CoChat
Sign Up
Cited by 114AI in cancer detectionPrivacy-Preserving Technologies in DataDomain Adaptation and Few-Shot Learning

Memory-aware curriculum federated learning for breast cancer classification

Open access
Jiménez-Sánchez, Amelia, Tardy, Mickael, González Ballester, Miguel A., Mateus, Diana, Piella, Gemma·2023·Computer Methods and Programs in Biomedicine
DOI 10.1016/j.cmpb.2022.107318
Open this paper in CoChat — free accountOpen PDF

Abstract

Background and objective For early breast cancer detection, regular screening with mammography imaging is recommended. Routine examinations result in datasets with a predominant amount of negative samples. The limited representativeness of positive cases can be problematic for learning Computer-Aided Diagnosis (CAD) systems. Collecting data from multiple institutions is a potential solution to mitigate this problem. Recently, federated learning has emerged as an effective tool for collaborative learning. In this setting, local models perform computation on their private data to update the global model. The order and the frequency of local updates influence the final global model. In the context of federated adversarial learning to improve multi-site breast cancer classification, we investigate the role of the order in which samples are locally presented to the optimizers.Methods We define a novel memory-aware curriculum learning method for the federated setting. We aim to improve the consistency of the local models penalizing inconsistent predictions, i.e., forgotten samples. Our curriculum controls the order of the training samples prioritizing those that are forgotten after the deployment of the global model. Our approach is combined with unsupervised domain adaptation to deal with domain shift while preserving data privacy.Results Two classification metrics: area under the receiver operating characteristic curve (ROC-AUC) and area under the curve for the precision-recall curve (PR-AUC) are used to evaluate the performance of the proposed method. Our method is evaluated with three clinical datasets from different vendors. An ablation study showed the improvement of each component of our method. The AUC and PR-AUC are improved on average by 5% and 6%, respectively, compared to the conventional federated setting.Conclusions We demonstrated the benefits of curriculum learning for the first time in a federated setting. Our results verified the effectiveness of the memory-aware curriculum federated learning for the multi-site breast cancer classification. Our code is publicly available at: https://github.com/ameliajimenez/curriculum-federated-learning.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a memory-aware curriculum federated learning method, combined with unsupervised domain adaptation, for multi-site breast cancer classification on mammograms. It reports that scheduling samples forgotten after global model deployment improves federated classification performance over conventional federated learning.

Key findings

  • The proposed Fed-Align-CL achieved average AUC 0.79 and PR-AUC 0.82 across Hologic, Siemens, and GE, compared with Fed at AUC 0.75 and PR-AUC 0.77 in Table III.
  • The abstract reports that AUC and PR-AUC improved on average by 5% and 6%, respectively, compared to the conventional federated setting.
  • The best initialization strategy was pretrained weights from Wu et al. [51], with average AUC 0.75 in Table II, versus local model 0.53, scratch 0.63, and DDSM 0.65.
  • Statistical testing reported significant improvements for Fed-Align-CL versus Fed (AUC p=1.46E-04; PR-AUC p=9.82E-06) and versus Fed-Align (AUC p=5.65E-02; PR-AUC p=1.42E-02) in Table V.

Methods

  • The method builds on federated averaging with federated adversarial domain adaptation, differential privacy via Gaussian noise on shared weights/embeddings, and a memory-aware curriculum data scheduler. The scheduler scores samples higher when local prediction was correct but global prediction became incorrect after aggregation, using ρ=2.0 for forgotten samples and 1.0 otherwise.
  • Experiments used three Full Field Digital Mammography datasets from different vendors: Hologic (1460 subjects; 730 benign/normal, 730 malignant), Siemens/INBreast (410 subjects; 287 benign/normal, 123 malignant), and GE (852 subjects; 421 benign/normal, 431 malignant). Each dataset was split approximately 70%:10%:20% for training, validation, and test.
  • The feature extractor was a ResNet-22 pretrained with Wu et al. [51] weights; models were trained 5 times with different seeds, using Adam for 50 epochs, initial learning rate 1e-5, Q=120 optimization iterations, and Gaussian noise variance σ²=0.001. Evaluation metrics were ROC-AUC and PR-AUC.

Limitations

  • The paper states there is still room for improvement when dealing with small datasets, such as Siemens, and that some sites trade a significant gain in generalization for a local performance drop.
  • The paper states that the computational cost of analyzing high-resolution images imposes resource limitations, that Grad-CAM interpretation is subjective, and that patient reidentification or information leakage cannot be discarded unless anonymization destroys data fidelity.
  • Evident from the text: the evaluation is limited to three mammography datasets from three vendors, with two private datasets and one public dataset, and the method relies on full mammograms and classification labels rather than region-wise ground truth.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.cmpb.2022.107318

Cites

80 of 80 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 26 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The individual models (Single) achieve an average AUC of 0.83 for the three sites.

    “Second, the individual models (Single) achieve an average AUC of 0.83 for the three sites.”
    Quote in the paperNot a substantial claim
  • Findingown data

    An ablation study showed the improvement of each component of our method.

    “An ablation study showed the improvement of each component of our method.”
    Quote in the paperSubstantial claim
  • Findingnone given

    The performance for the different domains of Fed and Fed-CL can be uneven.

    “However, the performance for the different domains of these two methods can be uneven.”
    Quote in the paperSubstantial claim
  • Findingnone given

    Classification results were better when all local models were initialized either randomly or pretrained on a single dataset (DDSM) than when each of them was pretrained on a small private dataset.

    “Interestingly as well, classification results were better when all local models were initialized either randomly or pretrained on a single dataset (DDSM) than when each of them was pretrained on a small private dataset.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Some qualitative results using Gradient-weighted Class Activation Mappings (Grad-CAMs) show the generalisation effect of the federated approaches.

    “Some qualitative results using Gradient-weighted Class Activation Mappings (Grad-CAMs) [58] show the generalisation effect of the federated approaches in Fig. 7 (Suppl.Material).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [58] Tse, 2007, Calcification in breast lesions: pathologists’ perspective, Journal of Clinical Pathology source

      Not checked in the free run.

  • May combine two claims. 19
  • Classified as probably not a substantial claim. 18

  • May combine two claims.

    Findingown data

    The AUC and PR-AUC are improved on average by 5% and 6%, respectively, compared to the conventional federated setting.

    “The AUC and PR-AUC are improved on average by 5% and 6%, respectively, compared to the conventional federated setting.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Our proposed Fed-Align-CL achieves on average the highest AUC and PR-AUC.

    “Finally, we can see that our proposed Fed-Align-CL achieves on average the highest AUC and PR-AUC.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    We combine our data scheduler with federated adversarial learning and show that the association is beneficial for the classification and the alignment between domain pairs.

    “We combine our data scheduler with federated adversarial learning and show that the association is beneficial for the classification and the alignment between domain pairs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    The best performing model is Mix, which is trained with mammography images and their corresponding annotations from all sites, thus, not preserving privacy.

    “As expected, the best performing model is Mix, which is trained with mammography images and their corresponding annotations from all sites, thus, not preserving privacy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    When comparing our performance to other works on the publicly available INBreast dataset (Siemens), we achieve an AUC comparable to Wu et al., but lower than two other works with an AUC of 0.95.

    “When comparing our performance to other works on the publicly available INBreast dataset (Siemens), we achieve an AUC comparable to [51], but lower than [50], [57] with an AUC of 0.95.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [51] Selvaraju, 2017, Grad-CAM: visual explanations from deep networks via gradient-based localization

      Not checked in the free run.

    • cites [50] Samuli, 2017, Temporal ensembling for semi-supervised learning

      Not checked in the free run.

    • cites [57] Tang, 2018, Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Fig. 6 in Supplementary Material shows a similar behaviour for the penultimate classification layer.

    “Fig. 6 in Supplementary Material shows a similar behaviour for the penultimate classification layer.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Some sites preserve the performance on the local test set after collaborative training, while others trade a significant gain in generalization for a local performance drop.

    “While some sites preserve the performance on the local test set after collaborative training, others trade a significant gain in generalization for a local performance drop.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    With the same collaborative training, there is no onsite performance drop for the GE and Hologic sites, which benefit from a significant generalization improvement at no cost.

    “With the same collaborative training, there is no onsite performance drop for the GE and Hologic sites, which benefit from a significant generalization improvement at no cost.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    The computational cost of analyzing high-resolution images imposes some resource limitations on the federated scenario.

    “The computational cost of analyzing high-resolution images imposes some resource limitations on our federated scenario.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The small size of the lesion requires high-resolution images showing tissue details to provide a precise diagnosis.

  • May combine two claims.

    Findingown data

    Fed-Align-CL reduces activations on the boundary or outside the breast and better focuses on lesions.

    “Our method (Fed-Align-CL) reduces activations on the boundary or outside the breast and better focuses on lesions.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductown data

    We applied one-way ANOVA followed by Bonferroni's post-hoc t-test comparisons tests to evaluate for differences in the performance of the federated approaches and reported the p-values in Table 5.

    “We applied one-way ANOVA followed by Bonferroni's post-hoc t-test comparisons tests to evaluate for differences in the performance of the federated approaches and reported the p-values in Table 5 (Suppl.Material).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    We train our models 5 times with different seed initialization for the classifier Cls and domain discriminator D.

    “We train our models 5 times with different seed initialization for the classifier Cls and domain discriminator D.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    We employ as feature extractor F the architecture proposed by Wu et al., a ResNet-22 that is adapted to take high-resolution images (∼ 4 megapixels) as input, and achieved state-of-the-art classification performance for breast cancer screening.

    “We employ as feature extractor F the architecture proposed by Wu et al. [51], a ResNet-22 [52] that is adapted to take high-resolution images (∼ 4 megapixels) as input, and achieved state-of-the-art classification performance for breast cancer screening.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [51] Selvaraju, 2017, Grad-CAM: visual explanations from deep networks via gradient-based localization

      Not checked in the free run.

    • cites [52] Sheller, 2018, Multi-institutional deep learning modeling without sharing patient data: a feasibility study on brain tumor segmentation

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    UDA methods force the model to learn domain-agnostic features through adversarial learning or a specific type of batch normalization.

    “UDA methods force the model to learn domain-agnostic features through adversarial learning [6] or a specific type of batch normalization [7].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [6] Carlucci, 2017, Just dial: domain alignment layers for unsupervised domain adaptation

      Not checked in the free run.

    • cites [7] Darzidehkalani, 2022, Federated learning in medical imaging: Part II: methods, challenges, and considerations, Journal of the American College of Radiology source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    FedAvg is a generalization of FedSGD, which allows local nodes to perform more than one batch update on local data and exchanges the updated weights rather than the gradients.

    “Federated averaging FedAvg is a generalization of FedSGD [46], which allows local nodes to perform more than one batch update on local data and exchanges the updated weights rather than the gradients.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [46] Ribli, 2018, Detecting and classifying lesions in mammograms with deep learning, Scientific Reports source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    This behaviour is expected because their model was trained with a very-large private dataset.

    “This behaviour is expected because their model was trained with a very-large private dataset.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Our scoring function is related to temporal ensembling and consistency training, which are used in semi-supervised setups.

    “Our scoring function is related to temporal ensembling [44] and consistency training [45], which are used in semi-supervised setups.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [44] D. Peterson, P. Kanani, V.J. Marathe, Private federated learning with domain adaptation, arXiv preprint arXiv:1912.06733(2019).

      Not checked in the free run.

    • cites [45] Redmon, 2016, You only look once: unified, real-time object detection

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    We do not use any site-specific image filtering to compensate the domain shift, and we apply the same preprocessing to the images from the three sites.

    “We do not use any site-specific image filtering to compensate the domain shift, and we apply the same preprocessing to the images from the three sites.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The later two works rely on regionwise ground truth: the first leveraging ROI localization and the second one using patch pretraining.

    “However, the later two works rely on regionwise ground truth: the first leveraging ROI localization and the second one using patch pretraining.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [50] Samuli, 2017, Temporal ensembling for semi-supervised learning

      Not checked in the free run.

    • cites [57] Tang, 2018, Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Differential privacy perturbs the parameters of each local model by purposely adding noise before uploading them to the server for aggregation.

    “Differential privacy perturbs the parameters of each local model by purposely adding noise before uploading them arXiv:2107.02504v2[cs.CV] 6 Jan 2023 to the server for aggregation.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [9] Dong, 2021, Federated contrastive learning for decentralized unlabeled medical images

      Not checked in the free run.

    • cites [10] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: transformers for image recognition at scale, International Conference on Learning Representations (2021).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    In the federated setting, data is stored locally and cannot be shared.

    “However, in the federated setting, data is stored locally and cannot be shared.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Institutional board approvals were obtained for each of the datasets.

    “Institutional board approvals were obtained for each of the datasets.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    CL has already demonstrated an improved performance in medical image classification tasks, such as thoracic disease, skin disease, proximal femur fractures and breast screening classification.

    “CL has already demonstrated an improved performance in medical image classification tasks, such as thoracic disease [36], skin disease [37], proximal femur fractures [12], [38] and breast screening classification [39].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [36] Maicas, 2018, Training medical image analysis systems like radiologists

      Not checked in the free run.

    • cites [37] M. Mancini, Z. Akata, E. Ricci, B. Caputo, Towards recognizing unseen categories in unseen domains, European Conference on Computer Vision. Springer, Cham, 2020. source

      Not checked in the free run.

    • cites [12] Dwork, 2014, The algorithmic foundations of differential privacy, Foundations and Trends in Theoretical Computer Science

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Such inconsistent performance is a known problem of deep learning methods for mammography classification.

    “Such inconsistent performance is a known problem of deep learning methods for mammography classification [56].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [56] Sung, 2021, Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA: A Cancer Journal for Clinicians

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    The new order of training samples has an impact on the local minimum achieved by the optimizer, leading to an improvement in the classification accuracy.

    “This new order has an impact on the local minimum achieved by the optimizer, leading to an improvement in the classification accuracy.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    One-way ANOVA followed by Bonferroni's post-hoc t-test comparisons tests were applied to evaluate for differences in the performance of the federated approaches and p-values were reported.

    “In Table V, we applied one-way ANOVA followed by Bonferroni's post-hoc t-test comparisons tests to evaluate for differences in the performance of the federated approaches and reported the p-values.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Medical images collected from different healthcare providers may originate from diverse devices or imaging protocols, leading to a domain shift problem when deploying algorithms trained on a single domain.

    “Medical images collected from different healthcare providers may originate from diverse devices or imaging protocols, leading to a domain shift problem when deploying algorithms trained on a single domain.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Domain alignment keeps the performance from dropping, but there are fewer corrections for the curriculum to account for after the update.

    “In this sense, domain alignment keeps the performance from dropping, but there are fewer corrections for the curriculum to account for after the update.”
    Quote in the paperNot a substantial claim
28 of 28

  • The paper relies on this without support.

    The limited representativeness of positive cases can be problematic for learning-based CAD systems.

    “The limited representativeness of positive cases can be problematic for learning-based Computer-Aided Diagnosis (CAD) systems.”

  • The paper relies on this without support.

    A potential solution to mitigate the limited representativeness of positive cases and to increase the size of the annotated dataset is to employ data coming from multiple institutions.

    “A potential solution to mitigate this problem and to increase the size of the annotated dataset is to employ data coming from multiple institutions.”

  • The paper relies on this without support.

    Sharing medical information across institutions is challenging in terms of privacy, technical and legal issues.

    “However, sharing medical information across (international) institutions is challenging in terms of privacy, technical and legal issues.”

  • The authors state this as an assumption.

    Each site has local data storage, a computing server, and a memory-aware CL module.

    “We assume that each site has local data storage, a computing server, and a memory-aware CL module.”

  • The paper relies on this without support.

    Both global and local models share the same architecture.

    “In this type of FL setting, both global and local models share the same architecture, and it is common to share the model weights and aggregate them at the central server.”

  • The authors state this as an assumption.

    Each site has mammography systems of different vendors.

    “In this work, we assume the most challenging scenario, in which we consider that each site has mammography systems of different vendors (see Fig. 1).”

  • The authors state this as an assumption.

    Data owners collaboratively train a global model without sharing their image data.

    “We assume that data owners collaboratively train a global model without sharing their image data.”

  • The paper relies on this without support.

    FL allows for collaboratively training a model without sharing private data from different sites.

    “FL allows for collaboratively training a model without sharing private data from different sites.”

  • The paper relies on this without support.

    The small size of the lesion requires high-resolution images showing tissue details to provide a precise diagnosis.

    “Indeed, the small size of the lesion requires high-resolution images showing tissue details to provide a precise diagnosis.”

  • The paper relies on this without support.

    Standard clinical protocols do not include the lesion delineation.

    “However, the level of annotation required would be much higher as standard clinical protocols do not include the lesion delineation.”
10 of 10

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Early cancer diagnosis accuracy has been shown to significantly improve with federated learning.

    “Early cancer diagnosis accuracy has been shown to significantly improve with FL.”

    From Federated Learning in Smart Healthcare: A Comprehensive Review on Privacy, Security, and Predictive Analytics with IoT Integration (2024) · cites this paper as [4]

    The Abstract reports improved breast-cancer classification metrics for the proposed method, and the results section (‘a)’, page not provided) says its improvement over other federated approaches was statistically significant. The source does not establish that federated learning generally improves early cancer diagnosis accuracy: it tests a specific breast-cancer classification method, primarily against other federated methods. (LLM, conf=0.88)

    Experimental — accuracy not yet measured

  • Using differential privacy, a memory-aware curriculum federated learning strategy increased the precision of local models in breast tumor prediction by up to 20% while protecting patient information confidentiality.

    “For example, using sophisticated privacy-preserving techniques like differential privacy, a memory-aware curriculum FL strategy increased the precision of local models in breast tumor prediction by up to 20% while protecting patient information confidentiality [4].”

    From Federated Learning in Smart Healthcare: A Comprehensive Review on Privacy, Security, and Predictive Analytics with IoT Integration (2024) · cites this paper as [4]

    The paper says it uses differential privacy for data protection (section B), and Table III reports improved average precision–recall AUC for its federated method, from 0.77 to 0.82. It does not report a 20% increase in local-model precision. The Discussion also cautions that patient reidentification or information leakage cannot be ruled out, so the confidentiality claim is too strong. (LLM, conf=0.89)

    Experimental — accuracy not yet measured

  • In a multi-hospital diabetes research, federated learning decreased the risk of data breaches by 40% while also improving predicted outcomes by 15% in electronic health record analysis as compared to centralized machine learning.

    “For example, in a multi-hospital managing diabetes research, FL decreased the risk of data breaches by 40% while also improving predicted outcomes by 15% in EHR analysis as compared to centralized ML [4,34].”

    From Federated Learning in Smart Healthcare: A Comprehensive Review on Privacy, Security, and Predictive Analytics with IoT Integration (2024) · cites this paper as [4]

    The Datasets section describes mammography images from three vendors, and the Discussion concerns breast-cancer classification and privacy considerations. Searches for diabetes, EHR analysis, data breaches, and the claimed 40% and 15% figures found no passage reporting the attributed comparison with centralized machine learning. (LLM, conf=0.78)

    Experimental — accuracy not yet measured

Cites

Flagged first, then by citations

Research works printed but not linked: 3

  1. Y. Freund, R. Schapire, N. Abe, A short introduction to boosting, Journal Japanese Society for Artificial Intelligence 14 (771–780) (1999) 1612.
  2. Paszke, 2019, PyTorch: an imperative style, high-performance deep learning library, Adv. Neural Inf. Process. Syst., 32
  3. Wang, 2021, Addressing class imbalance in federated learning, Vol. 35, 10165

Cited by

Sorted by citations

Related

Authors
Jiménez-Sánchez, A., Tardy, M., González Ballester, M. A., Mateus, D., & Piella, G.
Affiliations
  • Universitat Pompeu Fabra (ES)
  • École Centrale de Nantes (FR)
Corresponding author
Jiménez-Sánchez, Amelia
ORCID
Venue
Computer Methods and Programs in Biomedicine
Publisher
Elsevier BV
Published
2023-02
Volume
229
Pages
107318
ISSN
0169-2607 (ISSN-L)
DOI
10.1016/j.cmpb.2022.107318
OpenAlex
W3182336762
Cited by
114
References
80
Citations by year
2026: 9 · 2025: 33 · 2024: 34 · 2023: 33 · 2022: 4 · 2021: 1
FWCI
8.55
Type
journal article
Language
en
Open access
Licence
Locations
  • Computer Methods and Programs in Biomedicine (publishedVersion)
  • PubMed (publishedVersion)
  • HAL (submittedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • IT University Of Copenhagen (IT University of Copenhagen) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Nvidia
  • European Commission (713673)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
federated learning, curriculum learning, breast cancer classification, mammography, computer-aided diagnosis, domain adaptation, domain shift, data privacy, multi-site data, class imbalance
MeSH terms
Awareness, Cognition, Curriculum, Learning, Mammography, Neoplasms

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review