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

Abstract

Gene expression profiles play a crucial role in understanding cancer pathogenesis, guiding personalized treatment, and predicting clinical outcomes. However, traditional gene expression detection relies on invasive tissue sequencing, which is time-consuming and costly. Existing image-based gene prediction methods are often limited to specific cancer types and fail to adapt to the pan-cancer scenario with diverse histological features. To address these challenges, we propose a novel pan-cancer image-based gene expression prediction method (PIGP) that enables gene expression prediction without the need for prior acquisition of cancer type information. Specifically, the model directly extracts discriminative features from pan-cancer histopathological images to preliminarily determine the cancer type, and further combines pan-cancer networks and corresponding cancer-type-specific networks to predict the corresponding gene expression levels. Extensive experiments on the pan-cancer dataset (HEST) have shown that the Pearson correlation coefficient for specific genes reaches up to 0.743. Moreover, heatmaps and distribution maps confirm the biological rationality and interpretability of the predicted results. This approach provides a cancer-type-agnostic computational framework for pan-cancer gene expression prediction and may support future translational studies. Our code is available at https://github.com/oceanflyfly/PIGP

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4489_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/oceanflyfly/PIGP

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SheYiy_CancerTypeAgnostic_MICCAI2026,
        author = { Shen, Yiyang AND Li, Bochong AND Wang, Huitao AND Chen, Zhiyu AND Li, Zhe AND Li, Xiu},
        title = { { Cancer-Type-Agnostic Pan-Cancer Gene Expression Prediction from Histopathological Images } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution of the paper is the formulation of pan-cancer gene expression prediction from histopathological images without prior knowledge of cancer type, framed as an open-set generalization problem.

  • 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.

    A major strength of the paper is the extensive experimental evaluation. The authors validate their approach across a large number of datasets (10) and compare different embedding strategies using as many as 13 models, which is a substantial effort. This breadth of experimentation provides a solid and comprehensive assessment of the proposed method and strengthens the credibility of the reported results.

  • 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.

    A main weakness of the paper concerns clarity and presentation rather than the overall approach. In particular, some claims appear overstated, such as describing the method as a “universal solution” with strong clinical translation potential, which should be moderated given the current level of validation.

    Additionally, the manuscript would benefit from clearer positioning of the data modality. It should explicitly state early on that the approach is based on spatial transcriptomics, rather than generically referring to histopathological images, to avoid confusion with standard WSI-only settings.

    Finally, some methodological details are not sufficiently specified. For instance, the selection and role of the genes used in the experiments are not clearly described, which limits interpretability and reproducibility.

  • 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?

    I recommend acceptance as the methodological approach is well-grounded and clearly structured. The experimental setup is comprehensive and up-to-date, and the extent of the evaluation is particularly strong, involving a large number of datasets and embedding models. These aspects collectively support the overall quality and relevance of the work.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have significantly improved the manuscript during the rebuttal phase, further strengthening a work that was already of high quality in its original submission. The revised version provides clearer explanations and a more polished presentation of the methodology and results, making the overall contribution more solid and convincing. Overall, I believe the revisions positively enhanced both the clarity and the impact of the paper.



Review #2

  • Please describe the contribution of the paper

    The paper’s main contribution is a novel method (PIGP) to infer gene expression from H&E-stained histological slides. The method aims to address two major limitations in existing gene expression prediction methods: lack of spatial context and cross-cancer heterogeneity. It does so by using 2 specific modules, each designed to deal with one of the drawbacks: a graph convolutional layer is tasked to bring spatial awareness in each spot’s representation while a duplicated branch deals with the cross-cancer generalization (with one branch being cancer specific and the other global).

    Extensive experiments on an adapted HEST-1k benchmark demonstrate superior performance compared to the approach introduced in the benchmark paper, across both tasks and foundation models.

  • 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.

    While spatial context modelling is not novel itself in gene expression prediction (HisToGene, Hist2ST, iSTAR, MISO), to the best of our knowledge, this is the first cancer-aware architecture proposed for this specific task within the digital pathology community.

    Beyond the novelty, the major strength of the paper lies in the scale of its experiments: the benchmark covers 13 foundation models and 10 indications, with the paper’s method demonstrating shared improvements across tasks and extractors. This outlines the broad generalization capability of the method.

    Another strong point is the ablation study. It effectively highlights the importance of the dual-branch component; compared to using a GCN alone, this addition almost systematically increases performance (in 125 out of 130 task x model pairs).

  • 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.

    We noted several weaknesses:

    Regarding methodology, gene expression prediction from H&E images is not a new task, and several methods have tried to tackle the issue (e.g. ST-Net, HisToGene, Hist2ST, iSTAR, MISO, in addition to those cited by the authors). There is no comparison with any existing methods beyond the simple ridge regression used in the HEST-1k paper. If none of the existing methods are applicable to the setting chosen by the authors, it would be interesting to state it explicitly and detail why.

    The paper would also benefit from several clarifications, regarding missing information and inconsistencies:

    • The authors state that 13 highly variable genes were selected for the tasks, whereas the original HEST-1k paper uses 50 genes. Is the ’13’ a typo (perhaps a mixup with the 13 foundation models ?) or an actual change. If it is the latter, why was this decision made ? And are the genes computed indication-wise or across the full set ?

    • Given that the authors state that the data is adapted from HEST-1k, are the results in this paper directly comparable to those in the original HEST-1k publication? If the results are not comparable, what motivated the changes ? The training set and validation set used in the paper need to be explicitly detailed (eg. same indications in training and validation ?) and, more generally, all the differences between the original HEST-1k benchmark and the deviations from the original HEST-1k benchmark should be clearly outlined.

    • The method proposed by the paper is tailored to support a pan-cancer training set. Does that mean that a single model was trained (for each foundation model) for all indications ? The HEST-1k benchmark trained one model per foundation model x indication. Obtaining better performance with a single model across several indications deserves to be highlighted more.

    • Where does the unsupervised phenotypic label c come from ? Is there a set of randomly initialized vectors, optimized alongside the network’s weights ? Likewise, how is the learnable parameter trained, is it during training or at inference time (perhaps the most likely given that the authors state that it dynamically weights the the outputs of both branches) ? If it is the latter, what is the procedure to train it without using training labels ? Finally, why are weighting factors necessary when merging y^_ridge and y^_res ? If y^_res is optimized to fit the residuals, after y^_ridge has been fitted, a simple addition should work. The architecture and training procedure should be detailed more.

    • The authors state «Comparisons with same tasks in the HEST benchmark [7] demonstrate that PIGP achieves significant performance improvements across all pathology foundation models, proving the framework’s advantages in capturing the morphological and spatial associations of gene expression. H-Optimus-0 performed most robustly, securing the highest Pearson correlation coefficients in six out of ten tasks, such as SKCM (r = 0.7302) and LUNG (r = 0.7283). Comparative analysis shows that the performance gain is most significant in the PAAD (pancreatic cancer) task, where the r-value improved from 0.5391 to 0.6646.». However, there are several data discrepancies between these claims and the provided tables:
    • The text claims H-Optimus-0 secures the highest Pearson correlation in six out of ten tasks. However, the preceding table only shows H-Optimus-0 as the top-performing model in two tasks.
    • The text states the PAAD baseline r-value is 0.5391, but the original HEST-Benchmark reports 0.4932 for H-Optimus-0.It appears the numbers in the text may have been mistakenly pulled from the ablation study table rather than the primary results.

    • The results in Table 1 do not align with the corresponding “with Dual-Branch Decoupled Learning” rows in Table 2.Assuming Table 1 is meant to present the results of the full PIGP architecture, these values should be identical.

    • Finally, if the results are indeed comparable, the original HEST-Benchmark results should be included in the same table next to the proposed method’s results. Without this side-by-side comparison, the claim that « Comparisons with same tasks in the HEST benchmark [7] demonstrate that PIGP achieves significant performance improvements across all pathology foundation models, proving the framework’s advantages in capturing the morphological and spatial associations of gene expression. » is rather difficult for readers to verify.
  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The submission does not provide sufficient information for reproducibility.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While the authors present a novel method, benchmark it extensively and shared a convincing ablation study, the paper currently suffers from several flaws:

    • there are several data discrepancies, with in particular the unresolved mismatch between Table 1 and Table 2.
    • the experimental setup is unclear, and seemingly different from the HEST-Benchmark setup to which the authors compared their method, making the core claim of performance improvement over HEST difficult to verify.
    • there are missing methodological details, with several components of the architecture left unexplained.
    • there is no comparison with any state-of-the-art methods.

    I am fully open to raising my score during the rebuttal phase if the authors can satisfactorily address the three major clarification and consistency issues, as new state-of-the-art comparisons is out of scope for 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.

    The authors have addressed and fixed the inconsistencies highlighted in the review, have addressed and are willing to add the necessary clarifications regarding the method and the experimental setup, and are willing to reframe the direct comparison with HEST-1k due to setup differences.

    While the need for this multi-stage architecture is not entirely backed up by the experiments (given the lack of comparison beyond the simple baseline of HEST-1k), I believe it should be accepted nonetheless due to other strong points (most notably, the cross-cancer nature of the approach).



Review #3

  • Please describe the contribution of the paper

    Authors propose PIGP, a novel pan-cancer image-based gene prediction framework. The core architecture of the framework is a dual-branch decoupled learning module, which simultaneously captures universal expression trends and cluster-specific heterogeneity across different cancer types. Experiments and visualization results demonstrated the effectiveness of the proposed framework on a pan-cancer dataset.

  • 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.
    • Methodological soundness: Authors clearly define the problem and design targeted modules to address it. They model pan-cancer image-based gene expression prediction under unknown cancer types as an open-set generalization problem, and their dual-branch decoupled learning module aptly addresses this problem. Furthermore, the design of other modules is also clear and reasonable.
    • Extendability: Even though this is not investigated in this work, I believe that the design of decoupling universal regularities and cancer-specific heterogeneity can be similarly applied to other pan-cancer tasks.
  • 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.
    • Writing issue: The overall writing of the paper is clear, but one issue remains. When introducing the dual-branch decoupled learning module(subsection 2.2), the description between the dual-branch architecture and the structure within each branch should not be conflated.
    • Limited experiments: The “Visual Analysis” part effectively demonstrates that the proposed framework effectively compensates for the lack of spatial prediction sensitivity found in existing models. However, the “Evaluation of Generation Performance” part lacks sufficient analysis, and the “Ablation Study” part is limited (Only the dual-branch decoupled learning module is ablated).
  • 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

    Questions to authors:

    • In the “Evaluation of Generation Performance” part, you mention that “Although PRAD and HCC prediction tasks still face bottlenecks (r < 0.12) due to high tissue heterogeneity, foundation models still exhibit a generational performance advantage compared to traditional supervised ResNet50.” The latter part of this sentence (comparing foundation models to ResNet50) seems irrelevant to the main theme of the paper. Moreover, the poor prediction performance is simply attributed to “high tissue heterogeneity” without further analysis. Since the proposed framework already introduces cancer-specific branches to handle heterogeneity across different cancer types, why does this issue still persist? What do you believe are the deeper reasons behind this phenomenon? And do you have any potential solutions to address this problem?
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

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

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

    Overall, despite the limitations discussed above, I believe this work still has certain value. The following suggestions are provided: ·In Section 2, line 6 and Subsection 2.2, I suggest that authors present the dual-branch decoupled learning module with greater clarity and better organization. Authors should clearly state that the dual-branch architecture is designed to concurrently model universal expression trends and cluster-specific heterogeneity, while the structure within each branch (i.e., Ridge regression + MoE) is intended to capture fine-grained expression fluctuations. ·In the “Evaluation of Generation Performance” part, I recommend that authors provide a more detailed and in-depth analysis of why the prediction performance is poor on PRAD, HCC, and other challenging cancer types. ·In the “Ablation Study” part, given the page limitations, evaluating such a large number of encoders is not essential. A smaller set of representative encoders would adequately support the main claims, and the spared space would be better allocated to showing ablations of the other components. Since adding new experiments may not be allowed, I raise this point just as a suggestion for you rather than a mandatory request for revision.

  • 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 addressed all of my previous concerns.



Author Feedback

We thank the associate editor and reviewers for evaluating our manuscript and recognizing our dual-branch pan-cancer gene prediction structure. Below, we address the meta-review concerns. (a)Reported inconsistencies in the results 1.Response to data inconsistencies (Reviewer #2) The inconsistency originated from the manuscript compilation stage rather than the training pipeline itself. We have compiled updated data under a unified configuration for the revised manuscript, with conclusions and performance trends unchanged. 1.1 H-Optimus-0 performance: Performed best on two tasks: IDC(r=0.6798) and COAD(r=0.6705). 1.2 PAAD performance: Manuscript analyses used uni_v1, not H-Optimus-0.The 0.5391 result was reproduced from HEST-1k (differing from the original publication), while 0.6646 is our experimental result. 1.3 Table 1 vs. Table 2: Compiling mixed data versions caused inconsistencies. As Reviewer #2 noted, all “w/” rows in Table 2 have been revised to adopt Table 1’s data. (b) Benchmark setting compared with the original HEST setting 1.Response to Reviewer #2 HEST-1k trained separate networks per cancer type using larger gene sets with higher expression levels. Consequently, comparability with our pan-cancer study is limited. (c) Training/validation protocol and selection of top genes 1.Training/validation details (Reviewer #2) Datasets were pooled across all ten cancer types. 1/4 (rounded up) of samples were randomly selected for validation, with the remainder used for training. 2.Gene selection (Reviewers #1 & #2) The original reference selected top-50 highly expressed genes per cancer. We selected 13 genes common across all ten cancers. Gene roles: CD3E, CD79A, CD8A, GZMK, TNFRSF17 (T/B cell markers); CCR7, CXCR4, GPR183 (spatial navigation); KIT, KLRB1 (other immune); ACTA2, PDGFRA (stromal/structural); MKI67 (proliferation) (d) Architecture details and training procedure 1.Model framework (Reviewer #2) A single model was applied to all cancer types. Unsupervised clusteringbased phenotypic label c is concatenated with image features. The learnable parameter α updates during training loss and backpropagation. An 8:2 weight ratio balances high prediction stability (linear ŷ_ridge) with spatial feature gains (deep network ŷ_res). (e) Rationality of the experimental setup 1.Comparisons & HEST-1k baseline (Reviewer #2) Existing representative methods focus on specific tissues/cancers (mostly breast) and fine-tune on single-cancer datasets. Our framework emphasizes pan-cancer prediction. Direct comparison should be interpreted cautiously due to different optimization objectives and training settings.

Responses to additional reviewer comments: 1.Reproducibility (Reviewers #1, #2, #3) We have finalized partial code repository following ablation experiments. Raw data is publicly available and aligns with HEST-1k; our gene selection utilized this raw data. 2.Writing suggestions (Meta-reviewer, Reviewers #1, #3) The revised manuscript clarifies data modalities and softens the statements about clinical translation. Section 2.2 has been restructured to general-to-specific structure to better describe the dual-branch module. 3.Performance on complex cancers (Reviewer #3) Selecting genes based purely on cross-cancer commonality led to low expression and variance in cancers like PRAD and HCC, which reduces prediction performance. Future improvements include revising the gene list or dynamically adjusting the ŷ_ridge to ŷ_res weighting based on round-wise predictions to enhance spatial feature detection. 4.Ablation experiments (Reviewer #3) Our core modules are the dual-branch decoupled learning module and adaptive fusion module, which are interdependent. Current ablation experiments adequately support the rationality of our core architectural design. If required, we can test alternative weighted fusion strategies.

We hope these responses address all major concerns of the reviewers and the associate editor.



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.

    On the one hand, the paper is potentially useful in view of its (i) timely focus on a pan-cancer formulation, (ii) broad evaluation of different datasets and foundation models, (iii) encouraging preliminary results, and (iv) a promising dual-branch design. On the other hand, there are major concerns about the clarity and consistency of the experimental evaluation, as well as their verifiability. In particular, the following issues should be addressed in the rebuttal: (a) reported inconsistencies in the results, (b) benchmark setting compared to the original HEST setting, (c) the training/validation protocol as well as the choice of top genes (d) the architecture details and the training procedure. The lack of experiments/evaluation with existing methods for gene expression predictions should be explained. Also, the claim for universality and its clinical translation should be moderate. The weakness analysis in specific cancer types should be provided as well as more ablation studies on the architecture. The authors should provide a focused and well-explained rebuttal on these questions.

  • 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.

    clear 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.

    This work addresses the challenge of predicting gene expression from histopathological images across unknown cancer types using a dual-branch decoupled learning module , and is strengthened by its evaluation scale spanning 10 indications and 13 foundation models. The authors resolved major consistency flaws and implementation ambiguities highlighted in the initial round of reviews. The scale of their evaluation across multiple foundation models and indications provides a solid benchmark for pan-cancer spatial transcriptomics. Given the consensus among reviewers following the rebuttal, acceptance is recommended.



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 paper presents a timely pan-cancer gene expression prediction framework with a cancer-type-agnostic dual-branch design and broad evaluation across cancer types and foundation models; while the original reviews raised important concerns about result inconsistencies, benchmark comparability, gene selection, and methodological clarity, the rebuttal satisfactorily clarified or mitigated these issues, and all three reviewers updated their final opinions to accept. The remaining limitations, including limited comparison to prior single-cancer methods and the need to moderate clinical/universality claims, are important but do not outweigh the methodological value and potential impact of the work.



back to top