Abstract

Precise nuclei segmentation is crucial in histopathology but remains challenging due to variable tissue types, staining protocols, and imaging conditions. Traditional deep neural networks often perform inconsistently when presented with data distributions not seen during training. Existing approaches typically process multi-scale features sequentially but lack mechanisms to explicitly enforce robustness against distribution shifts. To address this, we propose a novel framework that integrates hierarchical feature learning with associative memory to enhance model adaptability. First, we extract multi-layer embeddings from an image encoder, then process them in reverse layer order (from deepest to shallowest) using Liquid Neural Network (LNN). This design allows the model to capture global features initially and then refine them with increasingly localized information. The image encoding and the LNN encoding is then concatenated in hidden space and passed through Hopfield layer that stabilizes and stores relevant patterns. This effectively enhances domain-invariant representations by filtering out spurious correlations. Our OOD experiments on nuclei segmentation benchmark datasets show that our approach achieves average improvement of 16.35% over baseline models. Our code will be released at https://github.com/CVPR-KIT/OOD-Nuclei-Segmentation-via-LNNs-with-MHN.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3815_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/CVPR-KIT/OOD-Nuclei-Segmentation-via-LNNs-with-MHN

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SwaBis_OutofDistribution_MICCAI2025,
        author = { Swain, Bishal Ranjan and Cheoi, Kyung Joo and Ko, Jaepil},
        title = { { Out-of-Distribution Nuclei Segmentation in Histology Imaging via Liquid Neural Networks with Modern Hopfield Layer } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15966},
        month = {September},
        page = {402 -- 412}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper presents a nuclei segmentation model designed to address the out-of-distribution (OOD) domain gap. To mitigate this issue, the authors introduce a memory module that helps align OOD features more closely with in-domain features.

  • 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. The paper aims to address the domain gap in nuclei segmentation, a task that plays a critical role in histopathology diagnosis. Given this context, the study is of notable relevance and significance.
    2. The motivations for incorporating feature fusion and a memory module are clearly articulated and well-justified within the scope of the proposed solution.
  • 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. While the proposed method demonstrates promising performance, it primarily integrates several well-established techniques, such as LNN and attention-based prototype learning. As a result, technical novelty in this paper appears limited.
    2. The use of LNN for feature fusion raises questions. It would be helpful to understand the specific advantages of using LNN over more conventional approaches such as MLP- or CNN-based multi-scale feature fusion. Are there any qualitative or quantitative comparisons to support this design choice?
    3. I have concerns regarding the approach of representing OOD features using in-domain features. While this strategy may help align OOD and in-domain samples, it also risks introducing bias that could obscure meaningful, informative patterns unique to the query data.
    4. In the experimental setup, the model uses LoRA-SAM as the backbone. However, it is unclear whether this backbone is kept frozen or fine-tuned during the optimization of the LNN and prototype components. Clarification on this point would strengthen the methodology section.
    5. Since the paper aims to address the domain gap problem, it is recommended to include comparisons with existing methods in unsupervised domain adaptation and domain generalization to better demonstrate the effectiveness of the proposed approach.
  • 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.

  • 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

    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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    My primary concern with this paper lies in its limited technical innovation and the relatively insufficient experimental evaluation.

  • 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.

    I appreciate the authors’ responses and the effort to address my comments. While some concerns have been clarified, I still find the motivation for using Logic Neural Networks (LNNs), particularly in comparison to conventional MLP/CNN-based fusion methods, insufficiently justified. Additionally, the experimental evaluation lacks completeness. For these reasons, I am inclined to recommend rejection at this stage, in the hope that a more thoroughly revised version can be submitted in the future.



Review #2

  • Please describe the contribution of the paper

    The authors present a novel framework for out-of-distribution (OOD) nuclei segmentation that integrates Liquid Neural Networks (LNNs) with the associative memory capabilities of modern Hopfield networks. Their approach processes hierarchical image features through time-aware liquid dynamics. Specifically, features are extracted from the deepest to the shallowest layers of the network and passed through an LNN. These representations are then concatenated with the encoder output and subsequently processed by a modern Hopfield network before the decoding stage. This architecture aims to enhance the model’s ability to capture complex temporal and contextual dependencies in challenging segmentation scenarios and thus handle OOD samples better.

  • 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 authors demonstrate that their proposed method greatly outperforms other SOTA models in handling OOD datasets.

    • The approach yields consistently greatly improved performance gains across multiple datasets.

    • The framework presents a novel and compelling strategy to address distribution shifts.

    • Ablation studies effectively illustrate how the two proposed components complement each other, leading to performance improvements.

  • 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 experimental scope is quite limited, the method is trained only on a single dataset. This raises concerns about the generalizability of the proposed approach. (The evaluation on 3 OOD datasets is sufficient. )

    • But it would have been particularly insightful to evaluate their performance across multiple datasets (or simply cross validate).

    • Baselines specifically designed to handle distribution shifts in nuclei segmentation are missing. Including such comparisons would have provided a more rigorous benchmark.

    • The paper only evaluates performance on OOD data. For completeness, it is essential to also include results on in-distribution test sets. While the motivation for focusing on OOD is understandable, omitting in-distribution evaluation weakens the overall analysis.

    • The paper is difficult to follow in places. Key motivations and core methodological explanations are sometimes overshadowed by lengthy discussions of less important details.

    • In Figure 3, it’s unclear why LoRA-SAM, the method with the closest results to the proposed approach, is not shown. Including it would provide a more meaningful comparison.

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

  • 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

    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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    While the consistently strong performance of the proposed method is promising, the limited and insufficient evaluation raises concerns about the generalizability of the results. Without broader testing across diverse datasets and conditions, it’s difficult to assess how well the approach would perform in real-world or varied scenarios.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    While this paper shows some weaknesses in terms of overall experiments and writing, the ood results are very promising and its a rather novel interesting direction/method. It should raise some interesting discussions.



Review #3

  • Please describe the contribution of the paper

    The paper introduces an architecture for OOD nuclei segmentation that combines Liquid Neural Networks for adaptive feature encoding and MHN for associative memory stabilization. The approach shows improved robustness against domain shifts across histology datasets and achieves good performance gains over both conventional and transformer-based 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 integration of LNN and MHN provides an elegant mechanism to improve feature robustness and domain invariance, especially under OOD conditions. The extensive experiments performed allow to convincingly assess the method.

  • 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 computational cost introduced by the LNN is not sufficiently quantified, which is important for practical deployment.

  • 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

    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?

    The paper addresses a relevant problem with a technically sound solution, combining neuromorphic dynamics and memory-based feature stabilization. Its experimental performance is compelling.

  • 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 all reviewers for taking their valuable time to review our paper and providing constructive feedback. We especially appreciate Reviewer 2 for recommending acceptance of our submission. Below we address the remaining concerns in detail.

  1. Technical novelty & contribution — R1 Our primary innovation lies in a novel strategy to tackle the challenging OOD problem in nuclei segmentation. We propose the use of LNNs for their inherent flexibility and dynamic feature processing capabilities, which are theoretically well-suited for adapting to unseen data domains. However, a known characteristic of LNNs can be their training instability, potentially hindering their ability to consistently learn robust and key discriminative features. Hence to stabilize the LNN, we integrate MHN that serves to specifically stabilize the LNN’s learning process and refine its feature representations through associative memory. This synergistic combination of LNNs and MHN is a key factor behind the observed performance gains.

  2. LNN for multi-scale fusion — R1 LNNs maintain time-varying internal dynamics that adapt to changing inputs even after training. This is particularly important for processing multi-scale features sequentially. Moreover, LNNs maintain a hidden state that evolves as each scale is processed, capturing inter-scale dependencies more effectively. Also, our CfC implementation provides all these benefits with minimal parameters compared to stacked MLP approaches.

  3. Risks in representing OOD using in-domain features — R1 We agree that forcing OOD features to match in-domain prototypes could potentially obscure unique patterns. However, the MHN performs soft association with learned patterns through attention mechanisms. The LNN component processes each image’s features individually, preserving unique characteristics before the MHN. The significant performance improvements across different OOD datasets suggest that our approach successfully captures essential nuclei characteristics while adapting to domain variations. We will clarify in the final paper.

  4. Unclear whether LoRA-SAM backbone is frozen or fine-tuned — R1 We apologize for this oversight. To clarify, during training, the SAM backbone was kept frozen while only the LoRA parameters were optimized. Our LNN and MHN components were trained along with the LoRA parameters. We will add this important detail to the revised manuscript.

  5. Lack of comparison with domain generalization methods — R1, R3 We recognize that explicit domain generalization baselines were missing from the submission. We have evaluated our method against HoVer-Next and Med-OoD models, where our method performed better in comparison. We also agree with the reviewers on the inclusion of in-distribution model results and we would like to include both in the final version of paper. Since we are not allowed to add new experiments, we would request the ACs to allow us to include these results in Table 1 as they will improve and substantiate our findings better.

  6. Limited experimental scope — R3 We selected PanNuke as our training dataset because it offers diversity and broad coverage for different tissue types. We then evaluated on three distinct OOD test datasets with different characteristics (CryoNuSeg, CoNIC, CoNSeP). Our approach represents a realistic scenario where a model is trained once on a diverse dataset and deployed across different clinical settings. We however agree with the reviewer that our experimentation need broader testing results to ascertain the generalization claims however we think its a promising start and could lead to impactful discussion at the conference.

  7. Other Concerns — R3

    • There was an oversight on our part where we misspelt the image labels. We actually included images for U-Net3+ and Unet3+ w LNN+MHN; LoRA-SAM and LoRA-SAM w/ LNN+MHN.
    • We appreciate reviewers feedback on the clarity of our methodology section and will improve it in the final version of the paper.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The paper presents a novel and promising approach by leveraging Logic Neural Networks (LNNs) for OOD detection. Despite some minor weaknesses in writing and experimental depth, the results are strong and the method offers a fresh perspective that could inspire further research.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top