List of Papers Browse by Subject Areas Author List
Abstract
Spatial transcriptomics enables profiling of spatial gene expression but is limited by high cost and low throughput, motivating prediction from H&E histopathology images. Existing context-aware methods mainly supervise absolute expression, while relative expression relationships between spots are rarely used explicitly.
We propose COAST, a context-aware differential learning framework for spatial gene expression prediction.
COAST conditions the local and global context features with type-specific modulation and aggregates the target and context spot tokens using a Transformer encoder to capture both fine-grained local patterns and slide-level structure.
It is trained with a joint objective that combines absolute expression regression with signed differential regression between the target and context spots.
Experiments on multiple spatial transcriptomics datasets show consistent improvements in correlation- and distribution-based metrics, demonstrating the effectiveness of context-aware differential learning for histology-based spatial gene expression prediction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5249_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/QuIIL/COAST
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ByeKeu_COAST_MICCAI2026,
author = { Byeon, Keunho AND Park, Sunhong AND Lim, Jeewoo AND Kwak, Jin Tae},
title = { { COAST: Context-Aware Differential Learning for Gene Expression Prediction in Spatial Transcriptomics } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes COAST, a framework for predicting spatial gene expression from H&E images. The method combines type-specific feature modulation for target and context spots, a Transformer encoder for joint context modeling, and an auxiliary differential regression loss that supervises expression differences between spots in addition to absolute prediction. Experiments are conducted on seven SpaRED datasets, along with a downstream survival analysis on TCGA-LUAD.
- 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 idea of explicitly modeling relative expression differences between nearby spots is biologically well motivated. The differential regression objective is a reasonable way to encode this prior, and the ablation results support its contribution. (2) The experimental evaluation is reasonably comprehensive, with results on seven datasets, multiple metrics, qualitative visualizations, and ablations of the main design choices. (3) The downstream survival analysis, while limited in scope, is a useful addition and suggests that the predicted expression profiles retain clinically relevant information.
- 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 overall framework is reasonable, but the methodological novelty appears somewhat limited. The main components, feature modulation, Transformer-based aggregation, and auxiliary differential supervision, are all established ideas, and the paper could better clarify what is uniquely enabled by their combination. (2) The survival prediction experiment is a useful addition, but the results are fairly modest and somewhat inconsistent across folds. As presented, this evidence only provides limited support for the claim that the learned representations capture clinically meaningful prognostic information. Reporting statistical testing or stronger downstream validation would make this part more convincing. (3) The baseline comparison could be strengthened by including more recent context-aware methods, such as TRIPLEX (CVPR 2024), M2OST (AAAI 2025), and MERGE (CVPR 2025), to better position the contribution relative to the current SOTAs.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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 paper studies a relevant problem and presents a method that is generally well motivated, with reasonably broad experimental evaluation across multiple datasets. I also found the differential supervision idea intuitively sensible, and the results suggest that it contributes to performance. However, my overall score is limited by the relatively modest methodological novelty, the lack of comparison with several recent context-aware baselines, and the fact that some of the supporting evidence, especially the downstream survival analysis, is not yet strong enough to fully substantiate the broader claims. Overall, I think the paper has merit, but in its current form, I place it slightly below the acceptance threshold.
- 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.
Thank you for the detailed rebuttal. I appreciate the clarification that the contribution is a task-specific and biologically motivated adaptation of signed target-context differential supervision for H&E-to-ST prediction, rather than a claim of a generic new differential objective. The additional discussion on biological relevance and fairness of implementation is also helpful. My main concern, however, remains only partially resolved. In particular, I still find the lack of direct comparison with several recent and highly relevant context-aware methods (e.g., TRIPLEX, M2OST, and MERGE) to be a notable weakness. Because the paper is also positioned around context-aware modeling of target-context relationships, these comparisons would be important for more convincingly establishing the contribution relative to the current state of the art. The conceptual distinctions raised in the rebuttal are reasonable, but in my view they do not fully replace direct empirical comparison. I also appreciate the authors’ clarification regarding the survival analysis. This experiment is a useful additional evaluation, but as currently presented it still offers only limited support for stronger claims about prognostic relevance. Overall, I think this is a relevant and promising paper, but under the final binary decision, I lean slightly toward rejection.
Review #2
- Please describe the contribution of the paper
The paper proposes COAST, a context-aware differential learning framework for predicting spatial gene expression from H&E histopathology images. The primary innovation is a joint learning objective that combines absolute expression regression with a signed differential regression between target spots and surrounding context spots, explicitly supervising relative spatial relationships. The framework also introduces a type-specific feature modulation mechanism to adaptively integrate heterogeneous local and global contextual information. The method is validated on seven spatial transcriptomics datasets and demonstrates clinical utility through downstream survival prediction on the TCGA-LUAD cohort.
- 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 proposed signed differential learning objective represents a conceptually elegant approach that captures stable spatial gradients and relative biological relationships, effectively addressing the sensitivity of traditional absolute regression to slide-level batch effects. 2.The framework utilizes a sophisticated type-specific contextual feature modulation mechanism that adaptively conditions local and global features through learnable parameters, successfully resolving signal conflation in multi-scale spatial data. 3.The evaluation is exceptionally rigorous, spanning seven diverse datasets and demonstrating consistent, substantial performance improvements over strong state-of-the-art baselines across multiple metrics. 4.The inclusion of thorough ablation studies systematically validates each component of the architecture, such as context sampling sizes and the differential loss weight, demonstrating that the performance gains result from the full integrated configuration.
- 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 baseline comparisons may be influenced by the choice of feature extractor. Since the proposed framework utilizes a modern foundation model (UNI), it might possess an inherent representational advantage over older baselines. Clarifying whether baseline hyperparameters were optimally tuned for these specific datasets, along with providing deeper implementation specifics (e.g., target gene selection criteria, inference-time global context grid), would enhance the fairness and reproducibility of the evaluation. 2.While the downstream survival prediction results are promising, the clinical validation is currently limited in scale to a single dataset cohort (TCGA-LUAD). Exploring the framework’s performance across additional clinical cohorts or broader clinical tasks would more comprehensively establish the translational robustness of the predicted representations. Additionally, there is a minor presentation oversight in Table 3 regarding this experiment: the “Avg.” row for the number of patients and slides incorrectly displays the total sum (e.g., 162 and 294) rather than the actual average, which should be corrected for clarity. 3.The paper lacks a deeper biological interpretability and theoretical analysis regarding the properties of the differential learning objective. Specifically, it is currently unclear how this supervision affects different gene categories (e.g., housekeeping vs. tissue-specific genes), nor does it provide attention map visualizations to verify if the model focuses on biologically relevant morphological regions rather than random areas.
- 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.
- 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?
The paper addresses a genuine limitation in histology-based spatial transcriptomics prediction with an elegant and effective differential learning objective. The methodological design is technically sound, and the experimental validation is highly rigorous, demonstrating consistent state-of-the-art performance across a wide variety of datasets. The inclusion of the downstream survival prediction task is a strong addition that fits perfectly with MICCAI’s translational goals. The weaknesses identified primarily relate to the scale of the clinical validation and nuances in baseline comparisons, neither of which fundamentally undermine the strong core contributions of this work.
- 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.
The authors have satisfactorily addressed the primary concerns raised in my initial review. Regarding baseline fairness, their justification for evaluating original architectures rather than forcing the UNI extractor is methodologically sound, especially since COAST still outperforms PEKA. Furthermore, the authors effectively resolved my critique regarding biological interpretability by providing a detailed breakdown of performance across specific gene categories (tumor-associated, immune/TME, and stromal/ECM). While clinical validation remains limited to a single cohort, acknowledging this limitation is acceptable.
Review #3
- Please describe the contribution of the paper
This paper introduces COAST, a context-aware differential learning framework for spatial transcriptomics gene expression prediction. Unlike existing methods that solely regress absolute expression values, COAST explicitly models relative expression differences between target and context spots. By integrating both local and global contexts via a Transformer encoder, it effectively captures spatial gradients and structural variations.The reconstructed expression representations demonstrate strong prognostic value in predicting overall survival for TCGA-LUAD, highlighting their practical biological and clinical significance.
- 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.Instead of solely relying on absolute expression regression, the paper introduces “signed differential regression.” By explicitly modeling relative expression differences and spatial gradients between regions, the method is more robust against batch shifts and better captures true biological structures, such as tissue boundaries and tumor-stroma transitions. 2.The framework doesn’t just blindly aggregate neighborhood data. It explicitly distinguishes between local and global contexts using type-specific feature modulation (learning specific scale and shift parameters). This ensures heterogeneous spatial signals are processed appropriately before being fed into the Transformer. 3.Beyond standard predictive metrics, the paper validates the method’s downstream utility. The reconstructed spatial expressions successfully predict overall survival in TCGA-LUAD cohorts (achieving a higher C-index than image-only baselines), proving the model captures clinically meaningful biological information rather than just statistical correlations. 4.COAST consistently outperforms multiple representative baselines (e.g., STNet, Hist2ST, PEKA) across seven datasets and diverse metrics (PCC, MI, AUC, NRMSE). This cross-dataset, cross-metric success proves the method’s robustness.
- 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.
First, the methodological novelty is somewhat limited. While the proposed method works, it acts more as a task-specific integration of existing technical modules rather than a strong paradigm shift. The authors claim that this direction is “largely unexplored” in histology-based ST; however, Differential Ordinal Regression has clear precedents in pathology image analysis. For instance, the authors themselves cite DIOR-ViT, which explicitly models ordinal relationships between samples using a differential objective. COAST essentially migrates this concept from classification/ranking to regression. The underlying mathematical logic of predicting differences between sample pairs, rather than absolute values, is a mature concept in computer vision and metric learning.
Second, the study lacks biological granularity in its performance evaluation. The paper reports the average prediction performance across 128 genes but fails to discuss or analyze which specific categories of genes (e.g., tumor markers vs. basal metabolic genes) are predicted more effectively. This omission limits the biological insights and interpretability of the proposed framework.
- 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.
- 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
1.I suggest slightly tempering the novelty claims regarding the “signed differential regression” mechanism in the final version. Given the maturity of this mathematical logic in metric learning and computer vision, accurately framing it as a “clever adaptation and integration” for the heterogeneous context of spatial transcriptomics—rather than a ground-up exploration—will elevate the academic rigor of the paper and better align with the expectations of top-tier medical imaging venues. 2.Although the paper demonstrates the potential for overall survival (OS) prediction on the TCGA-LUAD cohort, for spatial transcriptomics research, supplementing the analysis of specific biological processes in the appendix or future work (e.g., evaluating the model’s performance in predicting genes related to the tumor microenvironment or tumor-stroma boundary gradients) would greatly broaden the interdisciplinary impact of this work. Understanding exactly which specific biomarkers correspond to the “structural consistency” captured by the model is key to bridging the gap between AI and clinical medicine.
- 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?
1.Comprehensive Benchmarking and Robust Empirical Performance The authors conducted extensive evaluations across seven distinct SpaRED datasets encompassing diverse tissues and species (e.g., AHSCC, EHPCP1, MMBO). The proposed COAST framework consistently outperformed six competitive baseline models across five independent evaluation metrics, showing notable improvements in both correlation (e.g., PCC, MI) and distribution-based metrics (e.g., NRMSE). This rigorous benchmarking provides strong confidence in the framework’s generalizability and robustness against tissue-specific variations.2.Practical Clinical Translation via Downstream Task Validation A significant strength of this work is its demonstration of clinical feasibility. The authors did not stop at expression prediction but validated the prognostic value of the reconstructed gene expression representations on the TCGA-LUAD cohort for overall survival (OS) prediction. Achieving an average C-index of 0.6029—which is highly comparable to the actual bulk RNA sequencing baseline (0.6042) —is a compelling empirical result. This underscores the practical value of the proposed method in utilizing cost-effective H&E imaging for clinical prognostication.3.Sensible Architectural Design and Solid Ablation Although differential regression is an established concept in metric learning, its specific integration into the spatial transcriptomics domain is well-executed here. The architecture seamlessly handles heterogeneous contexts by applying type-specific scale and shift parameters to modulate local and global features prior to Transformer aggregation. Furthermore, the joint optimization of absolute and differential regression objectives forms a coherent engineering solution. The provided ablation studies adequately validate these structural choices.Summary Statement: Overall, despite the limitations regarding deep biological interpretability (e.g., gene-specific performance analysis) and the derivative nature of the core learning objective, the paper presents a highly effective, well-engineered solution. The solid experimental design and the demonstration of clinically relevant outcomes make this a valuable contribution to the computational pathology community, warranting acceptance.
- 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 clarification on baseline fairness is reasonable. Using original implementations and settings while matching target genes, splits, and metrics is an acceptable evaluation protocol, and I agree that forcing all baselines to use the same foundation-model features could distort their intended designs. The inclusion of recent methods such as PEKA and NH2ST also helps address the concern about missing context-aware or foundation-model-based baselines.
Author Feedback
We thank the reviewers for the comments. In revision, we will clarify novelty, baseline fairness, recent baselines, survival statistics, and biological interpretation.
A1.(R1, R3) Uniqueness and novelty. COAST does not claim a new generic differential objective. Its contribution is a tailored spatial context-aware differential supervision for H&E-to-ST prediction. Existing ordinal/differential objectives are used for class ordering or relative differences among samples within the same class, whereas COAST defines intra-slide target-context expression differences as signed gene-wise values. This explicitly models biological spatial changes in ST (e.g., tumor-stroma boundaries and glandular regions). COAST also differs from existing context-aware H&E-to-ST models. TRIPLEX and M2OST learn the multi-scale image context centered around the the target, while COAST directly incorporates both proximal and distal context patches as separate spatial references. Although MERGE can utilize patches within the same slide, it aggregates them via GNNs. In contrast, COAST learns signed relative differences. Thus, the novelty of COAST lies not in differential learning alone, but in adpating singed gene-wise differential supervision to spatially related proximal and distal tissue contexts for ST prediction. We will revise the manuscript to present COAST as a task-specific and biologically motivated integration of context modeling and differential regression.
A2.(R2) Fairness of baseline implementation. To ensure fairness, we evaluated all baselines using their original implementations and settings while matching target genes, splits, and metrics. Target genes were the dataset-specific Moran’s I-based HVGs provided by SpaRED benchmark. Since several baselines include the image encoder itself, forcing all methods to use the same foundation-model features could distort their intended architecture rather than improving fairness. To compare against a recent foundation-model-based method, we included PEKA, which uses two powerful foundation models (H-optimus-0 and scFoundation). COAST outperforms PEKA. We will clarify these in the final version.
A3.(R1, R3) Missing or recent context-aware baselines. While the comparison does not exhaustively cover all recent context-aware methods, it includes representative strong baselines from the most relevant directions. NH2ST (MICCAI 2025) is a graph-based context-aware method, while PEKA (MICCAI 2025) is a recent foundation-model-based method. NH2ST also reported higher PCC than TRIPLEX on several benchmarks, supporting its role as a strong context-aware reference. Thus, our comparison includes context-aware and foundation model based methods, not only classical baselines. For relationship to context-aware methods, please see A1. A4.(R1) Statistical interpretation of survival analysis. We agree that fold-wise performance shows some variance and stronger downstream tasks would further validate our method. Fold-wise variation may be ascribable to the limited sample size and disproportionate distribution of survival events. Moreover, k-fold cross-validation is designed to be evaluated holistically, as the aggregated performance provides a more stable assessment. The MICCAI rebuttal guideline strictly prohibits the inclusion of additional experiments. We will acknowledge the fold-wise variance and a single downstream task as limitations of our study.
A5.(R3) Biological interpretation and qualitative interpretability. We agree that the 128-gene average limits biological interpretation. Using existing test predictions, we computed gene-category PCC. COAST outperformed PEKA/NH2ST for tumor associated genes (0.4100 vs. 0.3959/0.3096), housekeeping/basal genes (0.4100 vs. 0.3896/0.3261), immune/TME genes (0.3436 vs. 0.3239/0.2581), and stromal/ECM or tumor-stroma boundary genes (0.3630 vs. 0.3153/0.2490), indicating gains across biologically meaningful groups.
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.
The paper covers an important problem and the reported results indicate that the empirical performance is promising. However, concerns are raised on the methodological novelty, the positioning with respect to recent context-aware state-of-the-art methods, the validity of baseline comparison with the use of the modern foundation-model feature extractor, and the strength of the clinical and biological validation. In particular, the authors are encouraged to make clear why the proposed differential learning framework is unique, to justify the baseline comparison in term of the implementation and tuning fairness, to justify the missing/baseline and possibly the recent baseline in the comparison, to interpret the survival analysis result with more rigorous statistical evidence, and to add more biological interpretation such as the gene-category level performance or qualitative interpretability of the model. Please make a careful and focused rebuttal to the reviewers since it will be important for the final decision.
- 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 vote acceptance, finding the clarifications on baseline fairness, biological interpretation, and the role of recent baselines sufficient, while R1 remains a borderline Reject mainly due to missing direct comparisons with TRIPLEX/M2OST/MERGE and limited survival evidence. Given the relative strong results and that most post-rebuttal concerns were reasonably addressed, I vote acceptance of this paper.
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 reviewers agreed that COAST addresses an important H&E-to-ST prediction problem, with a sound context-aware differential learning framework and consistently strong results across seven SpaRED datasets. The rebuttal clarified the intended novelty as a task-specific integration of signed target-context differential supervision rather than a generic new differential objective, and it reasonably addressed concerns on baseline fairness, biological interpretation, and limitations of the survival analysis. While the lack of direct comparison to some recent context-aware methods and the modest clinical validation remain weaknesses, the overall empirical evidence, ablations, and two post-rebuttal accept recommendations support acceptance.
Meta-review #3
- 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.
Two reviewers recommended accept. The third however had concerns on the lack of direct empirical comparisons with highly relevant context-aware methods (TRIPLEX, M2OST, and MERGE). Additionally, the survival analysis in the paper offers insufficient empirical evidence to fully substantiate the paper’s claims regarding prognostic relevance. These are valid concerns and I’m inclined to reject the paper in its current form.
