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

Abstract

While recent FMs have shown the power of large-scale pre-training to CT and MRI analysis, there remains significant room to optimize how these models learn from complex radiological volumes. Building upon the Curia framework, this work introduces Curia-2, which significantly improves the original pre-training strategy and representation quality to better capture the specificities of radiological data. The proposed methodology enables scaling the architecture up to billion-parameter Vision Transformers, marking a first for multi-modal CT and MRI FMs. Furthermore, we formalize the evaluation of these models by extending and restructuring CuriaBench into two distinct tracks: a 2D track tailored for slice-based vision models and a 3D track for volumetric benchmarking. Our results demonstrate that Curia-2 outperforms all FMs on vision-focused tasks and fairs competitively to vision-language models on clinically complex tasks such as finding detection. To foster further research, model weights are made publicly available at https://huggingface.co/raidium/curia-2.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6594_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://huggingface.co/raidium/curia-2

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SapAnt_Curia2_MICCAI2026,
        author = { Saporta, Antoine AND Callard, Baptiste AND Dancette, Corentin AND Khlaut, Julien AND Corbière, Charles AND Butsanets, Leo AND Prat, Amaury AND Manceron, Pierre},
        title = { { Curia-2: Scaling Self-Supervised Learning for Radiology Foundation Models } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces Curia-2, a refined self-supervised pretraining recipe for radiology foundation models. Key modifications to DINOv2: content-aware cropping, anatomically-guided masking, SigReg regularisation instead of KoLeo, lower pretraining resolution (256) with high-res fine-tuning (512). Scales to ViT-g (1.3B params), a first for multimodal CT/MRI FMs. Evaluated on CuriaBench (19 tasks, 2D and 3D tracks). Claims SOTA on vision tasks and is competitive with vision-language models on findings detection.

  • 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 incremental ablation is well-structured, and each modification is individually justified with a clear rationale for why DINOv2’s defaults are ill-suited for radiology (background-heavy crops, trivial masking, KoLeo forcing apart clinically similar representations)

    • Content-aware cropping seems to solve a real training collapse problem at lower resolutions and is the single most impactful contribution (+34.6 points), making the paper’s core insight clear and reproducible (if indeed the case - see weaknesses)

    • Comprehensive benchmark across 19 tasks spanning 7 clinical categories, both CT and MRI, with 2D/3D tracks and results averaged over 5 runs, strong evaluation methodology

    • Practically important: Lower-resolution pretraining (256 vs 512) with staged fine-tuning achieves comparable or better results at lower compute cost, and weights will be released

  • 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 34.6-point drop when moving from 512 to 256 resolution needs more explanation. Is the configuration identical to Curia B except for resolution? Token count, batch size, number of crops, and training iterations all change when resolution halves — the paper does not clarify what was held constant and what was adjusted to match compute budgets. Without this, it is unclear whether the collapse is a resolution problem or a configuration mismatch.
    • SigReg replaces KoLeo but contributes 0.0 points in the ablation. If it neither helps nor hurts on any measured task, the justification for introducing an additional optimisation objective is weak. The theoretical argument about preserving clinically similar representations is sensible but unsupported by the results presented.
    • The difference between Curia B and Curia-2 B is 0.2 points on average (84.8 vs 85.1). Without confidence intervals, p-values, or per-task variance across the 5 runs, this difference is not interpretable. Adding dataset sizes for each evaluation task and 95% CIs would help determine whether this is meaningful or noise.
    • The 3D track comparisons are dominated by VLMs (Merlin, Pillar-0, CT-CLIP) and MedImageInsight, none of which are natively 3D vision-only models. Comparing against 3D vision-only baselines such as SPECTRE, CT-FM, or 3DINO would make the 3D track results far more informative.
    • Scaling from ViT-L to ViT-g provides no consistent benefit. On the 2D track, both achieve an aggregate of 88.5%. On the 3D track, ViT-L (88.6%) actually outperforms ViT-g (87.7%), with a notable gap on M4.At 3x the compute cost, the scaling story is unconvincing. A more valuable contribution would be an analysis of optimal model size selection, where diminishing returns begin and why, rather than billion-parameter scaling.
  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    The paper presents Curia-2, a refinement of the Curia pretraining recipe for radiology foundation models, with modifications to DINOv2’s cropping, masking, regularisation, and resolution strategy. While the individual modifications are sensible engineering decisions and the ablation study is methodically structured, the core contribution is incremental. Curia-2 B improves over Curia B by only 0.2 points on average (with matching configs), and the proposed modifications (content-aware cropping, anatomically-guided masking, SigReg regularisation) are adaptations of existing techniques rather than novel contributions.

    The paper’s headline claim of scaling to billion-parameter ViTs is not supported by the results. ViT-g matches ViT-L on the 2D track (88.5% each) and underperforms it on the 3D track (87.7% vs 88.6%) at three times the compute cost. This undermines the scaling narrative, and the paper does not offer an analysis of why scaling stalls or what an optimal model size might be.

    The 3D track evaluation lacks comparison against established 3D vision-only models (SPECTRE, CT-FM, 3DINO), instead comparing primarily against VLMs, which makes the results difficult to contextualise.

    In its current form, this work reads as a solid engineering report on adapting DINOv2 to radiology at scale, but the contribution does not meet the threshold for a main conference paper and would be better suited for a workshop venue.

  • 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

    • This is the first publicly reported billion-parameter self-supervised model that jointly handles CT and MRI.

    • This paper introduces a set of domain-specific modifications to the DINOv2 recipe.

    • A benchmark that contains 19 tasks spread across different disease categories.

    • This paper includes extensive empirical validation, encompassing a full ablation study, few-shot learning curves, cross-modal similarity maps, scaling experiments, and comparisons against nine baselines

  • 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 research demonstrates a high degree of completeness. This paper includes extensive empirical validation, encompassing a full ablation study, few-shot learning curves, cross-modal similarity maps, scaling experiments, and comparisons against nine baselines

  • 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 paper thoughtfully adapts existing self-supervised learning ideas to the radiology domain, while the contributions are primarily engineering refinements rather than theoretical novelties.

    • The 3-D evaluation track is potentially unfair, as slice-based models are assessed by simply averaging slice features, which discards the spatial context that genuine volumetric models are designed to exploit.

    • The term “images” throughout the paper refers to individual 2D slices rather than full 3D scans, which may cause confusion for readers

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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 makes a significant contribution by introducing the first publicly reported billion-parameter self-supervised model for joint CT and MRI analysis

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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

    The primary contribution of this work is Curia-2, a large-scale self-supervised radiology foundation model scaling to 1.3B parameters, built on the Curia framework and DINOv2 with domain-specific adaptations for CT and MRI data. The authors introduce several targeted improvements, including content-aware cropping to remove non-informative regions, anatomically-guided masking to focus learning on relevant structures, and clinical-prior regularization to better preserve meaningful feature similarity. They also restructure evaluation through CuriaBench by separating 2D slice-based and 3D volumetric benchmarks, helping standardize assessment across settings. While there are open questions around 3D adaptation, diminishing returns at scale, and semantic representation compared to other paradigms, the scale of pretraining, demonstrated stability and efficiency, and commitment to releasing model weights make this work compelling, and I am leaning toward accept.

  • 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.
    • Massive pretraining with a billion parameter Vit-g. Also, large scale dataset consisting of over 400 million images.
    • Better training stability through their approaches compared to Dino-V2.* Data efficiency: Few-shot learning demonstrates strong results with Dino-style pretraining.
    • The authors have mentioned that they will make the model weights public, which will be a great contribution to the community.
  • 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.
    • How is curia adapted for 3D volumetric tasks if it’s trained on only 2D images? It seems to perform well but what does Curia miss compared to native 3D architectures.
    • Diminishing returns of scales. While the model scale, the difference between Curia2-L and Curia2-G are marginal and it may require rethinking exactly how pretraining needs to be done at large model scales.
    • Semantic information: Although Figure 3 showed some benefit of semantic understanding, it would be great to understand what is lacks compared to Merlin or Pillar-0 which are CLIP-style models.
  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    While there are open questions around 3D adaptation, diminishing returns at scale, and semantic representation compared to other paradigms, the scale of pretraining, demonstrated stability, high-performance, and efficiency, and commitment to releasing model weights make this work compelling, and I am leaning toward accept.

  • 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 #4

  • Please describe the contribution of the paper

    The paper proposes Curia-2, a scalable self-supervised radiology foundation model for CT and MRI. Its main contribution is to adapt the DINOv2/Curia pre-training framework to radiological imaging through content-aware cropping, anatomically guided masking, clinical-prior regularization, and low-resolution pre-training followed by high-resolution fine-tuning. These changes improve training stability and representation quality, enabling the model to scale to large Vision Transformers and achieve strong performance across 2D and 3D radiology benchmarks.

  • 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 has several strengths. First, it addresses an important and timely problem: scaling self-supervised foundation models for CT and MRI radiology data. Second, it benefits from a very large-scale pre-training dataset, consisting of hundreds of millions of CT and MRI images, which makes the scaling study more convincing. Third, the method includes several interesting training-strategy innovations, such as content-aware cropping, anatomically guided masking, and replacing KoLeo regularization with SigReg. These strategies are well motivated by the characteristics of radiological images. Overall, the paper provides clear ablation studies and a relatively broad evaluation across both 2D and 3D radiology benchmarks.

  • 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 paper has several weaknesses. First, the technical novelty is moderate, as the main framework builds on DINOv2/iBOT and Curia, with most improvements being training-strategy adaptations rather than fundamentally new architectures or learning objectives. Second, although the pre-training dataset is very large, its institutional diversity is not clearly demonstrated, which limits evidence for broad cross-site generalization. Third, clinical feasibility is not fully validated, as the evaluation mainly relies on downstream benchmark performance rather than real-world clinical workflow testing. Finally, the performance gain from Curia-2 L to Curia-2 g is limited, suggesting that the benefit of billion-parameter scaling in radiology remains insufficiently established.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    The paper has several weaknesses. First, the technical novelty is moderate, as the main framework builds on DINOv2/iBOT and Curia, with most improvements being training-strategy adaptations rather than fundamentally new architectures or learning objectives. Second, although the pre-training dataset is very large, its institutional diversity is not clearly demonstrated, which limits evidence for broad cross-site generalization. Third, clinical feasibility is not fully validated, as the evaluation mainly relies on downstream benchmark performance rather than real-world clinical workflow testing. Finally, the performance gain from Curia-2 L to Curia-2 g is limited, suggesting that the benefit of billion-parameter scaling in radiology remains insufficiently established.

  • 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?

    My recommendation is based on the paper’s clear practical value but moderate technical novelty. The work addresses an important problem, uses a very large CT/MRI pre-training dataset, and provides well-motivated radiology-specific training strategies with clear ablation and broad benchmark evaluation. However, the core framework largely builds on DINOv2/iBOT and Curia, and the main contributions are mainly training-recipe adaptations rather than fundamentally new methods. In addition, cross-site clinical generalization and real-world clinical feasibility are not fully demonstrated. Overall, I view the paper as a solid and useful scaling study for radiology foundation models, but not a highly novel methodological contribution.

  • 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



Author Feedback

N/A



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 paper presents Curia-2, a self-supervised pretraining recipe that adapts DINOv2 to the radiology domain and scales to a 1.3B-parameter ViT, evaluated across a 19-task benchmark spanning both CT and MRI. The contribution is best understood as a domain adaptation of training recipes rather than a fundamentally new framework, and certain limitations remain. The benefit of billion-parameter scaling is not consistently demonstrated, and the empirical justification for some design choices, including SigReg and the resolution transition, is partial. That said, the breadth of evaluation, the empirical rigor of the 5-run averaging protocol, and the value of the planned model weight release to the community speak in favor of acceptance.



back to top