Abstract

The high cost of spatial transcriptomics (ST) has driven extensive studies into predicting gene expression directly from H&E histology images. However, this prediction task faces an inherent limitation, as tissue morphology alone provides insufficient information to fully resolve underlying gene expression. To address this limitation, a recent study leverages partial gene expression to guide the prediction process alongside histology images. Building on this paradigm, we approach the prediction task as a spatial imputation problem, employing a Masked Autoencoder (MAE) to utilize a small fraction of gene expression as genetic anchors for inferring whole-slide gene expression profiles. Specifically, we propose a bio-saliency score and a learning-to-rank strategy to adaptively identify the most informative spots within the tissue. Based on these identified spots, our framework selects contiguous regions as genetic anchors to ensure suitability for real-world ST profiling hardware. To effectively leverage these anchors, we design a cross-modal joint encoder that integrates visual and genetic modalities. By aligning the selected anchors with their corresponding visual features via contrastive learning, the encoder generates robust joint representations to accurately predict gene expression across the whole slide. Notably, our framework consistently surpasses existing methods in both histology-only prediction and spatial imputation, achieving superior accuracy even without genetic anchors and further excelling with as little as 10% transcriptomic coverage. Our code is available at https://github.com/Kyyle2114/CAMMST.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1131_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Kyyle2114/CAMMST

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KimJoo_Contrastive_MICCAI2026,
        author = { Kim, Joohyeok AND Jeong, Taejin AND Kim, Jinyeong AND Hwang, Seong Jae},
        title = { { Contrastive and Adaptive Multi-modal Masked Autoencoder for Spatial Transcriptomics } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a novel multi-modal spatial transcriptomics prediction framework (CAMMST) that reformulates gene expression prediction as a spatial imputation problem under the masked autoencoder (MAE) paradigm and introduces an adaptive, biologically-informed sampling strategy to effectively utilize limited gene expression measurements.

  • 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 motivation to sample contiguous regions instead of randomly scattered points is highly practical, as it perfectly aligns with the physical constraints of real-world ST profiling hardware (2) Formulating the spatial imputation task within a masked autoencoder (MAE) framework is intuitive. Furthermore, the introduction of a scale-aware ranking loss to mitigate cross-slide batch effects and a soft-target contrastive loss to handle adjacent biologically similar spots are clever technical adaptations. (3) As demonstrated in Table 2, CAMMST significantly outperforms the direct baseline (PH2ST) in terms of computational overhead. Reducing model parameters by ~85% and training time from hours to minutes makes this method highly scalable for clinical environments.

  • 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 paper demonstrates improvements in numerical metrics (MSE, MAE, PCC), predicting gene expression is ultimately a proxy task. The paper lacks downstream biological analysis (e. g. , spatial clustering, cell type deconvolution, or pathway enrichment analysis) to prove that the improved numerical accuracy actually translates into more reliable biological or clinical insights. (2) The proposed bio-saliency–guided sampling strategy is intuitive but appears somewhat incremental. The idea of prioritizing informative or high-variance regions is conceptually related to existing approaches (e. g. , SEPAL), and the empirical improvements brought by the sampler are relatively modest according to the ablation study. In addition, the sensitivity of the method to the specific design of the saliency score and related hyperparameters is not thoroughly analyzed. (3) Figure 2 only provides visual predictions for two specific genes (FASN and GNAS). Given the complexity of the whole-slide transcriptomic profile, visualizing spatial distributions for spatial domains or a broader set of marker genes would make the results much more convincing.

  • 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

    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 proposed CAMMST framework tackles a meaningful clinical problem with a highly practical and hardware-aware design. By sampling contiguous regions, the method naturally aligns with the physical constraints of real-world spatial transcriptomics hardware. Furthermore, formulating the imputation within a masked autoencoder paradigm, which is cleverly augmented by a scale-aware ranking loss and soft-target contrastive fusion, yields impressive computational efficiency gains and performance improvements over the baseline. However, a higher score is precluded by several limitations in the experimental evaluation. Specifically, the manuscript lacks downstream biological validation, such as spatial clustering or pathway enrichment analysis, to demonstrate that the improved numerical metrics actually translate into reliable biological insights. Additionally, the framework introduces several empirical hyperparameters without providing a sensitivity analysis to verify model robustness. Finally, expanding the qualitative visual results beyond two genes and providing a more granular evaluation of anchor ratio dependencies, like testing performance degradation below the ten percent coverage mark, would significantly strengthen the paper.

  • 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.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper presents a practical and well-motivated framework for spatial transcriptomics prediction that goes beyond pure histology-based inference. Instead of assuming that morphology alone is sufficient to fully recover gene expression, the authors formulate the task as a spatial imputation problem, where a small fraction of real measured gene expression is used as genetic anchors to guide whole-slide prediction. I think this is the main contribution of the paper: it moves the problem setting closer to biological reality and clinical practice, since gene expression is only partially observable in many realistic scenarios and relying on a limited number of true measurements is more reasonable than expecting histology alone to explain the full transcriptomic landscape. Methodologically, the paper combines adaptive anchor-region selection, contiguous region sampling, and a multimodal MAE-style fusion framework, which together provide a coherent solution for this setting. The empirical results further suggest that this design is effective across both histology-only prediction and partial-imputation settings.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper does not simply propose another image-only predictor, but explicitly addresses the limitation that morphology and gene expression are not perfectly coupled. Reframing the task as prediction with a small amount of real transcriptomic guidance is both biologically sensible and practically meaningful. The paper evaluates on multiple datasets, compares against both histology-only baselines and a recent imputation-oriented baseline, and reports gains under different anchor ratios.

  • 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.My main concern is that the comparison space for the imputation setting could be broader. As it stands, the paper shows improvement over PH2ST and image-only baselines, but the value of the proposed formulation would be even more convincing with a more comprehensive imputation-oriented comparison. 2.Some ablation variants are numerically quite close to the full model, especially in PCC/MAE/MSE, so the practical importance of certain design choices is a bit difficult to judge from raw averages alone. In this context, adding statistical significance analysis, confidence intervals, or at least a clearer discussion of variance across folds would make the conclusions more credible.

  • Please rate the clarity and organization of this paper

    Good

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

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

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

    N/A

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

    N/A

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

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

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

    I gave this paper a weak accept because the overall idea is well motivated and practically relevant: using a small amount of real gene expression to guide prediction is more realistic than relying on histology alone, and the method is supported by solid experiments. My main reservations are that the imputation-related comparisons could be broader and that some ablation gains are fairly small, so stronger statistical support would make the conclusions more convincing.

  • 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 #3

  • Please describe the contribution of the paper

    This paper proposes a spatial transcriptomics imputation framework that reconstructs gene expression across tissue sections by leveraging histological images and a subset of observed gene measurements (referred to as genetic anchors). The method introduces a bio-saliency-guided sampling strategy to select informative spatial regions and employs a masked autoencoder (MAE) with contrastive learning to infer missing gene expression. Experiments are conducted on three spatial transcriptomics datasets, demonstrating improved performance over image-only baselines, particularly under limited observation ratios.

  • 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 problem formulation The paper reformulates spatial transcriptomics prediction as a spatial imputation problem rather than a purely image-to-gene regression task. This is a meaningful shift, as incorporating partial gene observations alleviates the ill-posed nature of cross-modal prediction.

    2.Integration of multimodal information The method effectively combines: histological image features, and partially observed gene expression. This joint modeling allows the network to leverage both morphological cues and direct molecular signals, leading to improved predictive performance.

    3.Practical motivation The idea of selecting a subset of spatial regions for measurement aligns with real-world constraints in spatial transcriptomics, where full-resolution profiling is costly.

    4.Empirical improvements The method consistently outperforms image-only baselines across multiple datasets, especially at low observation ratios (e.g., 10%), demonstrating its practical advantage.

  • 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.Limited evidence for generalization across tissues Although the method is evaluated on three datasets, the experimental setup does not convincingly demonstrate generalization. Two of the datasets are derived from similar breast cancer tissues, and the third remains within tumor histology. Moreover, all evaluations are conducted within each dataset rather than across datasets. Given that the relationship between morphology and gene expression is highly tissue-dependent and patient-specific, the lack of cross-dataset or cross-domain evaluation makes it difficult to assess whether the proposed approach captures generalizable principles or merely adapts to dataset-specific patterns. This limitation is particularly important because the method relies on learned statistical associations rather than mechanistic modeling.

  • 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

    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?

    The paper presents a well-motivated reformulation of spatial transcriptomics prediction as a spatial imputation problem, and the incorporation of partial gene observations (genetic anchors) provides a practical and effective way to improve over image-only approaches. The method is technically sound, and the experimental results show consistent gains, particularly under limited observation settings.

    However, my primary concern—also the main factor limiting my score—is that the experimental evaluation does not sufficiently demonstrate generalization across different tissues or domains. Given that morphology–gene relationships can vary substantially across biological conditions, this aspect is critical for assessing the real-world applicability of the method. While the approach performs well within each dataset, the current evidence is not sufficient to determine whether the learned mapping and sampling strategy would remain effective under distribution shifts, particularly since the evaluated datasets may share similar characteristics.

    Overall, I find the idea promising and the empirical results encouraging, but stronger evidence on cross-domain robustness would be necessary to fully support the claims. My recommendation is therefore borderline and dependent on whether this concern can be adequately addressed in the rebuttal.

  • 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.

    While I find the use of three cancer-related datasets somewhat redundant, given that many gene-associated diseases exist and broader validation across different disease types would further strengthen the generalizability of the method, I still believe the current experimental setup is acceptable. Although the performance improvements are not particularly significant, the proposed framework nevertheless demonstrates a meaningful level of methodological novelty.



Author Feedback

We thank the reviewers for their constructive comments and for recognizing the practicality and efficiency of CAMMST. We address the main concerns below.

R1.Biological Insight: We acknowledge that direct downstream validation would further strengthen this point and consider it important future work. Our primary goal is to introduce a spatial imputation framework beyond histology-only prediction, and we therefore prioritized the formulation, method, and comparative evaluation. While we do not claim that numerical improvements guarantee superior downstream performance, numerical fidelity can still provide a meaningful foundation for downstream analyses. For example, SpaIM (Nature Communications, 2025) reported that methods with higher PCC also tended to perform better in downstream analyses such as spatial domain identification and DEG analysis. This suggests a positive correlation between numerical fidelity and biological validity. The consistent gains in Table 1 and the more faithful spatial patterns in Fig. 2 suggest that CAMMST better captures the spatial gene expression landscape and provides a stronger basis for downstream analyses.

R1.Sampling Strategy: As discussed in the paper, we acknowledge the conceptual link to SEPAL. However, utilizing variability itself is not our primary novelty. Instead, we use biological variability as a proxy for informativeness and leverage it as a target for learnable anchor selection. Table 3 shows that replacing the sampler network with random sampling causes the largest performance drop among the ablations, demonstrating that the sampling strategy is an important component. We also tested a global deviation-based design instead of the current local deviation-based design, as well as related hyperparameters such as tau_C and the number of neighbor spots. The current setting performed best, but these results were omitted due to space limitations.

R1.Visualization: We agree that visualizing more marker genes would be informative. Due to the page limit, we selected FASN and GNAS as representative tumor- and cancer-related biomarkers that are commonly examined in prior studies. We will update the visualization code for broader qualitative inspection.

R2.Comparison Scope: We agree that broader imputation comparisons are valuable. To our knowledge, PH2ST was the only directly comparable baseline, as it uses both histology and partial gene measurements. Gene-only imputation methods address a different setting because they do not use histological images. We therefore used PH2ST as the main comparison.

R2.Ablation Study: We agree that some ablation variants are numerically close to the full model. In particular, replacing the soft-target contrastive loss with a standard contrastive objective yields the closest result. We believe this may be because standard contrastive alignment remains reasonably effective, while soft targets further reduce false conflicts among biologically similar spots. At the same time, the full model achieves the best overall performance, and removing any single component leads to a consistent drop. We view this trend as supporting our design motivation. To further clarify this point, we will also report fold-wise variance for the ablation variants.

R3.Generalization: We agree that robustness across tissues and domains is an important consideration. Our claim in this paper is not tissue-agnostic generalization, but consistent improvement under the current experimental protocol. In our setting, a direct cross-dataset evaluation is not straightforward, as each dataset uses different target genes. That said, CAMMST is designed to rely less on morphology alone, as it uses not only morphology but also measured gene expression from the target tissue as genetic anchors. These anchors provide information about the target tissue, which can help the model make more robust predictions under cross-dataset or cross-domain shifts than image-only methods.




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.

    Paper Summary The paper addresses the problem of predicting spatial transcriptomics by reformulating it as a spatial imputation task rather than relying solely on histology-based prediction. It proposes a multimodal framework that leverages both histological images and a subset of observed gene expression measurements as anchors. The method adopts a masked autoencoder paradigm, combined with a biologically informed sampling strategy that prioritizes contiguous and informative regions, along with tailored loss functions to handle batch effects and biological similarity. The goal is to provide a more practical and biologically grounded solution for reconstructing gene expression across tissue sections under limited measurement settings.

    Key Strengths The work is consistently recognized as well motivated and practically relevant, particularly in moving beyond purely image-based prediction to incorporate partial gene measurements, which better reflects real-world constraints. The reformulation as a spatial imputation problem is viewed as meaningful and biologically sensible. The integration of multimodal information and the use of contiguous region sampling aligned with hardware constraints further strengthen its applicability. Reviewers also note solid empirical performance across datasets, improvements over baselines especially in low observation regimes, and strong computational efficiency gains, suggesting scalability for clinical use.

    Key Weaknesses The main concerns center on the depth and scope of experimental validation. Reviewers highlight the lack of downstream biological analyses to demonstrate that improved numerical metrics translate into meaningful biological insights. The evaluation is also considered limited in terms of generalization, with insufficient evidence across diverse tissues or cross-domain settings. Some methodological components appear incremental or contribute modest gains, and the importance of certain design choices is not fully supported by statistical analysis or sensitivity studies. Additionally, comparisons to other imputation-focused methods and qualitative visualizations are seen as insufficient.

    Review Summary All reviewers agree that the paper proposes a meaningful and practical reformulation of the problem, with a technically sound framework and consistent empirical improvements over baseline methods. There is also consensus that the approach is well aligned with real-world constraints and offers efficiency advantages. However, they collectively express concerns about the experimental evaluation, particularly the lack of biological validation, limited generalization analysis, and insufficient depth in ablation and comparison studies. Differences arise in how strongly these issues affect confidence, with some reviewers viewing the contributions as sufficient while others see the limitations as significant barriers. Given these mixed perspectives, the positive aspects are acknowledged but are not yet fully supported by comprehensive validation, justifying further clarification.

    Decision Invite for rebuttal: the paper shows promising and practical contributions, but key concerns about biological validation and generalization need to be addressed.

    Rebuttal Focus In the rebuttal, the authors should focus on clarifying how improved prediction metrics translate into meaningful biological or clinical insights, potentially through additional justification. They should address concerns about generalization by discussing robustness across tissues or domains. Strengthening the empirical evidence with clearer justification of design choices would help resolve uncertainties. Clarifying sensitivity to key components would further improve confidence in the method.

  • 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 adequately addresses the main post-review concerns by clarifying that the paper’s claims are limited to consistent improvement under the submitted experimental protocol, not tissue-agnostic generalization, and by explaining why PH2ST is the most directly comparable multimodal imputation baseline. While downstream biological validation and broader cross-domain evaluation remain limitations, the reviewers’ final opinions are predominantly positive, and the paper offers a practical, efficient, and methodologically meaningful contribution; I therefore recommend acceptance.



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.

    The paper addresses the financial and physical constraints of full-resolution spatial transcriptomics by introducing a multi-modal masked autoencoder that uses a hardware-aware sampling strategy to impute missing expression profiles from partial gene anchors and histology. Its key strengths lie in moving beyond flawed histology-only predictions and offering significant gains in training efficiency. The rebuttal effectively clarified baseline selection limitations and highlighted the technical barriers to cross-dataset gene matching. While downstream biological benchmarking remains an open avenue for future work, the technical novelty and the practical relevance justify accepting the paper.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    After rebuttal, R2 and R3 support acceptance, with R3 noting remaining limitations on dataset diversity and modest gains but finding the methodological novelty sufficient, while R1 maintains Reject. Given the remaining concerns do not appear blocking, I vote for acceptance.



back to top