Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Foundation models (FMs) pre-trained via self-supervised learning (SSL) on large, diverse datasets are widely expected to offer superior generalization—and, by extension, reduced bias—compared to classically supervised (CS) models. Yet whether this holds in practice, and whether the choice of pretraining data matters, remains largely unresolved. We present a controlled comparison between SSL-pretrained FMs and a CS baseline fine-tuned on identical data across two prenatal ultrasound tasks: spontaneous preterm birth (sPTB) prediction and fetal scan weight estimation. Using visual radar-based bias profiles spanning a comprehensive set of clinical and acquisition subgroups, we show that FMs reproduce the bias profile of their CS counterpart—best- and worst-performing subgroups largely coincide, and mean performance gaps are comparable in magnitude. Pretraining data strategy does influence bias: oversampling cervical images during pretraining outperformed scaling to 13 million images for sPTB in both global performance and subgroup fairness, while naïve oversampling of rare devices degraded overall performance. These results challenge the assumption that large-scale diverse pretraining uniformly reduces bias, and suggest that data alignment with the downstream task matters more than scale.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4565_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{FouJor_Are_MICCAI2026,
        author = { Fournel, Joris AND Ambsdorf, Jakob AND Jürgensen, Benjamin Jønch AND Boland, Christopher AND Elgebaly, Aya AND Mikolaj, Kamil AND Pegios, Paraskevas AND Sejer, Emilie Pi Fogtmann AND Tolsgaard, Martin AND Christensen, Anders Nymark AND Nielsen, Mads AND Feragen, Aasa},
        title = { { Are Foundational Models Less Biased Than Specialized Models? An Ultrasound Study } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper tests whether SSL-pretrained foundation models are less biased than classically supervised models, using two prenatal ultrasound tasks (preterm birth prediction and fetal weight estimation) with a controlled comparison across various factors and underlying subgroups. The answer is suprising as FMs reproduce nearly identical bias profiles, with the same best- and worst-performing subgroups and comparable performance gaps. Another crucial finding is that aligning pretraining data with the downstream task (e. g. , oversampling cervical scans for sPTB) matters more than scaling to 13M images. Scale over quality in pretraining matters more.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Directly tests a widely held but underexamined assumption that SSL pretraining reduces bias using a controlled experimental design using identical fine-tuning data across FM and CS models.
    • Well designed bias quantification framework: four complementary metrics (absolute gap, relative gap, improved/worsened subgroups, DSC bias similarity index) plus radar plot visualization across clinical and acquisition factors
    • Valuable finding that naively using pretrained FMs reproduce the bias profile of CS models (17 improved, 17 worsened)
    • The pretraining data strategy comparison is practically valuable showing that targeted cervical oversampling (2M images) outperforms generic scaling (13M images) has direct implications for how institutions should invest compute resources
    • Scale of the study is substantial for this kind of analysis: 13M pretraining images, 7,862 sPTB samples, 433K scan weight images
    • Very clear and intuitive explanation of the study aims, hypothesis and conclusions.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The FM uses a ViT-B architecture while the CS baselines are CNNs. Although the paper’s central claim is about pretraining strategy (SSL vs supervised), the architecture difference is an uncontrolled confounder. A ViT-B trained from scratch on the data would add additional insight as well - is pre-training even necessary?
    • The bias similarity index (DSC on radar plot areas) is not sufficiently justified for a novel metric. It is not axis-order invariant: rearranging factors on the radar changes polygon areas and therefore the DSC, even with identical underlying data. The paper does not address how this metric behaves as the number of factors increases, how subgroup count within each factor influences the score, or what constitutes a meaningful DSC threshold. New metrics introduced as collateral contributions require more rigorous validation and technical exposition.
    • The subgroup-level analysis would benefit from more granular reporting. Extended versions of Figures 3 and 4 covering all factors (not just ethnicity, device, and birth year) would help readers assess where bias persists and where it shifts. Even an exploratory supplementary figure showing per-subgroup performance across all models would strengthen the paper.
    • The choice of 2M images as the base pretraining size could be justified better. Is it plausible that a smaller subset could yield similar performance? Especially given that DINOv2 is known to exhibit performance drift on long training horizons with low-variance images like ultrasound. DINOv3’s Gram-based teacher-student loss was designed to address this, and the authors use DINOv2, raising the question of whether the 2M-to-13M scaling comparison is confounded by training dynamics rather than reflecting a true data volume effect. A comment on this would be helpful.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    LLMs were used solely for polishing review as per acceptable reviewer policy.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses a timely and practically important question — whether SSL-pretrained foundation models reduce bias compared to classically supervised models — with a well-controlled experimental design. The controlled comparison uses identical fine-tuning data and comprehensive subgroup stratification (55 and 38 subgroups across clinical and acquisition factors). The finding that pretraining data alignment with the downstream task outperforms naive scaling to 13M images is actionable and directly relevant to institutions deciding how to allocate compute resources for foundation model training. The scale of the study (13M pretraining images, two distinct tasks, four pretraining variants, bootstrap significance testing) is substantial and the comprehensive bias quantification (sans the bias similarity index) represents a meaningful effort to move beyond single-metric fairness evaluation.

    The architecture confound (ViT-B vs CNN) and the lack of robustness analysis for the bias similarity index are real limitations, but they do not undermine the paper’s central finding: that FMs reproduce the bias structure of CS models fine-tuned on the same data, with the same best- and worst-performing subgroups. This conclusion is supported by multiple independent metrics beyond the DSC, including the symmetric 17/17 improved-worsened split and the convergent relative gaps. The practical implications that naïve scaling is not a substitute for thoughtful data curation, and that SSL pretraining should not be assumed to confer fairness benefits are important messages for the community.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper experimentally explores if fine-tuned foundation models are less biased/unfair models trained from scratch on the same fine-tuning data. The analyses are carried out on two US downstream tasks: spontaneous preterm birth prediction and scan weight prediction. The results show that foundation models (mostly) share the same biases as standard model, which (could) indicate that the fine-tuning step is where the bias is induced, but that is not further explored.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Interesting topic that definitely deserves a thorough analysis
    • US downstream tasks is an interesting shift from more traditionally used scenarios
    • Visualizations in the paper are quite nice
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The title’s paper suggests a broader and very thorough analysis of the bias problem in foundation models, which the paper does not provide. It is fully focused on the two US downstream tasks and it remains unclear how generalizable the results are.
    • The written text is not really extensive in terms of details and a lot is missing to really assess how things are done, which I think is problematic for an evaluation study. Crucially, it remains totally unclear how the fine-tuning is done (full network?, only a head?, …). I think this is an extremely important point to really assess where the biases are introduced. It also remains unclear how the standard baselines differ architecturally from the foundation models. I would assume that this also affects how and where biases occur, but this is neither studies nor discussed.
    • The study uses quite large datasets for fine-tuning, which makes sense for the standard models, but does not that kind of defeat the purpose of a foundation model? In other words: In reality, would not I want to use the foundation primarily in fewer shot scenarios, where only 100 samples or less are available?
    • Many subgroup attributes are analyzed, but I am missing any thoughts about intersectionality and how that affects the outcomes.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I think that the study raises a critical question and is of interest to our audience, but that it is not developed enough in its current form to meet the standards for evaluations (without methodological contributions) at MICCAI. Please also see my other comments above.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper provides a controlled comparison between self-supervised foundation models and classically supervised models, showing that foundation models largely reproduce similar bias profiles. It introduces a comprehensive subgroup-based bias evaluation framework combining visual and quantitative analysis. The study further demonstrates that pretraining data composition, particularly task-aligned curation, has a greater impact than dataset scale. Overall, it challenges the assumption that large-scale SSL pretraining inherently reduces bias and offers practical insights for data-centric model design.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper presents a well-controlled experimental design that enables a fair comparison between foundation and supervised models. It introduces a comprehensive and nuanced subgroup-based bias evaluation framework, going beyond standard aggregate metrics. The work provides a novel and important empirical finding that challenges the assumption that SSL pretraining reduces bias. Additionally, it highlights a valuable data-centric insight: task-aligned pretraining is more effective than naive scaling. Finally, the study is conducted at meaningful scale on clinically relevant tasks, strengthening its practical impact.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The study is limited to a single architecture (ViT-B) and two ultrasound tasks, which restricts generalizability to larger foundation models and other imaging modalities.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper addresses an important and timely question regarding bias in foundation models and provides a well-controlled and carefully designed experimental study. The strongest aspects are the rigorous comparison setup, the comprehensive subgroup-level bias evaluation, and the practical insight that task-aligned pretraining can outperform naive scaling. These findings challenge common assumptions and are highly relevant to the medical imaging community.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank the reviewers and AC for their time and feedback.

Here are our answers and clarifications for the reviewers; the final version will be modified accordingly.

== Title of the paper ==

In order to avoid any confusion about the scope of our study, we decided to modify the title (“Are Foundational Models Less Biased Than Specialized Models?”) with the following addition “An Ultrasound Study”.

== Implementation details and comparison of architectures==

We will clarify in the manuscript that the foundation model (FM) was finetuned in all its parts: encoder, decoder, up to the final classification layer. Regarding the exact architecture of the specialized model, it is entirely available in the referenced work of Pegios et al [1]. We will further emphasize that we reimplement exactly their architecture (SA-Sononet).

== Are foundation models only for few-shots scenarios? ==

Deep learning models have some prominent domain shift issues such as scanner bias, where foundation models can hopefully help. FM have already been shown to improve robustness and mitigate biases on larger finetuning datasets [2,3]. And our results confirmed that these models can improve performance over specialized models on two different tasks.

Therefore, FM, beyond their special fitness for few-shots scenarios, can also be valuable in other scenarios.

== Effect of intersectionality on the outcome ==

We believe the intersectionality between factors cannot be expected to fundamentally alter the outcome of this paper. This outcome was, given two models, to answer this question: is the first model considerably less biased than the other? Given enough factors, two very similar bias profiles already tell us that the two bias behaviors correspond in their major directions, which was what we wanted to know.

But the intersectional analysis could indeed add to that in revealing some additional and more subtle differences.

Hence, we did acknowledge the absence of intersectional analysis as a limitation in the discussion, but we will make this part even more clear in the final version.

== Results without pretraining for ViT-B ==

We recognize this comparison of architecture can be of interest. However, as our purpose here was to compare foundation models to specialized models, we did not consider the ViT-B architecture without the pretraining as it would not be considered a foundation model anymore.

== Bias similarity index ==

Our goal with this index to associate some familiar metric value of overlapping (Dice Similarity Index) next to the two visuals of the bias profiles. (We didn’t intend to introduce a new metric with specific properties.) We will clarify this in the methods sections.

== Granular reporting of subgroup analysis ==

We agree with the reviewer that an exhaustive reporting of all subgroups and biases would be valuable in appendix. Unfortunately, supplementary figures are not allowed, and this reporting would have to replace another important part of the paper. Hence, we chose to select three important subgroups for granular reporting (Ethnicity, Device, Birth Year).

[1] Pegios, P. et al. (2023). Leveraging Shape and Spatial Information for Spontaneous Preterm Birth Prediction. ASMUS 2023.https://doi.org/10.1007/978-3-031-44521-7_6 [2] Goyal, Priya, et al. “Vision models are more robust and fair when pretrained on uncurated images without supervision.” arXiv preprint arXiv:2202.08360 (2022). [3] Sofia Yfantidou, et al. “Using Self-supervised Learning Can Improve Model Fairness.” In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ‘24). https://doi.org/10.1145/3637528.3671991



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviews are conflicting; After reviewing the paper and reviewers’ comment, I am inclined to sugget the authors to present it on the conference. however, it would be better to address the major concerns raised by the negative reviewers when preparing for the camera ready version



back to top