Abstract

Developing generalizable predictive models for brain disease detection critically depends on the availability of a large amount of fMRI data. To enable large-scale analysis, it is often necessary to aggregate data from multiple acquisition sites/domains. However, differences in scanners and acquisition protocols across sites lead to unfavorable heterogeneity in data distribution, which hinders reliable biomarker identification and undermines the generalizability of predictive models. In this paper, we propose URLCF, a novel few-shot learning framework that learns a set of universal representations by distilling knowledge from multiple separately trained models for cross-domain autism detection. Specifically, we first train multiple domain-specific networks to serve as teacher models for guiding the universal model. Subsequently, domain-specific data are simultaneously fed into both the universal and teacher models. By randomly masking the output features of the universal model and employing domain-specific generators to reconstruct these masked features, we align the learned representations of the universal model with those of the teachers. Concurrently, a domain discriminator is introduced to perform domain alignment, jointly facilitating the learning of universal representations. Finally, few-shot fine-tuning is conducted on the unseen domain to enable rapid adaptation to the unseen distribution. Experimental results on the ABIDE dataset demonstrate that URLCF achieves superior performance in autism detection under both zero-shot and five-shot settings, confirming its effectiveness in generalizing to unseen domains with limited data. The code is available at https://github.com/dwd111/URLCF.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2881_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/dwd111/URLCF

Link to the Dataset(s)

ABIDE dataset: http://fcon_1000.projects.nitrc.org/indi/abide/

BibTex

@InProceedings{MaoZhe_URLCF_MICCAI2026,
        author = { Mao, Zhenlin AND Sun, Qiyu AND Wan, Zhuo AND Huang, Jiashuang AND Wang, Mingliang},
        title = { { URLCF: Universal Representation Learning for Cross-Domain Few-Shot Autism Spectrum Disorder Detection } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a two-stage framework (URLCF) for cross-domain few-shot ASD detection from rs-fMRI functional connectivity. The method first distills knowledge from multiple site-specific teacher models into a universal student representation through masked feature reconstruction, logit distillation, and domain discrimination, and then performs lightweight few-shot adaptation on unseen sites. The experiments are conducted on the ABIDE dataset under a leave-one-site-out setting with both zero-shot and five-shot evaluation.  

    I view this as a clinically motivated engineering paper with moderate novelty. The problem is important and realistic, but the technical contribution is largely a combination of existing ingredients, including multi-teacher distillation, masking-based reconstruction, adversarial/domain alignment, and lightweight target adaptation, rather than a fundamentally new representation-learning method. The manuscript somewhat over-packages this as “universal representation learning. ”

  • 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 addresses an important practical problem in multi-site neuroimaging, and the evaluation protocol is well aligned with the motivation. The leave-one-site-out setting is appropriate for studying unseen-site generalization, and the additional five-shot setting is also clinically relevant. The overall framework is coherent and the reported empirical gains over the included baselines are encouraging. The paper is also generally well organized and includes ablation analysis and qualitative visualization.

  • 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.

    Global limitation: The paper does not truly establish “universal” representations in a strong scientific sense; it mainly shows improved performance on the ABIDE benchmark. The biological and interpretability claims are also overstated. The presented FC visualization, t-tests, and t-SNE plots are post-hoc and correlational, and do not justify claims about reliable biomarkers or mechanistic interpretability. 

    Method-induced limitation: The main concern is limited novelty. Once unpacked, the method is a relatively standard combination of teacher-student distillation, masked reconstruction, domain discrimination, and a lightweight adapter. In addition, the framework is fairly heavy, requiring multiple teacher models, generators, and a discriminator, yet the paper provides little analysis of computational cost, hyperparameter sensitivity, or training stability. The ablation study is useful but still not enough to show why this specific combination is necessary. 

    Dataset-induced limitation: The empirical validation is restricted to ABIDE only. Although leave-one-site-out is stronger than random splitting, it is still internal validation on a single benchmark. This is not sufficient to support strong claims about robust cross-domain generalization. In addition, the five-shot setting is inherently fragile in medical imaging due to substantial subject-level and site-level heterogeneity. 

    Another concern is reproducibility. Although the paper provides an anonymized code link, the anonymous repository does not yet appear sufficiently complete for full verification. These missing details should be explicitly addressed in the rebuttal. In particular, the authors should clarify what exactly is included, whether the full training and evaluation pipeline can be reproduced, and whether all implementation details, scripts, and dependencies needed for leave-one-site-out and few-shot experiments are available.

  • 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 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

    This is a technically competent paper on an important problem, but the claims should be better calibrated. In particular, the authors should tone down the language around “universal representation” and “reliable biomarkers,” better position the work as an integration of existing techniques, and clarify the completeness of the anonymous repository in the rebuttal.

  • 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?

    Overall, I find this paper borderline but positive. It addresses a meaningful practical problem, uses a suitable evaluation protocol, and reports consistent gains over the listed baselines on ABIDE. However, the novelty is moderate, the biological and interpretability claims are overstated, and the evidence for broader generalizability remains limited because the validation is confined to a single benchmark. In addition, although the submission provides an anonymized code link, the current repository content does not yet appear sufficient for full reproducibility, and this should be clarified and completed in the rebuttal.

  • 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.

    Thank you for your response. I understand that some raised issues are difficult to sufficiently address during the rebuttal stage. As no additional supportive evidence can be provided, I will keep my original rating.



Review #2

  • Please describe the contribution of the paper

    In this work, a novel few-shot learning framework is introduced for ASD detection. The framework follows the student-teacher design to obtain a universal student model from datasets across multiple domains. In addition, a masked generation and a discriminator are integrated to promote universal learning. The proposed model, URLCF, demonstrates superior performance.

  • 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 URLCF effectively addresses the multi-site deployment bottleneck by utilizing a Teacher-Student distillation strategy, which extracts and transfers knowledge from multiple domain-specific models into a single, unified student network to mitigate fMRI data heterogeneity.

    2.The authors introduce a highly robust alignment mechanism that combines masked generative reconstruction with a conditional domain discriminator, actively forcing the network to learn generalized, universal representations rather than site-specific artifacts.

    3.The framework’s efficacy is rigorously validated through extensive comparative experiments on the ABIDE dataset against diverse baselines, while a detailed ablation study successfully demonstrates the critical contribution of each individual module to the model’s overall zero-shot and five-shot performance.

  • 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.How is the performance score ( s_i ) computed? If it relies on training accuracy, how do you prevent overfitted teacher models—especially those trained on small ABIDE sites—from being given overly large weights?

    2.Were the baseline models (especially MAML, Prototypical Networks, and the domain-invariant approaches) built using the same 3-layer MLP architecture as your feature extractor?

    3.Minor: a missing citation on page 6 (ComBat[?]).

  • 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?

    Even though the methodology’s novelty is limited, the proposed method offers a practical solution to few-shot autism spectrum disorder detection, highlighting promising performance together with an ablation study. Therefore, I suggest a weak accept.

  • 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.

    All my concerns have been addressed, and I believe the added revisions will greatly improve the clarity of the manuscript. I therefore recommend accepting this paper.



Review #3

  • Please describe the contribution of the paper

    1.The authors introduce a few-shot learning framework that distills knowledge from multiple domain-specific teacher models into a single universal student model to alleviate multi-site fMRI heterogeneity. 2.It uses masked generative alignment combined with a conditional domain discriminator to align feature distributions. 3.Extensive experiments on the ABIDE dataset demonstrate model effectiveness.

  • 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 manuscript is well-structured, logically organized, and clearly written. 2.It provides biological interpretability for evaluation.

  • 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.A shallow MLP is inadequate for capturing the complex topological features of FC networks, lagging behind current SOTA methods that favor GNNs or Transformers. 2.Does masking 1D latent features differ fundamentally from standard dropout? Furthermore, what is the specific advantage of masking in the latent space rather than directly on the raw FC matrices? 3.Relying exclusively on the ABIDE dataset is insufficient to prove generalizability. The authors are strongly encouraged to validate URLCF on additional multi-site datasets, such as REST-meta-MDD. 4.The selected baselines and reference list are notably outdated (mostly pre-2022). Excluding recent SOTA methods [1-2] significantly weakens the persuasiveness of the results. [1] Yang Y, Ye C, Su G, et al. Brainmass: Advancing brain network analysis for diagnosis with large-scale self-supervised learning[J]. IEEE transactions on medical imaging, 2024, 43(11): 4004-4016.[2] Yu S, Jin S, Li M, et al. Long-range brain graph transformer[J]. Advances in Neural Information Processing Systems, 2024, 37: 24472-24495.

  • 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.

    (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 methodological innovation is insufficient, and the comparative experiments rely on outdated baselines. Furthermore, evaluating on a single dataset significantly weakens the overall persuasiveness of the results.

  • 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.

    The authors have adequately addressed my primary concerns regarding the backbone architecture, masking strategy, and model generalizability.



Author Feedback

We thank all reviewers for their comments and enthusiasm for our work. R#1 Q1: Global limitation A1: In this work, the term “universal representation” is intended to describe representations that are transferable across heterogeneous imaging domains, rather than biologically universal representations in a strict neuroscientific sense. We will revise the manuscript to avoid overly broad interpretations. Also, we agree that the current visualization analyses are post-hoc and correlational. These analyses were used to provide qualitative insight into the learned feature structure and cross-domain consistency. We will revise the manuscript to present these findings as preliminary observations that may motivate future neuroscientific validation. Q2: Method limitation A2: Although the individual modules are inspired by existing techniques, the main contribution of this work lies in the principled integration of these modules into a unified framework for cross-domain few-shot fMRI representation learning under severe site heterogeneity. Importantly, ablation results indicate that these components play complementary roles within the framework, as removing any individual module leads to degraded unseen-domain generalization performance. Regarding computational costs, we would like to mention that multi-teacher training is performed only once in an offline stage. During inference and target domain adaptation, only the student model and a lightweight adapter are used, making the framework computationally efficient for few-shot fine-tuning. Q3: Generalization A3: The current study primarily focuses on ASD-related analysis and evaluation. We will further validate our method on additional multi-site datasets (e.g., MDD and ADHD) in future work. Q4: Reproducibility A4: As suggested, we have updated the anonymized repository to clearly present the full training and testing pipeline, ensuring the reproducibility of our implementation. R#2 Q1: Calculating of performance score A1: For each non-held-out site, we split the data into training and validation sets, and compute the validation accuracy as the performance score using a single fully connected layer. To prevent overfitted teacher models, we apply a temperature scaling parameter τ in Eq. (1), which smooths the weight distribution, ensuring that no single teacher model dominates the distillation process. Q2: Architecture of baselines A2: For a fair comparison, all few-shot learning and domain-invariant methods were implemented using the same three-layer MLP architecture as our feature extractor. Q3: Missing citation A3: We apologize for the typo error, and will revise it in the revised version. R#3 Q1: Deficiencies of MLP A1: We agree that shallow MLPs are insufficient to fully capture the complex topological structure of FCs. In this study, we use shallow MLPs to avoid over-parameterization and ensure stable optimization, especially under limited neuroimaging data. This lightweight backbone also allows us to better evaluate the effectiveness of the proposed framework itself, independent of sophisticated network architectures. Q2: Masking A2: Masking is a feature corruption strategy that hides parts of the input to promote robust representation learning, whereas dropout is a regularization technique that randomly drops neural activations during training to reduce co-adaptation and prevent overfitting. Compared with directly masking raw FC matrices, masking semantically enriched high-level latent representations enables more robust feature learning and improves downstream clinical prediction performance. Q3: Generalization A3: A similar concern was also raised by R#1.A detailed response is provided in our reply to Q#3 of R#1.Q4: Comparison methods A4: We have already compared against 3 categories and 9 baseline methods. In future work, we will further expand the comparisons by incorporating additional SOTA methods to provide a more comprehensive evaluation.




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.

    Based on the three reviews, the paper addresses an important practical problem and uses a reasonable evaluation protocol. However, all reviewers raise substantial concerns. R1 notes limited novelty and the claim of “universal” representations. R2 questions the weighting mechanism and baseline comparisons. R3 delivers a strong critique: the shallow MLP backbone is inadequate compared to current GNN/Transformer methods, masking in latent space lacks clear advantage, the validation is restricted to a single dataset (ABIDE), and the baselines are notably outdated. Given these issues, the authors must reply the concerns regarding architectural limitations, dataset generalizability, baseline recency, and reproducibility.

  • 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 rebuttal does a good job in clarifying the “universal” representations in the paper. The reviewers and I agree that the paper merits acceptance to MICCAI in its current form.



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.

    After carefully reviewing the authors’ rebuttal, I have revised my initial assessment to Accept. The authors directly addressed key concerns including clarifying the intended meaning of “universal representation,” explaining the principled integration of modules with ablation evidence, justifying MLP choice for stable optimization under limited data, distinguishing latent masking from dropout, and committing to expanded validation on additional multi-site datasets.



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.

    This paper is recommended for acceptance. The reviewers converged positively after rebuttal, and the authors adequately addressed the main concerns regarding the backbone choice, masking strategy, few-shot setting, and generalizability.



back to top