Processing (pp. 71- 82). Cham: Springer Nature Switzerland

    work page 2023

  51. [51]

    and Iglesias, J.E., 2018, August

    Morrell, S., Wojna, Z., Khoo, C.S., Ourselin, S. and Iglesias, J.E., 2018, August. Large-scale mammography CAD with deformable conv-nets. In International Workshop on Reconstruction and Analysis of Moving Body Organs (pp. 64-72). Cham: Springer International Publishing

    work page 2018

  52. [52]

    and Zheng, B., 2022

    Jones, M.A., Faiz, R., Qiu, Y . and Zheng, B., 2022. Improving mam- mography lesion classification by optimal fusion of handcrafted and deep transfer learning features. Physics in Medicine & Biology, 67(5), p.054001

    work page 2022

  53. [53]

    Synthesizing lesions using contextual GANs improves breast cancer classification on mammograms

    Wu, E., Wu, K. and Lotter, W., 2020. Synthesizing lesions using contextual GANs improves breast cancer classification on mammograms. arXiv preprint arXiv:2006.00086

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  54. [54]

    and Brentnall, A.R., 10 GENERIC COLORIZED JOURNAL, VOL

    Damiani, C., Kalliatakis, G., Sreenivas, M., Al-Attar, M., Rose, J., Pudney, C., Lane, E.F., Cuzick, J., Montana, G. and Brentnall, A.R., 10 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2023

    work page 2023

  55. [55]

    Radiology, 307(5), p.e222679

    Evaluation of an AI model to assess future breast cancer risk. Radiology, 307(5), p.e222679

    work page

  56. [56]

    and Warren, L.M., 2024

    Ellis, S., Gomes, S., Trumble, M., Halling-Brown, M.D., Young, K.C., Chaudhry, N.S., Harris, P. and Warren, L.M., 2024. Deep learning for breast cancer risk prediction: application to a large representative UK screening cohort. Radiology: Artificial Intelligence, 6(4), p.e230431

    work page 2024

  57. [57]

    and Matela, N., 2025

    Mendes, J., Oliveira, B., Ara ´ujo, C., Galr˜ao, J., Garcia, N.C. and Matela, N., 2025. You get the best of both worlds? Integrating deep learning and traditional machine learning for breast cancer risk prediction. Computers in Biology and Medicine, 187, p.109733

    work page 2025

  58. [58]

    and Wang, M., 2021

    Lotter, W., Diab, A.R., Haslam, B., Kim, J.G., Grisot, G., Wu, E., Wu, K., Onieva, J.O., Boyer, Y ., Boxerman, J.L. and Wang, M., 2021. Robust breast cancer detection in mammography and digital breast tomosynthesis using an annotation-efficient deep learning approach. Nature medicine, 27(2), pp.244-249

    work page 2021

  59. [59]

    and Marti, R., 2020

    Agarwal, R., Diaz, O., Yap, M.H., Llado, X. and Marti, R., 2020. Deep learning for mass detection in full field digital mammograms. Computers in biology and medicine, 121, p.103774

    work page 2020

  60. [60]

    and Wei, F., 2022

    Liu, Z., Hu, H., Lin, Y ., Yao, Z., Xie, Z., Wei, Y ., Ning, J., Cao, Y ., Zhang, Z., Dong, L. and Wei, F., 2022. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 12009-12019)

    work page 2022