Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Pretraining domain, not objective, sets the privacy–utility trade-off in medical imaging

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

2026-08-03 07:33 UTC pith:76CDKXCA

load-bearing objection Big, careful DP-SGD benchmark, but the domain-vs-objective headline is unidentifiable from the three arms; narrower comparisons are solid. the 4 major comments →

arxiv 2601.19618 v2 pith:76CDKXCA submitted 2026-01-27 cs.CV cs.AIcs.LG

The pretraining domain outweighs the training objective in setting the privacy-utility trade-off of differentially private medical image analysis

Soroosh Tayebi Arasteh , Mina Farajiamiri , Mahshad Lotfinia , Behrus Hinrichs-Puladi , Jonas Bienzeisler , Mohamed Alhaskir , Mirabela Rusu , Christiane Kuhl
show 2 more authors
Sven Nebelung Daniel Truhn
This is my paper
classification cs.CV cs.AIcs.LG
keywords differential privacyDP-SGDpretrainingself-supervised learningchest radiographsmedical image analysisinitializationfairness
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 tries to establish that, when training medical image models with differentially private SGD, the choice of pretraining data matters more than the choice of pretraining objective. Across five external chest radiograph datasets and three privacy budgets, supervised pretraining on chest X-rays (MIMIC-CXR) ranked first in 24 of 25 dataset–budget combinations, beating both ImageNet and self-supervised DINOv3 initializations. The advantage grew as privacy became stricter, from 2.5 to 14.6 points of macro-averaged area under the ROC curve, and the domain effect was 2.2–3.4 times larger than the objective effect. A sympathetic reader would care because initialization is currently the strongest practical remedy for the accuracy loss that differential privacy inflicts, and this result points to domain-specific pretraining as the lever that matters most.

Core claim

The central claim is that under full-model DP-SGD, what a model was pretrained on outweighs how it was pretrained. Using ConvNeXt classifiers fine-tuned with DP-SGD, the authors compare three initializations: supervised ImageNet, self-supervised DINOv3, and supervised pretraining on MIMIC-CXR using the same five diagnostic labels as the downstream tasks. The chest-domain supervised initialization achieves the highest macro-averaged AUROC across nearly all settings, remains closest to non-private performance, keeps demographic disparities smallest, and also yields the best cross-dataset generalization; privately pretraining the chest corpus costs about five points but still beats every public

What carries the argument

The central instrument is a three-way initialization contrast run under DP-SGD: supervised ImageNet weights, self-supervised DINOv3 weights, and supervised weights pretrained on MIMIC-CXR with the downstream labels. The authors read the first contrast (DINOv3 vs ImageNet) as an objective-only change and the second (MIMIC-CXR vs the two generic arms) as a domain-only change, allowing them to estimate the 'domain effect' and 'objective effect' on the privacy–utility trade-off. The comparison is carried out on more than 590,000 chest radiographs across five external datasets, four countries, three privacy ranges, two model sizes, and a training-set-size sweep.

Load-bearing premise

The three initialization arms are assumed to isolate 'domain' and 'objective' independently, but the chest-domain arm is also task-aligned and the two generic arms differ in corpus scale and pretraining recipe, so the measured 'domain effect' may partly be an effect of label alignment and data volume.

What would settle it

Train a self-supervised model on MIMIC-CXR (chest domain, no labels) and compare it with the supervised MIMIC-CXR arm under the same privacy budget and downstream training setup. If the self-supervised chest model matches the supervised chest model, the advantage is domain; if it falls to ImageNet-level performance, the advantage is task-aligned labels rather than domain alone.

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

If this is right

  • If the claim holds, practitioners with access to a large public labeled medical dataset should prefer supervised in-domain pretraining over generic or self-supervised weights when training under DP.
  • Self-supervised DINOv3 initialization is a viable fallback in the absence of labeled domain data, consistently beating ImageNet under privacy but never closing the gap to domain-specific supervision.
  • Pretraining the domain corpus itself under DP is practical: it costs roughly five AUROC points but still outperforms every public initialization when the downstream model is trained under privacy.
  • In-domain pretraining also mitigates demographic fairness degradation under DP, especially for age-based disparities, and improves cross-dataset generalization.
  • The relative benefit of domain-specific pretraining grows as the privacy budget tightens and training data shrink, making it the most valuable lever in high-privacy, low-data settings.

Where Pith is reading between the lines

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

  • The paper's 'domain effect' likely bundles the pretraining domain with task-aligned labels and corpus scale: MIMIC-CXR pretraining uses the same five diagnostic labels as the downstream tasks, and the generic arms differ in data volume and SSL recipe, so the 2.2–3.4× factor may partly reflect label alignment rather than domain alone.
  • A natural testable extension is to pretrain a self-supervised model on MIMIC-CXR and compare it with the supervised MIMIC-CXR arm; if the self-supervised chest model matches the supervised one, the advantage is domain, and if it drops toward ImageNet, the advantage is supervision or label alignment.
  • For clinical domains that lack a large public labeled cohort, the practical takeaway is inverted: self-supervised generic initialization remains the best available option, and the paper's hierarchy may not transfer to such domains.
  • The finding that initialization beats model capacity and training-set size under DP suggests that future privacy research should invest at least as much in pretraining strategies as in DP optimization mechanics.

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 / 4 minor

Summary. The paper presents a large-scale empirical evaluation of three initialization strategies—supervised ImageNet, self-supervised DINOv3, and supervised pretraining on MIMIC-CXR—for ConvNeXt classifiers trained with DP-SGD on chest radiograph classification. Across five external datasets, four privacy regimes, two model sizes, and several robustness probes, the authors report that domain-specific supervised pretraining consistently outperforms the two natural-image initializations, that DINOv3 outperforms ImageNet under DP, and that in-domain pretraining also improves fairness and cross-dataset generalization. The headline claim is that the pretraining domain matters more than the pretraining objective, quantified as a 2.2–3.4× larger effect.

Significance. If the decomposition were valid, this would be an important practical result: it would identify the pretraining domain, not the self-supervised/supervised objective, as the key lever for preserving utility under DP-SGD in medical imaging. The study's strengths are substantial: five external test cohorts, four countries, 1,000 patient-level bootstrap resamples, paired statistical tests with FDR correction, reproducible code, and public pretrained weights. These features make the descriptive comparisons—e.g., DINOv3 > ImageNet, MIMIC-supervised > DINOv3—credible within the measured cells. However, the central decomposition claim is not supported by the experimental design, so the paper's primary contribution as stated cannot be accepted.

major comments (4)
  1. [§4.4, Figure 1b] The claim that 'pretraining domain outweighs training objective' requires two clean contrasts: objective held fixed while domain varies, and domain held fixed while objective varies. The three arms provide neither. DINOv3 (trained on LVD-1689M) vs ImageNet (trained on ImageNet-1M) confounds objective with pretraining corpus scale and SSL recipe; MIMIC-CXR supervised pretraining vs ImageNet confounds domain with task-label alignment, since MIMIC uses the same five downstream diagnostic labels. The Discussion's fourth limitation (unlabeled medical SSL not evaluated) concedes the missing cell. Without a chest-domain self-supervised arm and a same-corpus supervised/self-supervised pair, the observed hierarchy is equally consistent with label/task alignment or corpus scale driving the gains.
  2. [Abstract; §4.6] The quantitative factor '2.2 to 3.4' for domain vs objective effect appears only in the abstract and is never defined or derived in Methods. No equations, no model, no sensitivity analysis. The paper reports raw AUROC differences but never states how 'domain effect' and 'objective effect' are computed from the three cells. This is not reproducible and cannot be checked. Even if the confounding were fixed, the effect-size definition would need to be explicit and justified.
  3. [Table 1; §4.3] Privacy budgets are not matched across initialization strategies; ε is reported at each model's convergence step and grouped into broad ranges (0<ε<1, 1<ε<3, 3<ε<10). Within a range, DINOv3, ImageNet, and MIMIC-supervised models may have different achieved ε, potentially biasing the comparison. This is especially problematic for the claim that the MIMIC lead 'grew from 2.5 to 14.6 points as the budget tightened'—the tightened cells may not be at the same ε. The paper should either match ε across arms (e.g., by tuning noise) or include ε as a regressor/covariate.
  4. [Abstract; §4.4] The abstract states 'five initializations that vary the objective and the domain independently,' but the paper evaluates three initialization regimes. This is not a typo: it suggests a design that does not exist. The missing two cells are exactly the ones needed for the claimed independent variation. Please correct the abstract and either add the missing initialization arms or explicitly retract the 'independently' language.
minor comments (4)
  1. [Table 1 vs §4.2] MIMIC-CXR is described as n=213,921 in Table 1 and the Methods, but §4.2 and §4.4 give the MIMIC training set as n=170,153. Clarify whether the total includes test/validation or whether these are different splits.
  2. [Supplementary Tables 2–3] The spelling 'DinoV3' is inconsistent with 'DINOv3' in the main text. Also, in Supplementary Table 2, the Non-DP row for 'Age' appears misaligned: values like '5.9 0.8' and '27.2 5.9' are not formatted as the other rows.
  3. [Figure 2 caption] The caption says 'Privacy budgets (ε) are shown on the x-axis using the achieved ε values for each trained model,' but the main text and Table 1 report only ranges. The figure appears to plot a continuous x-axis; it would be clearer to show actual ε values for each model, or explicitly state why ranges are used.
  4. [§4.3] The statement that the clipping norm was 'set to 4 for ConvNeXt-Small and 3.5 for ConvNeXt-Tiny' is given without justification. Since clipping interacts with both privacy accounting and optimization, please report the noise multiplier and effective sampling probability, and discuss the sensitivity to clipping.

Circularity Check

0 steps flagged

No significant circularity; the paper is an empirical benchmark with rankings measured on held-out external data. The domain/objective decomposition is confounded, but confounding is a validity issue, not circular reasoning.

full rationale

The paper's central claims are empirical rankings obtained from held-out test sets across five external chest radiograph datasets; no derivation chain is offered that reduces a predicted quantity to a fitted input. The three initialization arms (ImageNet supervised, DINOv3 self-supervised, MIMIC-CXR supervised) are compared by direct AUROC measurement under DP-SGD, so there is no equation in which an output is defined in terms of the input or in which a fitted parameter is renamed as a prediction. The headline 'domain outweighs objective' is not identifiability-clean: the MIMIC-CXR arm differs from ImageNet in both domain and label alignment, and DINOv3 differs from ImageNet in objective, corpus scale, and SSL recipe, so the decomposition is confounded. The paper's own fourth limitation concedes that medical-domain self-supervised pretraining was not evaluated, which is the missing cell needed for a clean domain-versus-objective contrast. However, a missing baseline and a confounded interpretation are threats to construct validity, not circularity: the observed numbers are still independently measured rather than being forced by definition or by a self-citation chain. Self-citations appear in motivational framing and in support of prior results about the importance of initialization under DP, but the present experiments independently generate the central ranking, so those citations are not load-bearing. No uniqueness theorem, ansatz, or known-result renaming is used to close the argument. The abstract's claims about 'five initializations' and private pretraining of MIMIC-CXR are not reflected in the methods, but that is an internal consistency problem, not circular reasoning.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim depends on standard DP-SGD privacy accounting, the harmonized label schema, and two interpretive assumptions that are not tested: that DINOv3-vs-ImageNet differs only in objective, and that MIMIC-CXR supervised pretraining represents domain rather than task alignment. No new entities are invented; the free hyperparameters (clipping norm, learning rate, privacy binning) are standard but could interact with initialization.

free parameters (3)
  • DP-SGD clipping norm C = 4.0 (ConvNeXt-Small), 3.5 (ConvNeXt-Tiny)
    Chosen by hand; controls noise-to-signal ratio and thus utility under DP; not shown to be neutral across initializations.
  • DP learning rate = 5e-6 to 1e-5
    Selected per convergence behavior under privacy; not fixed across runs, could interact with initialization.
  • Privacy budget binning / achieved epsilon = three ranges 0-1, 1-3, 3-10; achieved epsilons vary by dataset and init
    Comparisons at 'the same privacy' are within ranges rather than at matched epsilon, reducing precision of the trade-off comparison.
axioms (5)
  • standard math Rényi differential privacy accounting correctly tracks DP-SGD privacy loss.
    Invoked in §4.3 as the privacy accountant for all reported epsilon values.
  • domain assumption Harmonized five-label schema and treating uncertain labels as negative preserves meaningful comparisons.
    Introduced in §4.2/4.4; this mapping is necessary for cross-dataset pooling but can alter label semantics.
  • domain assumption DINOv3 weights loaded from official repo represent the self-supervised objective.
    Used in §4.4/4.8; the arm also differs from ImageNet in corpus scale and training recipe, so it is not objective-only.
  • domain assumption MIMIC-CXR supervised pretraining is a representative of 'domain-specific pretraining' rather than label-aligned pretraining.
    This is the key interpretive assumption; because the same five labels are used, the arm conflates domain with task alignment.
  • domain assumption Patient-wise splits prevent information leakage.
    Stated in §4.2 and §4.4; required for the external evaluation to be valid.

pith-pipeline@v1.3.0-alltime-deepseek · 38682 in / 13621 out tokens · 139423 ms · 2026-08-03T07:33:00.638675+00:00 · methodology

0 comments
read the original abstract

Differential privacy protects the patients whose images train medical imaging models, but it lowers diagnostic accuracy, and the initialization is the strongest known remedy. Practice increasingly favors large generic self-supervised encoders. Yet the pretraining objective and the pretraining domain are confounded in existing comparisons, so which one preserves utility under privacy is unknown, and the pretraining corpus is treated as public even when it holds patient images. We trained ConvNeXt classifiers with differentially private stochastic gradient descent from five initializations that vary the objective and the domain independently, at four privacy budgets and without privacy, and evaluated them locally on more than 590,000 chest radiographs from five external datasets in four countries. Supervised pretraining on chest radiographs ranked first in 24 of 25 dataset and budget combinations. Its lead over ImageNet grew from 2.5 to 14.6 points of macro-averaged area under the receiver operating characteristic curve as the budget tightened, and the domain effect exceeded the objective effect by a factor of 2.2 to 3.4. Pretraining that corpus privately cost about 5 points and, under privacy, still beat every public initialization. Low-rank adaptation removed about half the residual gap, and in-domain pretraining raised the worst-performing demographic subgroup. Under privacy, what a model was pretrained on outweighs how it was pretrained.

Figures

Figures reproduced from arXiv: 2601.19618 by Behrus Hinrichs-Puladi, Christiane Kuhl, Daniel Truhn, Jonas Bienzeisler, Mahshad Lotfinia, Mina Farajiamiri, Mirabela Rusu, Mohamed Alhaskir, Soroosh Tayebi Arasteh, Sven Nebelung.

Figure 1
Figure 1. Figure 1: Study overview and experimental design. a Training deep learning models on sensitive medical images poses privacy risks, while real-world deployment requires robustness to cross-domain dataset shifts. The central open question addressed in this study is how model initialization influences utility, fairness, and generalization under differential privacy (DP). The study leverages a large and diverse collecti… view at source ↗
Figure 2
Figure 2. Figure 2: Diagnostic utility under DP across chest radiograph datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Demographic fairness under DP across chest radiograph datasets and initialization strategies. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-dataset generalization under DP. Absolute changes in diagnostic performance relative to the corresponding non￾private baseline (|ΔAUROC|) are shown for ConvNeXt-Small models trained on a combined multi-dataset cohort (total training n = 471,897) and evaluated independently on (a) VinDr-CXR (test n = 3,000), (b) ChestX-ray14 (test n = 25,596), (c) PadChest (test n = 22,045), (d) CheXpert (test n = 29,… view at source ↗
Figure 5
Figure 5. Figure 5: Generalization across model capacity under differential privacy. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Joint effects of training set size on diagnostic utility and demographic fairness under differential privacy on [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Differential privacy representation geometry for medical image analysis

    cs.CV 2026-03 unverdicted novelty 7.0

    DP-RGMI shows differential privacy creates a consistent utilization gap in chest X-ray models even when linear separability holds, while geometric measures like displacement and spectral dimension change in non-monoto...

Reference graph

Works this paper leans on

48 extracted references · 1 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Tayebi Arasteh, S. et al. Preserving fairness and diagnostic accuracy in private large-scale AI models for medical imaging. Commun Med 4, 46 (2024)

    2024

  2. [2]

    Kaissis, G. et al. End-to-end privacy preserving deep learning on multi-institutional medical imaging. Nat Mach Intell 3, 473–484 (2021)

    2021

  3. [3]

    & Hayes, J

    Balle, B., Cherubin, G. & Hayes, J. Reconstructing Training Data with Informed Adversaries. in 2022 IEEE Symposium on Security and Privacy (SP) 1138–1156 (IEEE, San Francisco, CA, USA, 2022). doi:10.1109/SP46214.2022.9833677

    arXiv 2022

  4. [4]

    Tayebi Arasteh, S. et al. Securing Collaborative Medical AI by Using Differential Privacy: Domain Transfer for Classification of Chest Radiographs. Radiology. Artificial Intelligence 6, e230212 (2024)

    2024

  5. [5]

    Differential Privacy

    Dwork, C. Differential Privacy. in Automata, Languages and Programming (eds Bugliesi, M., Preneel, B., Sassone, V. & Wegener, I.) vol. 4052 1–12 (Springer Berlin Heidelberg, Berlin, Heidelberg, 2006)

    2006

  6. [6]

    De, S., Berrada, L., Hayes, J., Smith, S. L. & Balle, B. Unlocking High-Accuracy Differentially Private Image Classification through Scale. Preprint at http://arxiv.org/abs/2204.13650 (2022)

    Pith/arXiv arXiv 2022

  7. [7]

    Kurakin, A. et al. Toward Training at ImageNet Scale with Differential Privacy. Preprint at http://arxiv.org/abs/2201.12328 (2022)

    Pith/arXiv arXiv 2022

  8. [8]

    Tayebi Arasteh, S. et al. Differential privacy enables fair and accurate AI-based analysis of speech disorders while protecting patient data. npj Artificial Intelligence 1, 37 (2025)

    2025

  9. [9]

    Mohammadi, M. et al. Differential privacy for medical deep learning: methods, tradeoffs, and deployment implications. npj Digit. Med. 9, (2026)

    2026

  10. [10]

    Abadi, M. et al. Deep Learning with Differential Privacy. in Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security 308–318 (ACM, Vienna Austria, 2016). doi:10.1145/2976749.2978318

    arXiv 2016

  11. [11]

    Ziller, A. et al. Reconciling privacy and accuracy in AI for medical imaging. Nat Mach Intell 6, 764–774 (2024). 33

    2024

  12. [12]

    & Topol, E

    Krishnan, R., Rajpurkar, P. & Topol, E. J. Self-supervised learning in medicine and healthcare. Nat. Biomed. Eng 6, 1346–1352 (2022)

    2022

  13. [13]

    & Song, D

    Hendrycks, D., Mazeika, M., Kadavath, S. & Song, D. Using self-supervised learning can improve model robustness and uncertainty. in NIPS’19: Proceedings of the 33rd International Conference on Neural Information Processing Systems vol. 1403 15663– 15674 (2019)

    2019

  14. [14]

    & Zhu, L

    Yang, S., Wang, H., Xing, Z., Chen, S. & Zhu, L. SegDINO: An Efficient Design for Medical and Natural Image Segmentation with DINO-V3. Preprint at https://doi.org/10.48550/arXiv.2509.00833 (2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.00833 2025

  15. [15]

    & Yang, X

    Li, Y., Wu, Y., Lai, Y., Hu, M. & Yang, X. MedDINOv3: How to adapt vision foundation models for medical image segmentation? Preprint at https://doi.org/10.48550/arXiv.2509.02379 (2025)

    doi:10.48550/arxiv.2509.02379 2025

  16. [16]

    N., Truhn, D

    Tayebi Arasteh, S., Misera, L., Kather, J. N., Truhn, D. & Nebelung, S. Enhancing diagnostic deep learning via self-supervised pretraining on large-scale, unlabeled non-medical images. Eur Radiol Exp 8, 10 (2024)

    2024

  17. [17]

    Tayebi Arasteh, S. et al. Resolution scaling governs DINOv3 transfer performance in chest radiograph classification. Preprint at https://doi.org/10.48550/arXiv.2510.07191 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.07191 2025

  18. [18]

    Deng, J. et al. ImageNet: A large-scale hierarchical image database. in 2009 IEEE Conference on Computer Vision and Pattern Recognition 248–255 (IEEE, Miami, FL, 2009). doi:10.1109/CVPR.2009.5206848

    arXiv 2009

  19. [19]

    Ke, A., Ellsworth, W., Banerjee, O., Ng, A. Y. & Rajpurkar, P. CheXtransfer: performance and parameter efficiency of ImageNet models for chest X-Ray interpretation. in Proceedings of the Conference on Health, Inference, and Learning 116–124 (ACM, Virtual Event USA, 2021). doi:10.1145/3450439.3451867

    arXiv 2021

  20. [20]

    Vaswani, A. et al. Attention Is All You Need. in NIPS’17: Proceedings of the 31st International Conference on Neural Information Processing Systems 6000–6010 (2017)

    2017

  21. [21]

    Dosovitskiy, A. et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. Preprint at http://arxiv.org/abs/2010.11929 (2021)

    Pith/arXiv arXiv 2010

  22. [22]

    Liu, Z. et al. A convnet for the 2020s. in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition 11976–11986 (2022)

    2022

  23. [23]

    Tayebi Arasteh, S. et al. The Treasure Trove Hidden in Plain Sight: The Utility of GPT-4 in Chest Radiograph Evaluation. Radiology 313, e233441 (2024)

    2024

  24. [24]

    Johnson, A. E. W. et al. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Sci Data 6, 317 (2019)

    2019

  25. [25]

    Irvin, J. et al. CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison. AAAI 33, 590–597 (2019)

    2019

  26. [26]

    Siméoni, O. et al. DINOv3. Preprint at https://doi.org/10.48550/arXiv.2508.10104 (2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.10104 2025

  27. [27]

    Nguyen, H. Q. et al. VinDr-CXR: An open dataset of chest X-rays with radiologist’s annotations. Sci Data 9, 429 (2022)

    2022

  28. [28]

    Wang, X. et al. ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 3462–3471 (2017). doi:10.1109/CVPR.2017.369

    doi:10.1109/cvpr.2017.369 2017

  29. [29]

    & de la Iglesia-Vayá, M

    Bustos, A., Pertusa, A., Salinas, J.-M. & de la Iglesia-Vayá, M. PadChest: A large chest x- ray image dataset with multi-label annotated reports. Medical Image Analysis 66, 101797 (2020)

    2020

  30. [30]

    Khader, F. et al. Artificial Intelligence for Clinical Interpretation of Bedside Chest Radiographs. Radiology 307, e220510 (2022)

    2022

  31. [31]

    Tayebi Arasteh, S. et al. Collaborative training of medical artificial intelligence models with non-uniform labels. Sci Rep 13, 6046 (2023). 34

    2023

  32. [32]

    & Truhn, D

    Tayebi Arasteh, S., Isfort, P., Kuhl, C., Nebelung, S. & Truhn, D. Automatic Evaluation of Chest Radiographs – The Data Source Matters, But How Much Exactly? in RöFo- Fortschritte auf dem Gebiet der Röntgenstrahlen und der bildgebenden Verfahren vol. 195 ab99 (Georg Thieme Verlag, RheinMain CongressCenter (RMCC) in Wiesbaden, 2023)

    2023

  33. [33]

    & Tayebi Arasteh, S

    Lotfinia, M., Tayebiarasteh, A., Samiei, S., Joodaki, M. & Tayebi Arasteh, S. Boosting multi- demographic federated learning for chest radiograph analysis using general-purpose self- supervised representations. European Journal of Radiology Artificial Intelligence 3, 100028 (2025)

    2025

  34. [34]

    Ziller, A. et al. Medical imaging deep learning with differential privacy. Sci Rep 11, 13524 (2021)

    2021

  35. [35]

    Caron, M. et al. Emerging Properties in Self-Supervised Vision Transformers. in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) 9650– 9660 (2021)

    2021

  36. [36]

    Oquab, M. et al. DINOv2: Learning Robust Visual Features without Supervision. Preprint at http://arxiv.org/abs/2304.07193 (2023)

    Pith/arXiv arXiv 2023

  37. [37]

    Sabottke, C. F. & Spieler, B. M. The Effect of Image Resolution on Deep Learning in Radiography. Radiology: Artificial Intelligence 2, e190015 (2020)

    2020

  38. [38]

    Haque, M. I. U. et al. Effect of image resolution on automated classification of chest X-rays. J Med Imaging (Bellingham) 10, 044503 (2023)

    2023

  39. [39]

    & Hutter, F

    Loshchilov, I. & Hutter, F. Decoupled Weight Decay Regularization. in Proceedings of Proceedings of Seventh International Conference on Learning Representations (ICLR) 2019 (New Orleans, LA, USA, 2019)

    2019

  40. [40]

    R., Mijani, A

    Rezaei-Dastjerdehei, M. R., Mijani, A. & Fatemizadeh, E. Addressing Imbalance in Multi- Label Classification Using Weighted Cross Entropy Loss Function. in 2020 27th National and 5th International Iranian Conference on Biomedical Engineering (ICBME) 333–338 (IEEE, Tehran, Iran, 2020). doi:10.1109/ICBME51989.2020.9319440

    arXiv 2020

  41. [41]

    Rényi Differential Privacy

    Mironov, I. Rényi Differential Privacy. in 2017 IEEE 30th Computer Security Foundations Symposium (CSF) 263–275 (IEEE, Santa Barbara, CA, 2017). doi:10.1109/CSF.2017.11

    doi:10.1109/csf.2017.11 2017

  42. [42]

    Defining an Optimal Cut-Point Value in ROC Analysis: An Alternative Approach

    Unal, I. Defining an Optimal Cut-Point Value in ROC Analysis: An Alternative Approach. Comput Math Methods Med 2017, 3762651 (2017)

    2017

  43. [43]

    & Srebro, N

    Hardt, M., Price, E. & Srebro, N. Equality of opportunity in supervised learning. Advances in neural information processing systems 29, (2016)

    2016

  44. [44]

    & Verwer, S

    Calders, T. & Verwer, S. Three naive Bayes approaches for discrimination-free classification. Data Min Knowl Disc 21, 277–292 (2010)

    2010

  45. [45]

    & Zemel, R

    Dwork, C., Hardt, M., Pitassi, T., Reingold, O. & Zemel, R. Fairness through awareness. in Proceedings of the 3rd Innovations in Theoretical Computer Science Conference 214–226 (ACM, Cambridge Massachusetts, 2012). doi:10.1145/2090236.2090255

    arXiv 2012

  46. [46]

    & Pauly, M

    Konietschke, F. & Pauly, M. Bootstrapping and permuting paired t-test type statistics. Stat Comput 24, 283–296 (2014)

    2014

  47. [47]

    Tayebi Arasteh, S. et al. Enhancing domain generalization in the AI-based analysis of chest radiographs with federated learning. Sci Rep 13, 22576 (2023)

    2023

  48. [48]

    Tayebi Arasteh, S. et al. RadioRAG: Online Retrieval–Augmented Generation for Radiology Question Answering. Radiology: Artificial Intelligence 7, e240476 (2025). 35 Supplementary information Supplementary Table 1: Utility of differentially private ConvNeXt -Tiny models across datasets, privacy levels, and initialization strategies . Model utility is evalu...

    2025