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

Abstract

Implicit neural representations(INRs) have achieved remarkable successes in learning expressive yet compact signal representations. However, they are not naturally amenable to predictive tasks such as segmentation, where they must learn semantic structures over a distribution of signals. In this study, we introduce MetaSeg, a meta-learning framework to train INRs for medical image segmentation. MetaSeg uses an underlying INR that simultaneously predicts per pixel intensity values and class labels. It then uses a meta-learning procedure to find optimal initial parameters for this INR over a training dataset of images and segmentation maps, such that the INR can simply be fine-tuned to fit pixels of an unseen test image, and automatically decode its class labels. We evaluated MetaSeg on 2D and 3D Brain MRI segmentation tasks and report Dice scores comparable to commonly used U-Net models, but with 90% fewer parameters. MetaSeg offers a fresh, scalable alternative to traditional resource-heavy architectures such as U-Nets and vision transformers for medical image segmentation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3113_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/kushalvyas/MetaSeg

Link to the Dataset(s)

OASIS-MRI dataset: https://github.com/adalca/medical-datasets/blob/master/neurite-oasis.md

BibTex

@InProceedings{VyaKus_Fit_MICCAI2025,
        author = { Vyas, Kushal and Veeraraghavan, Ashok and Balakrishnan, Guha},
        title = { { Fit Pixels, Get Labels: Meta-Learned Implicit Networks for Image Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15962},
        month = {September},
        page = {197 -- 207}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This work introduces MetaSeg, a meta-learning framework to train INRs for medical image segmentation. This method is based on optimizing initial weights for an INR in the meta-learning regime on the set of training samples consisting of raw images and segmentation masks. The INR has two output heads, one that yields pixel intensity and the other the pixel class. Upon inference the meta-initialized INR is fitted to the image at hand and once the fit has good enough quality, an optimization stops and segmentation mask is retrieved through the second output head. Authors show how this approach can drastically limit number of model parameters while obtaining on-par results with UNet.

  • 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.

    1) Very interesting and novel approach to INR-based segmentation. It it is inspired by NISF [20], but instead of freezing INR weights and optimizing latent code, it meta-learns INR initialization and fine-tunes it on new sample. By that it achieves better performance as observed in Table 1. 2) Study of relationship of reconstruction quality (Fig. 2) with respect to segmentation quality is a very informative addition. Indeed this behaviour of INR overfitting has been commonly raised in INR and Weight Space learning works. 3) Vastly decreasing number of parameters while maintaining model’s quality is a good motivation for future practical usage - especially for embedded systems.

  • 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.

    1) Structures in OASIS-MRI dataset are quite homogenous across the samples - they lack anatomical variability. Authors already mention in discussion that their method is quite sensitive to spatial misalignment. It would make the work much stronger if the authors would use some other dataset that has a foreground class that can appear in different places of the image e.g. tumor segmentation. In principle network like this would need to adapt much more to a new sample than it is for OASIS-MRI. That would prove that the fine-tuning mechanism is robust enough to account for anatomical variability. 2) I believe that reproducing the method would be very challenging based on provided implementation details, while the authors do not mention making their code public upon acceptance. 3) Authors should mention how much time does it take to fine-tune to the specific sample. Decreasing number of parameters is important for practical usage but the inference time as well.

  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    1) I wonder how does one choose how many iteratation of fine-tuning to run to obtain the best-quality segmentation? Would you use fixed number of iterations? Or maybe compute the best PSNR threshold on training data and then use it for early stopping the optimization of test INR? 2) Would you consider whether yielding multiple segmentation masks during fine-tuning could be a form of uncertainty estimation?

  • 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?

    I find the proposed MetaSeg model to be a very interesting and novel addition to INR-based segmentation architectures. The authors have studies its feasibility for 2D & 3D problems, as well as problems with different amounts of segmentation classes. The work could be made stronger by including additional datasets that are less spatially homogenous e.g. tumor segmentation; and studying the effects of data augmentation in more detail. However, in my opinion the body of this work is solid and these additional experiments can be considered for future work.

  • 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 authors propose a strategy to obtain segmentations using implicit neural representations (INRs). In contrast to prior work, a joint function for the instensities and the segmentation labels is learned. At test time, the IRNs parameters can be optimized to match the input image. The same parameters can then be used to predict the corresponding segmentation. The work is evaluated on 2D and 3D datasets and shows comparable performance to established NN-based segmentation techniques.

  • 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 was a pleasure to read. It is very clearly written and easy to follow.
    • This is a very innovative novel approach to segmentation that has large potential to inspire the community.
    • Using meta-learning to find an optimal initialisation is a very elegant and effective idea. It is also empirically demonstrated that it is crucial for obtaining high Dice scores.
  • 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 authors do not mention that they are willing to share the code upon acceptance.
  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    Beyond my comments above, I only have some general discussion points (no need to answer those in a potential rebuttal):

    • Seeing that the algorithm is sensitive to the initialisation strategy (as shown in Table 2) do the authors see potential in using more advanced meta learning strategies to further improve performance?
    • A potential drawback (as pointed out by the authors) is that the method appears to be not robust to spatial alignment. Would this limit the methods applicability in applications beyond anatomical segmentation with more variable objects, for example, tumor or MS lesion segmentation?
    • Minor nitpick: I believe Fig. 4 is not discussed in the results (only later in the discussion). When reading the experiments the reader feels like they have missed something.
  • 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.

    (6) Strong Accept — must be accepted due to excellence

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

    Innovative and novel method with excellent presentation and evaluation on multiple datasets.

  • 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

    Fit Pixels, Get Labels introduces a novel and highly exciting perspective for medical image segmentation - the insight that using meta-learning for simultaneously fitting an image and its corresponding segmentation labels in the form of neural fields / implicit neural representations can lead to a generalized segmentation function. Based on MAML, the paper demonstrates that meta-learned (!) INRs learn a semantic structure in their layers in contrast to vanilla INRs, and that this structure may be leveraged to segment unseen images by just overfitting the meta-initalized INR to the new scan (just intensities, without the new for a segmentation mask). The proposed method achieves similar segmentation scores as current SOTA CNN-based methods, and outperforms the current implicit segmentation baseline NISF by higher margins.

  • 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.

    This paper exhibits many stengths:

    • Firstly, the authors introduce a learning strategy based on MAML, but expand it with a segmentation head, that is simple and easy to adapt to many medical imaging problems, and thus is an extremely exciting perspective for the whole field, especially since it is computationally cheap and easy to adapt to other segmentation problems.
    • While traditional CNNs have been optimized for many years (starting with the original U-Net paper and ultimately leading to e.g. nnUNET and ConvNeXt) , INRs for segmentation have been explored much less frequently. It is thus even more impressive that MetaSeg performs on par or challenges SOTA (c.f. 2D task) for these tasks while using less model parameters.
    • The paper is extremely well written, and flows well. It’s a fun and interesting read.
    • The authors set their paper perfectly into context of the INR and segmentation literature, conduct appropriate experiments for a proof-of-concept.
    • I really enjoyed reading the Methods and Discussion part, especially the ablation with shifting and rotating images, since INRs are typically used to center aligned images (e.g. MNI space) for many tasks.
  • 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 only possible weakness I could see is that the authors demonstrate this on the OASIS dataset only. However, the dataset is reasonably sized for training a competitive U-Net baseline (i.e. it is a fair comparison), and the fine-grained OASIS segmentation labels constitute a challenging segmentation use case that INRs can deal with quite well (that#s actually where they shine). It would be very interesting to see how well INRs can deal with sparse segmentation maps (e.g. single class lesions in MS, BraTS tumors), where the segmentation map is very sparse and does not occupy regions of the whole brain. For instance, it would be interesting to see if INRs can attain similar scores, if surrounding regions of a certain label are masked out. However, I feel this is future work, and likely out of the scope of this paper.

  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    I would encourage the authors to open source their implementation to encourage other researchers to leverage MetaSeg for their segmentation problems, and to let other people work with their proposed method to advance it even further.

  • 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.

    (6) Strong Accept — must be accepted due to excellence

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

    This paper introduces as simple, yet effective strategy for joint image reconstruction and segmentation with implicit neural representations. This simple, yet powerful approach is able to obtain segmentations with less resources, and is ultimately relevant to the INR and segmentation research fields. It’s exciting!

  • 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



Author Feedback

We appreciate and are grateful for the extremely positive and enthusiastic responses by the reviewers of our work, and the fact that all three reviewers felt that the paper should be accepted. We address the comments and questions raised by the reviewers below.

Regarding reproducibility and open-sourcing code(R1,R2,R3): We will make our code open-sourced on GitHub, with all the needed scripts to recreate our experiments and figures. We apologize for not mentioning this in the initial manuscript.

Regarding homogeneity of OASIS-MRI(R1,R2,R3): We agree with reviewers that the OASIS-MRI dataset has some limitations regarding homogeneity. Nevertheless, MetaSeg’s segmentation performance and compactness on OASIS-MRI are impressive, especially given its comparable metrics to U-Nets, which have been developed over many years (as reviewer R3 points out). MetaSeg is an exciting starting point to explore more complex segmentation tasks with INRs, such as tumor and lesion segmentation, which exhibit higher variability.

Handling spatial misalignment in images(R1,R2,R3): INRs fit a signal in a layerwise coarse-to-fine fashion [1], conditioned on its input coordinates. Hence, learned INR features are highly attuned to the underlying signal morphology and tightly coupled to the input coordinate ordering. As a result, handling spatial misalignment is non-trivial and an exciting future direction to build upon using MetaSeg.

Figure 4(R1): We appreciate the reviewer’s suggestion to reference Figure 4 early on. We will incorporate this change in our final submission.

References: [1] Vyas et.al., Learning Transferable Features for Implicit Neural Representations. NeurIPS 2024



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A



back to top