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

Abstract

The interaction between brain structure and genetic influences is key to understanding neuropsychiatric disorders. However, most large-scale datasets are unimodal, providing either neuroimaging or genetics data. We propose CALM, a framework that learns interpretable associations between brain ROIs and genetic pathways from \textit{completely disjoint} populations. CALM aligns the two modalities in a shared latent space via linear projections that simultaneously match the class-conditional latent distributions and ensure group separability. These projections provide interpretable pathway–ROI associations. When trained on unimodal imaging and genetics datasets, CALM generalizes to an unseen paired dataset, outperforming several state-of-the-art methods and ablation baselines. We also demonstrate stability of the learned associations against a paired baseline. Our experiments on autism spectrum disorder reveal immune and metabolic pathways linked to specific cortical regions and are consistent with established literature. Thus, CALM opens the door to leveraging large unimodal repositories for studying cross-modal interactions in brain disorders across disparate datasets.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2961_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/jueqiw/CALM

Link to the Dataset(s)

ABIDE dataset: https://fcon_1000.projects.nitrc.org/indi/abide/ SSC dataset: https://www.sfari.org/resource/simons-simplex-collection/ ACE dataset: https://nda.nih.gov/edit_collection.html?id=2021

BibTex

@InProceedings{WanJue_CALM_MICCAI2026,
        author = { Wang, Jueqi AND Jacokes, Zachary AND Van Horn, John Darrell AND Pelphrey, Kevin A. AND Schatz, Michael C. AND Venkataraman, Archana},
        title = { { CALM: Interpretable Cross-Modal Alignment for Biomarker Discovery from Unpaired Data } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper
    1. This work introduces a multimodal learning framework CALM that can be trained on entirely non-overlapping cohorts, enabling the modeling of associations between neuroimaging features and gene pathways without requiring paired multimodal data from the same subjects.
    2. A shared latent space is constructed via linear projections to align imaging and gene pathway modalities. In addition, a class-conditional distribution alignment strategy is employed to enforce consistent discriminative structures across modalities, facilitating effective cross-modal integration.
    3. The proposed projection mechanism is inherently interpretable.
    4. Even when trained solely on unimodal datasets (imaging or genetic) drawn from completely disjoint subject groups, the model generalizes well to unseen paired datasets and outperforms several existing approaches as well as ablation baselines.
    5. The empirical findings highlight immune- and metabolism-related pathways associated with specific cortical regions, providing meaningful evidence toward understanding the molecular basis of brain disorders.
  • 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. Unlike conventional approaches that rely on paired multimodal data from the same individuals, the proposed method learns cross-modal relationships from disjoint cohorts, improving its applicability in scenarios where paired data are scarce.
    2. By leveraging a shared latent space together with class-conditional distribution alignment, the model is able to establish meaningful inter-modal relationships even in the absence of explicit pairing.
    3. The use of linear projections allows for direct interpretation of region–pathway associations, which is advantageous over black-box models for biomarker discovery and mechanistic insights.
    4. The model trained on fully disjoint unimodal datasets can be transferred to unseen paired datasets, demonstrating strong generalization capability.
    5. The identified associations between immune/metabolic pathways and specific brain regions offer cross-scale (brain–gene) insights relevant to the study of brain disorders.
  • 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. The paper does not clearly specify the form of the final classifier (e. g. , whether a linear classifier or another architecture is used). This omission makes the overall pipeline less transparent and may affect both reproducibility and interpretability.
    2. The biological validation of the findings remains limited. It would be beneficial to incorporate external resources (e. g. , the Allen Brain Atlas) to further corroborate the reported brain region–gene pathway associations, thereby strengthening the reliability and biological relevance of the conclusions.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

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

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

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

    This paper addresses a central challenge, the scarcity of paired data, in multimodal neuroimaging and genetic data integrationby proposing a framework that can be trained on disjoint cohorts. This design choice is practically relevant and well-motivated. By learning a shared latent space and incorporating class-conditional distribution alignment, the model achieves cross-modal integration while preserving discriminative structure. Moreover, the use of linear projections contributes to interpretability, enabling explicit characterization of associations between brain regions and gene pathways.

    A notable strength is the model’s ability to generalize to unseen paired datasets despite being trained only on unimodal data from non-overlapping subjects. This property is both novel and potentially impactful for real-world applications. In addition, the reported associations between brain regions and immune/metabolic pathways provide biologically meaningful insights.

    While there is room for improvement in terms of clearer specification of model components and more comprehensive biological validation, the paper presents a well-motivated approach with solid methodological contributions, good interpretability, and promising cross-modal generalization performance.

  • 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.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper proposes CALM, an interpretable framework for learning cross-modal associations between neuroimaging and genetics from completely unpaired and disjoint datasets.

    The key contribution is a class-conditional alignment strategy that combines modality-specific encoders with linear projection matrices, enabling the model to align the two modalities in a shared latent space while preserving a direct correspondence between brain ROIs and genetic pathways. This design allows the extraction of an interpretable ROI-pathway association matrix rather than relying on black-box latent fusion.

    Experimentally, the method is trained on independent unimodal datasets and evaluated on an external paired dataset, where it achieves the best classification performance among the compared baselines. The paper further validates the learned associations by showing strong agreement with a paired-data baseline and by linking the identified pathway-ROI patterns to established ASD-related findings in the literature.

  • 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.Novel and practically relevant problem setting.
    The paper addresses the important and underexplored problem of learning imaging-genetics associations from completely unpaired datasets. This is practically meaningful because many large public datasets are unimodal and disjoint, making paired multimodal analysis difficult in real-world settings. 2.Interpretable model design.
    A major strength of the work is that interpretability is built into the architecture itself. By using per-entity encoders and linear projections, the method preserves explicit correspondence between brain ROIs and genetic pathways, allowing direct extraction of an association matrix. This is more meaningful than typical latent fusion approaches where feature interactions become difficult to interpret. 3.Well-motivated objective formulation.
    The combination of class-conditional distribution alignment, supervised contrastive learning, and orthogonality regularization is conceptually coherent. The objective is designed not only to align the two modalities, but also to maintain class discrimination and avoid degenerate solutions in the projection matrices.
    1. Reasonable experimental validation.
    The paper includes comparisons against several relevant baselines as well as ablation studies that isolate the contribution of the key loss terms. In addition, the method is evaluated on an independent paired external dataset, which strengthens the empirical validation. 5.Validation of biological consistency.
    The learned association matrix shows high correlation with a paired-data baseline (r = 0.87), and the identified pathway-ROI relationships are discussed in the context of established ASD literature. This gives additional support that the learned cross-modal associations are not purely algorithmic artifacts.
  • 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.Important hyperparameter choices are under-justified.
    The method depends on several potentially sensitive hyperparameters, including the weights of the CMMD, contrastive, and orthogonality losses, as well as the contrastive temperature. Although these values are reported, the paper does not clearly explain how they were selected, nor does it provide a sensitivity analysis. This makes it difficult to assess how robust the reported performance is to these design choices. 2.The performance gain over the ablations is relatively modest.
    CALM achieves the best overall results, but the margin over the strongest ablation variants is small in Table 1.While the method is conceptually appealing, the empirical advantage of using the full objective over simpler variants is not yet entirely convincing. A stronger discussion of when each component materially helps would improve the paper. 3.Reproducibility is limited by insufficient architectural detail.
    The paper states that pretrained autoencoders and classifiers are used, but the exact architecture and pretraining protocol of these encoders are not described in sufficient detail in the main text. Since the encoder design is important for preserving interpretability and for the second-stage alignment, clearer implementation details would improve reproducibility. 4.Biological interpretation remains largely qualitative.
    The discussion of the discovered pathway-ROI associations is interesting, but the biological interpretation is mostly narrative and literature-supported rather than quantitatively validated. Additional statistical or biological validation would make the interpretability claim stronger.
  • 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.

  • 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

    Overall, I found the paper interesting and technically well motivated. The problem setting is relevant, and the effort to build interpretability directly into the alignment framework is commendable. I encourage the authors to improve the paper by clarifying the encoder architecture and pretraining details, providing more justification for hyperparameter selection, and strengthening the empirical discussion around the relatively small gains over some ablations. These additions would make the paper easier to reproduce and would further support the practical value of the proposed method.

  • 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 recommend Weak Accept. The paper addresses a meaningful and challenging problem—learning interpretable imaging-genetics associations from fully unpaired datasets—and proposes a method that is both conceptually coherent and practically relevant. The interpretability-oriented design, the external paired-data evaluation, and the comparison with a paired baseline are all positive aspects. At the same time, I do not view the paper as a clear accept because some important implementation details are under-specified, the hyperparameter choices are insufficiently justified, and the empirical gains over the strongest ablations are relatively modest. Overall, I think the paper is above the acceptance threshold due to the novelty of the problem setting and the quality of the methodological idea, but I would expect the authors to address the concerns above in the rebuttal and revision.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 rebuttal addresses several of my technical concerns in a serious and constructive way. In particular, the authors provide substantially clearer detail on the encoder/classifier architecture and pretraining protocol, and they clarify that hyperparameters were selected by grid search with 5-fold CV on the unpaired training cohorts without using the ACE dataset. The rebuttal also offers a more explicit explanation of why the association matrix can be interpreted as reflecting ROI-pathway relationships in the aligned latent space.

    However, after reading the rebuttal, I remain unconvinced on several points that are central to the paper’s overall strength. First, the empirical advantage of the full model over the strongest ablations is still relatively modest, and the rebuttal explains the intended role of each loss term more clearly than it demonstrates a strong practical necessity of the full objective. Second, the biological validation remains largely qualitative. The authors point to future validation against external resources such as SFARI and the Allen Brain Atlas, but this does not resolve the current limitation of the submitted manuscript. Third, while the baseline and ablation rationale is better explained, I still find the experimental evidence somewhat limited relative to the strength of the interpretability claims.

    Overall, I appreciate the novelty of the problem setting and the authors’ thoughtful rebuttal, but I do not think the rebuttal fully resolves the remaining concerns about the strength of empirical support and biological validation. For that reason, I would not change my final opinion to acceptance.



Review #3

  • Please describe the contribution of the paper

    The paper proposes CALM, a framework for learning interpretable cross-modal associations between brain region morphometry and genetic pathways from entirely unpaired and disjoint populations. CALM employs a class-conditional linear transport model to align group-wise distributions across modalities in a shared latent space. Validated against a paired baseline and on an unseen external dataset, it yields stable associations and outperforms state-of-the-art 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 main strengths of the paper are: (1) The paper tackles a clinically relevant and practical problem by learning cross-modal associations from completely unpaired imaging and genetics data. (2) The method is conceptually simple and well-structured, combining distribution alignment and contrastive learning in a clear and reasonable way. (3) The experimental evaluation is relatively comprehensive, including external validation on a paired dataset and analysis of association stability.

  • 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 main weaknesses of the paper are: (1) The paper has many formatting issues. For example, the reference formats are inconsistent; the citation of Fig. 3(b) is problematic; the subsection ‘Regularization (L_{orth})’ is missing a colon; the font styles around symbols S(i) and P(j) are inconsistent, etc. The authors need to thoroughly check the manuscript themselves. (2) There’s a lot of unnecessary italic styling throughout the paper. Italics should really be used only when you want to emphasize a key point. (3) There are too few baseline comparisons, only three. And the performance gain of CALM over its own ablations is very limited. Also, why is the Ablation L_{OT} using the loss from Ref [3]? What’s the reason? (4) The association matrix A = W_I^T W_G is just the product of two parameter matrices. Why does that directly quantify pairwise relationships between brain regions and genetic pathways? What’s the interpretability there? Also, why set the latent dimension d = 6? There’s no justification for this choice, and no parameter analysis for this value.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

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

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (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?

    Although the CALM tackles a clinically relevant and practical problem by learning cross-modal associations from completely unpaired imaging and genetics data, I would recommend a weak reject due to the technical concerns highlighted above. A strong rebuttal addressing these issues could be highly beneficial.

  • 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 thank the reviewers for their constructive feedback and for recognizing the important problem of learning cross-modal associations from unpaired imaging and genetic data. We address the major critiques below.

1.Construction of the Association Matrix (Rev 3): Each column of W_I (resp. W_G) defines how a brain ROI (resp. genetic pathway) contributes to each shared latent dimension. L_orth keeps these rows near-orthonormal to maximize the variance captured by these dimensions. L_cmmd and L_con align the two modalities along these shared dimensions. A large entry A_{r,p} implies that ROI “r” and pathway “p” load similarly onto the latent dimensions, indicating that they co-vary in the aligned space. This is analogous to CCA but without paired samples. Columns of W_I / W_G correspond one-to-one to ROIs / pathways, so A can be viewed as an ROI-pathway association matrix.

2.Hyperparameter Selection (Revs 2 & 3): Hyperparameters were tuned via grid search with 5-fold CV on the unpaired ABIDE and SSC datasets, covering latent dimension d, batch size, learning rate, loss weights (λ_cmmd, λ_con, λ_orth), and contrastive temperature. The ACE dataset was never seen during hyperparameter selection, and Table 1 confirms CALM’s generalizability to it. Latent dimension d had the largest impact on performance: larger values tend to overfit the modest training set. Loss weights and contrastive temperature were robust across reasonable ranges and can serve as defaults in future studies.

3.Clarity of Model Architecture (Revs 1 & 2): Both the imaging and genetics branches use the same autoencoder design. Each brain ROI (n_I = 246, 4 morphometric values) and genetic pathway (n_G = 177, 6 phenotypic traits) is encoded independently (preserving a one-to-one correspondence) by a 3-layer per-entity MLP into a 6-dim embedding. A mirrored decoder reconstructs the input. Each encoder also has a 2-layer MLP classifier. Encoders are pretrained with MSE reconstruction and BCE losses. Code will be released upon acceptance.

4.Comparison and Ablation (Revs 1 & 3): Our baselines span key alternative approaches: SUE (NeurIPS 2025) for unpaired cross-modal alignment, UNSEEN (NeurIPS 2023) for missing-modality settings, and G-MIND as an imaging-genetics-specific baseline. SUE requires paired data for fine-tuning, yet CALM still achieves statistically better accuracy and AUC than all three. None of the baselines outputs interpretable ROI/pathway associations, the central deliverable of CALM. Although classification performance is similar across CALM and its ablations, each loss imposes a specific constraint on A. L_cmmd aligns intra-class imaging and genetic distributions, giving A biological relevance. L_con pulls the intra-class samples together across modalities while pushing the inter-class samples apart, anchoring the alignment to the diagnostic axis. Together, they “organize” the latent space, making the associations stable and reproducible across CV folds and consistent with paired results. We note that AUC only evaluates single-modality classification and does not reflect the cross-modal alignment achieved by ‭L_cmmd‬ and ‭L_con‬. L_OT in Table 1 replaces L_cmmd with the Sinkhorn OT loss as a separate ablation.

5.Biological Validation (Revs 1 & 2): As our paper introduces a new framework to link unpaired datasets, we focus on methodology and provide a preliminary demonstration in autism, using single-modality cohorts (ABIDE, SSC) and a matched cohort with both modalities (ACE). Results show (1) high concordance between associations in unpaired and paired scenarios (Fig. 2a) and (2) stability and reproducibility across CV folds (Fig. 2b). As suggested, we are pursuing quantitative validation against the SFARI gene database and the Allen Brain Atlas.

6.Formatting (Rev 3): We will fix inconsistent superscript fonts, a missing colon in the regularization subheading, broken Fig. 3(b) citation, extra italic emphasis, and reference list format.



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.

    I recommend an Invite for Rebuttal based on the mixed reviewer scores. All reviewers commended the paper for tackling a highly practical and novel problem setting by learning cross-modal associations from completely unpaired imaging and genetics datasets. The proposed method was praised for its conceptually coherent design, its strong generalization to unseen external paired datasets, and its built-in interpretability achieved through linear projections. However, several technical and presentation concerns must be rigorously addressed in the rebuttal. First, the authors need to provide comprehensive implementation details regarding the exact architecture and pretraining protocols of the autoencoders and classifiers used in the pipeline. Second, critical design choices and hyperparameters, particularly the chosen latent dimension size and the various loss weights, require thorough justification and a sensitivity analysis. Reviewers also questioned the modest performance gains of the full model over its ablation variants and requested justification for the limited number of baseline comparisons. Furthermore, the rebuttal must explicitly clarify the mathematical rationale behind using the product of the parameter matrices to directly quantify pairwise relationships and defend its interpretability. Finally, the authors must commit to resolving the manuscript’s formatting inconsistencies and excessive italicization, and they should discuss potential avenues for stronger quantitative biological validation, such as incorporating external resources like the Allen Brain Atlas.

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

    Reject

  • Please justify your recommendation.

    Although the proposed framework addresses the highly practical problem of learning cross-modal associations from completely unpaired imaging and genetics datasets , reviewers remain unconvinced due to the modest empirical gains over ablation baselines. While the authors provided helpful architectural details and hyperparameter justifications during the rebuttal, the biological validation of the discovered associations remains largely qualitative rather than quantitative. Consequently, given these lingering concerns regarding the strength of the experimental evidence and the lack of robust biological validation, the paper ultimately falls below the acceptance threshold.



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.

    This manuscript presented a method to align unpaired cross modal data for biomarker discovery. Reviewers found this method to be a practical solution to a novel problem set with a conceptually coherent design.

    The weaknesses were missing implementation details, some design choices not being justified, and there being only a few baseline comparison methods.

    Overall the reviewers found the rebuttal addressed most points except for the modest gains in the full model over ablation studies which lead to questions about if the full model is really necessary. Additionally the biological validation of the method was qualitative only.

    I recommend for acceptance because on balance the negatives (modest gains in ablation experiment, limitations on comparisons) are outweighted by the strength of the novel design which is conceptually interesting.



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.

    The rebuttal addresses the reviewer’s concerns well. The work addresses a core challenge by learning cross-modal associations from completely unpaired imaging and genetics data. Overall, the reviewers are positive and I recommend acceptation.



back to top