Abstract

Dense self-supervised learning (SSL) is a powerful paradigm for learning without annotations the local descriptors required to solve dense medical imaging tasks. We present Pix2Rep-v2, a framework for SSL of pixel- and voxel-level representations suitable for few-shot downstream applications. Pix2Rep-v2 addresses the main challenges of dense SSL by leveraging a redundancy reduction objective at the pixel-level with a principle of equivariance of dense representations, that scales efficiently to 3D or wide field-of-view applications. We evaluate our method on four datasets, across multiple tasks, multiple modalities and anatomical structures using multiple backbones in 2D and 3D, and under various data regimes. As an alternative to linear probing or full fine-tuning on the downstream task, we also propose an in-context variant, without downstream training, based on a dense prototype approach. Pix2Rep-v2 shows substantially higher data-efficiency in few-shot scenarios compared to fully supervised baselines, and is competitive with the state-of-the-art e.g., +9.3 Dice points in one-shot segmentation on the M&Ms-2 dataset. Our code and pre-trained models are publicly available at https://github.com/BioMedTP/pix2rep-v2.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0971_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/BioMedTP/pix2rep-v2

Link to the Dataset(s)

ACDC dataset: https://humanheart-project.creatis.insa-lyon.fr/database/#collection/637218c173e9f0047faa00fb M&Ms dataset: https://www.ub.edu/mnms/ M&Ms2 dataset: https://www.ub.edu/mnms-2/ AMOS dataset: https://zenodo.org/records/7262581

BibTex

@InProceedings{SifSof_Pix2Repv2_MICCAI2026,
        author = { Sifaoui, Sofiane AND Angelini, Elsa AND Toupin, Solenn AND Pezel, Théo AND Le Folgoc, Loïc},
        title = { { Pix2Rep-v2: Data-Efficient Representation Learning for Dense Medical Imaging Applications } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a self-supervised learning framework for medical image analysis based on redundancy reduction of learned features, unlike its predecessor Pix2Rep that uses a contrastive loss. The method computes a cross-correlation matrix between pixel-level representations extracted from two augmented views of the same image, encouraging the learned features to be invariant to these transformations while minimising redundancy across feature dimensions.

    A particularly elegant aspect of the architecture is the interleaving of the feature extraction network and the spatial augmentation pipeline before the projection heads. This allows the model to learn representations that are inherently robust to geometric variability at the feature level. Additionally, the authors introduce a random patch size strategy during training, where patches of varying sizes are extracted and subsequently resized to a fixed resolution before being fed to the network. This design encourages the model to generalise across different input scales.

    The framework is evaluated on both 2D and 3D medical datasets across multiple regimes: linear probing, fine-tuning, and low-data scenarios (one-shot, few-shot, and many-shot). The motivation for developing self-supervised methods tailored to medical imaging is well-grounded, given the chronic scarcity of annotated data in clinical settings.

  • 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 motivation for self-supervised pretraining in medical imaging is compelling and clearly presented, given the high cost of expert annotation.

    • The interleaving of the feature extraction network and spatial augmentation before the projection heads enables the use of stronger spatial augmentations, allowing the model to capture geometric invariance at the feature level.

    • The random patch size strategy during training is a practical and elegant solution for generalising across input scales, while keeping inference simple with a fixed patch size.

    • The experimental evaluation is commendably broad, covering both 2D cardiac and 3D abdominal datasets, which demonstrates the generalisability of the approach across different imaging modalities.

    • The low-data regime evaluation — spanning one-shot, few-shot, and many-shot settings — is thorough and particularly relevant for real-world clinical scenarios where labeled data is scarce.

    • They tested the method for standard fine-tuning and linear probing and prototype-based in-context segmentation that uses features learned from a set of prototypes (or atlases) for which the ground-truth is available.

    • The reported results are strong and suggest the framework is practically effective.

  • 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.
    • My major concern is that this work is incremental with respect the proposed approach in the Pix2Rep paper, where they already introduce Pix2Rep-v2 without giving too much details. Noteworthy, in the original paper, the authors already claimed that Pix2Rep-v2 performed slightly better than the original (contrastive) Pix2Rep. Probably, Pix2Rep-v2 have been improved since then (e. g. , using a projection head, multi-scale sampling and interleaving feature representation with spatial augmentations). But this discussion is missing from the current submission

    • The paper would benefit from a dedicated ablation or comparison directly against previous Pix2Rep-v2 from the original paper to quantify the impact of the proposed adaptations in isolation.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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
    • FPN acronym appears undefined
  • 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 work makes a genuine and substantive contribution to self-supervised learning for medical imaging. The transition from a contrastive (Pix2Rep) to a redundancy-reduction objective (Pix2Rep-v2) at the pixel level is not a superficial change — it reflects a fundamentally different approach to representation learning that carries well-understood theoretical and practical advantages. Combined with the thoughtful architectural choices (interleaved spatial augmentation, scale-agnostic training and in-context segmentation using prototypes) and a thorough experimental evaluation across modalities and data regimes, this paper presents a cohesive and impactful body of work.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    I believe it represents a good contribution to the community, providing comprehensive evaluation and strong results.



Review #2

  • Please describe the contribution of the paper

    This paper presents Pix2Rep-v2, a dense self-supervised learning (SSL) framework for learning pixel- and voxel-level representations for few-shot medical image segmentation. The method extends prior dense contrastive SSL approaches by replacing the contrastive loss with a redundancy reduction objective (Barlow Twins) applied at the pixel level, combined with an equivariance-based formulation to spatial transformations. Additional contributions include a multiscale patch sampling strategy, an implicit 3D U-Net backbone for scalable 3D applications, and an in-context segmentation variant based on dense nearest-neighbor prototypes that requires no task-specific fine-tuning. The framework is evaluated on four datasets (ACDC, M&Ms, M&Ms-2, AMOS) across cardiac MRI segmentation, video propagation, and abdominal CT segmentation tasks.

  • 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.
    • Comprehensive evaluation: The paper evaluates across four datasets, two imaging modalities (cine MRI, CT), three downstream tasks (segmentation, in-context segmentation, video propagation), two backbones (U-Net, Swin-UNETR), and multiple data regimes (1%–100%). This breadth of evaluation is a clear strength and provides confidence in the generality of the approach.
    • Strong few-shot performance: The reported gains are substantial in low-data regimes. The ×25 data-efficiency improvement over training from scratch is a strong result.
  • 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 proposed method has limited technical novelty. The method applies Barlow Twins (Zbontar et al., ICML 2021) at the pixel level with equivariance-based consistency regularization under spatial transformations. Each individual component (redundancy reduction loss, equivariance to spatial augmentations, and multiscale cropping) is well-established. Especially when compared with Pix2Rep, the proposed method can be seen as replacing the contrastive loss with Barlow Twins loss, which makes the proposed method less technically novel.
    • In figure 4 and figure 2, Pix2Rep-v2 uses both SwinUNETR and Unet as backbone network. It would be valuable if Pix2Rep also used a similar baseline network for fair comparison. Additionally, in many experiment setting, Unet performs better than the SwinUNETR. However, many prior works show that SwinUNETR has stronger performance than Unet. It would be valuable to discuss these results.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    I lean towards weak accept. The paper is well-written with a thorough experimental evaluation, and the practical results are convincing. A limitation of this paper is that the core technical contribution is incremental: it primarily applies Barlow Twins at the pixel level with spatial equivariance. Especially compared with Pix2Rep, which uses contrastive loss, the proposed technical novelty is relatively limited.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    I incline to accept the paper. The authors clarified most of my concerns. The paper is well-written, and the practical results are convincing. Despite the technical limitation, the problem the paper aims to solve is valuable. And overall, I incline to accept the paper.



Review #3

  • Please describe the contribution of the paper

    The authors present a self-supervised learning method called Pix2Rep-v2, which leverages redundancy reduction objective at the pixel-level with a principle of equivariance of dense representations, that scales efficiently to 3D or wide field-of-view applications. The method is evaluated on four datasets, across multiple tasks, multiple modalities and anatomical structures using multiple backbones in 2D and 3D.

  • 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 is well structured and written
    • I really like the in-context approach the authors proposed and it is a nice idea, however the results are not satisfactory to actually being able to use it (compared to the fine-tuned one)
    • Results look good for zero-shot but still are far away from performances that can be used. Using 10% already gives significantly better results (Figure 5). –. So it gets Clea training on 100% gets better but only marginally which is good in terms of efficiency and computational resources.
  • 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.
    • Cardiac MRI segmentation results (figure 2) is very cluttered and therefore difficult to extract information from –> bar plots might not be the ideal way to visualize
    • Having Pix2Rep-v2 outperforming SAM3 seems a bit obvious, as SAM never saw any data from medical domains. On the contrary; against MedSAM2, it got outperformed which is the real comparison one should take.
    • I am missing some ablation results on the memory and footprint of the model during training / inference. Do we have a gain there as well ?
    • Figure 5: The proposed method seems suspicious. It consist of Pix2Rep-v2 + Implicit U-Net and compares against Implicit U-Net –> the joint method/ensemble works better (which is not a surprise)
    • For Dice scores: Is the Dice score averaged per organ/label or is the dice calculated on the joined segmentation mask ? It should be per label and then averaged
    • Method is still a bit difficult to understand and the Figure 1 is also not that informative in terms of the architecture and how the method works opposed to the compared baselines.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    The preliminary rating is based on my review with particular emphasis on the identified weaknesses outlined in the sections above. These weaknesses highlight several critical gaps in the current proposal.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    The authors answered properly my questions, however given their previous work (Pix2Rep) as pointed out by other reviewers, I am not really convinced of the “new” v2 method given its current state and the proposed results.



Author Feedback

We thank reviewers for their constructive feedback and positive comments on the paper’s technical solidity (R1,AC), clarity (All), strong convincing results (R1,R2,AC) and substantive practical value (R1,AC). We address concerns below.

[R1,AC] Novelty vs. MIDL24 Pix2Rep paper The MIDL24 paper is positioned as a proof-of-concept with a single use case on fixed-size 2D slices, on a center-cropped version of ACDC, using GT labels to get the ROI (cf. Pix2Rep public codebase). This is inapplicable for SSL pretraining on real-world use cases with large fields of view. Instead, the current paper proposes:

  • A multiscale patch sampling strategy, that makes the method widely applicable, with fully label-free pretraining, and representations more scale-agnostic
  • A novel in-context segmentation method that outperforms SOTA fine-tuned SSL models in 1-shot
  • A new strategy to scale up to 3D applications with a thoughtful Implicit U-Net backbone
  • A video propagation method

[R2,AC] Novelty vs. SOTA outside of Pix2Rep’s line of work Although Barlow Twins (BT), equivariance, multiscale sampling exist in the broader AI & medical imaging literature, the Pix2Rep paradigm exploits these concepts in a markedly distinctive way absent from the dense SSL SOTA:

  • It is the first to implement mathematical equivariance, via “interleaving feature representation with spatial augmentation” (R1), instead of aligning 2 separate crops with limited overlap. This addresses a real technical barrier, allowing for stronger spatial augmentations
  • v2 is the first to use BT at the pixel-level instead of a contrastive objective. As per R1, it is a “fundamentally different approach” with “well-understood theoretical and practical advantages” (no hard negative sampling issue, trains on full -not sparse- set of pixels, lower memory cost)
  • It enables unprecedented results in few-shot segmentation. Pixel-level representations are intrinsically good enough to allow for an in-context approach downstream, which outperforms SOTA fine-tuned SSL methods
  • Besides, we are not aware of other in-context methods scaling up to ~40 support scans, with further performance gains (cf. Fig2, M&Ms-2, 10% column).

[R2,AC] U-Net vs. SwinUNETR backbones for baselines: Preliminary experiments with baselines (incl. Pix2Rep) showed consistently lower performance for SwinUNETR. For conciseness we only included the best configuration (w/ U-Net). [1] also hints to strong performance of U-Net-based models still today vs. transformer-based. We included SwinUNETR for Pix2Rep-v2 to demonstrate flexibility w.r.t. backbone. [1] Isensee et al. nnU-Net revisited, MICCAI 2024

[R3] “really like the in-context approach” but performance insufficient vs. fine-tuned Pix2Rep-v2 In-context segmentation is not intended to be a drop-in replacement for fine-tuning w.r.t. performance. In-context inference aims to:

  • Avoid supervised fine-tuning for users without expertise or resources
  • Potentially segment any organ with the same model, without prior bias towards an organ or annotation protocol
  • Assess the intrinsic quality of SSL representations

[R3] “results look good” for 1-shot, but not good enough to “be used” As per Fig2, we show a large improvement over the SOTA. We leave it to future work to make 1-shot segmentation clinical-grade.

[R3,AC] Clarifying evaluation settings

  • “Pix2Rep-v2+Implicit U-Net” vs. “Implicit U-Net”: we evaluate specifically the impact of Pix2Rep-v2 SSL pretraining vs. a fully-supervised baseline trained from scratch with the same Implicit U-Net backbone. We will clarify Fig5 accordingly (legend, caption & color code)
  • Performance vs. MedSAM2 on video propagation: MedSAM2 is trained on all of ACDC, M&Ms, M&Ms-2 GT labels, unlike Pix2Rep-v2, which operates in a zero-shot regime. We include it to provide an upper bound on performance. We surpass non data-contaminated models: Pix2Rep and SAM3’s foundational mask propagation module
  • Dice is indeed computed per label then averaged




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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.

    This paper introduces Pix2Rep-v2, a self-supervised learning framework for representation learning based on redundancy reduction instead of contrastive learning, with learned spatial equivariance and robustness to multiple resolutions.

    The paper is well-motivated, clearly written, and extensively evaluated, with particularly strong performance in low-data regimes. The experimental section is consistently highlighted as a major strength, including: broad validation across anatomies (with 2D/3D datasets), modalities, tasks, and inclusion of multiple evaluation protocols (fine-tuning, linear probing, in-context segmentation). Moreover, the results are generally strong and explained by insightful ablations. Finally, code is released to support reproducibility.

    The main weakness of the paper is its questionable novelty, which is mainly based on the association of already existing components (Barlow Twins objective, equivariance, multi-scale sampling). More troubling is the fact that a version of this work supposedly already appeared in the original Pix2Rep work, thus further worsening technical novelty. The authors need to provide more explanations about the positioning of the current submission with respect to these previous works, and include them in the evaluation.

    Further concerns to be addressed include commenting on network backbones across baselines, discussing zero-shot results, clarifying some evaluation settings, and improving figures and visualisation.

    Overall, this is a technically solid and empirically strong paper, with convincing results and clear practical relevance. However, the concerns about incremental novelty and insufficient positioning relative to prior Pix2Rep variants need clarifications.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The authors have satisfyingly responded to novelty concerns, which in my opinion is sufficient for acceptance at MICCAI. Moreover, the authors have clarified the positioning of this work with respect to their previous method Pix2Rep and a preliminary version of Pix2Rep-v2.The authors have also clarified raised points about experimental setup and backbone issues.



    These clarifications combined with the initial strengths of the paper (strong and thorough evaluation, impactful message) make it a good contribution to MICCAI, and I thus recommend acceptance. 



    However, I very strongly encourage the authors to improve the quality of result presentation, especially of Figures 2, 4 and 5, which can be very crowded (thus difficult to read), and use bar plots instead of box/violin plots. I also note the total lack of statistical analysis, which is a major weakness of this paper and should be added upon final acceptance. I will also recommend oral, conditionally on the addition of statistical analysis.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Authors carefully addressed all the reviewers’ concerns and clarified the novelty of the presented work with respect to the prior work. It would be helpful to clearly state what the contributions are and the improvements over the prior work in the paper. Reviewers agree regarding the practical utility and the strong results shown on multiple datasets, albeit all point to the incremental novelty, that should be clarified.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After reading the reviews and the rebuttal, I recommend Accept. The paper is not a clean break from Pix2Rep, and the novelty concern is real. The main change is to replace the contrastive objective with a pixel-level redundancy-reduction objective, together with the spatial equivariance design, multi-scale sampling, and the 3D extension. These pieces are not individually new. However, the submission makes a complete and useful case for this design in dense medical representation learning, rather than only presenting it as a small loss-function swap.

    The evidence is the main reason for acceptance. The method is tested across cardiac MRI and abdominal CT, in 2D and 3D settings, with linear probing, fine-tuning, low-data training, video propagation, and prototype-based in-context segmentation. The results are not uniformly dominant against every medical foundation model baseline, but they are consistent enough to show that the learned dense features are useful under limited annotation. This is the setting where the paper’s contribution is strongest.

    The rebuttal also resolves several reporting concerns, including the Dice computation and the comparison setup. Some presentation issues remain: the relationship to the earlier Pix2Rep-v2 mention should be stated more explicitly, and the figures could make the experimental message easier to read. I would not recommend the paper as a strong novelty paper. Still, as a method paper with broad evidence and a practical representation-learning recipe for dense medical tasks, it is above the acceptance threshold.



back to top