Abstract

Matched multi-omics can improve WSI-based biomarker and prognosis prediction, but most existing pipelines use omics as a parallel feature stream or textual context rather than as an explicit retrieval constraint. HERO asks whether observed omics can be a testable morphology hypothesis: a sparse pathway-to-morphology prior maps DNA methylation and miRNA into a K-dimensional intent vector m (K=16), TF-IDF retrieval over structured 10× captions selects endpoint-relevant regions, and a cosine gate c=cos(m,v) triggers deterministic deficit-driven repair when c<τc. This closed-loop design bounds VLM calls, reduces reliance on embedding-based semantic matching, and makes every retrieval and verification step lexically auditable. On TCGA-BRCA (930 WSIs, patient-level 5-fold CV), HERO sets new state-of-the-art across ER, PR, HER2, subtype, and risk prediction, outperforming both multimodal fusion and VLM-based baselines.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4607_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiXia_HERO_MICCAI2026,
        author = { Li, Xiangyu AND Su, Ran},
        title = { { HERO: Hypothesis-Driven Evidence Retrieval from Omics for Multi-task Breast Cancer Analysis } },
        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

    This paper presents an interesting and potentially impactful idea: rather than treating omics as a conventional parallel feature stream, it uses omics as both a retrieval prior and a consistency constraint for pathology evidence selection. In particular, DNA methylation and miRNA are first transformed into an intent vector $m$ and a keyword set $Q$, which guide patch retrieval through caption-based lexical matching. The retrieved evidence is then verified against the molecular hypothesis via a consistency score $c=\cos(m,v)$, with an additional deficit-driven repair step when the alignment is insufficient. This closed-loop design is conceptually novel and distinguishes the work from standard multimodal fusion or post-hoc interpretability approaches.

    At the same time, I am not yet fully convinced by the current presentation and experimental support. While the overall framework is well motivated and the reported results are promising, several important aspects remain insufficiently specified or insufficiently validated. In particular, the method relies heavily on text-mediated intermediate representations, yet the prompts, structured outputs, and captioning details are not described in enough detail for reproducibility. More importantly, the paper does not fully rule out alternative explanations for the observed gains, especially given the strong correlation between the omics input and the prediction targets. Additional clarification on label provenance, stronger shortcut-control ablations, and more rigorous validation of the caption-based intermediate pipeline would substantially strengthen the paper. Overall, I view this work as promising, but I believe the current manuscript still leaves several key methodological and experimental questions unresolved.

  • 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.

    This paper presents an interesting and well-motivated framework for multi-task breast cancer analysis from whole-slide images and matched multi-omics. The main strength of the work is its clear methodological position. Instead of treating omics as a standard parallel feature stream, the proposed method uses omics to guide two key parts of the pipeline: where to look in the slide and whether the retrieved visual evidence is sufficient to support the molecular hypothesis. This is an important distinction from prior approaches, which mainly rely on post-hoc MIL localization, VLM-based reading without specific priors, or conventional multimodal fusion. In this sense, the paper offers a clear and recognizable contribution. Another strong point is the coherence of the overall design. The pipeline forms a closed loop from the omics-derived intent vector m, to the keyword set Q, to Top-64 patch retrieval, to the visual evidence vector v, the consistency score c, the deficit term d=m−v, and finally the repair step. This information flow is logically connected and easy to follow. I also appreciate that the method uses checklist matching and TF-IDF retrieval instead of relying only on embedding-based retrieval. This makes the retrieval and verification process more transparent and easier to inspect.

  • 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.Insufficient methodological detail limits reproducibility. The paper presents a clear four-stage framework, but many details that are important for reproduction are still missing. For example, the prompts, output formats, decoding settings, and keyword extraction rules are not fully described. These are not minor implementation choices, because they directly affect retrieval quality and evidence verification.

    2.Eq. (1) and several Stage 1 definitions remain unclear. The formulation of Eq. (1) is not fully precise, especially the normalization term, whose exact meaning is difficult to interpret from the current text. In addition, the paper does not provide the full definitions of the 16 morphology axes, the checklist terms, or the detailed pathway-to-axis assignments in W. As a result, the proposed prior is conceptually appealing, but still difficult to reproduce in practice.

    3.The implementation of Stage 3 repair is still under-specified. Stage 3 is one of the most important parts of the method, but some key steps are not explained clearly enough. In particular, it is not clear how the final “best 64” patches are selected after the repair patches are added back to the pool. It is also unclear whether the model captions the full 8×8 mosaic, each 20× patch separately, or both.

    4.The text-based bridge is interesting, but it may introduce a same-model bias. The method uses text as an intermediate bridge between omics and image evidence, which is an interesting design choice. However, the same model family is used both to generate intermediate captions and to make the final prediction. The paper shows that Qwen2.5-VL can generate captions, but it does not yet show strongly enough that these captions are pathologically accurate and stable enough to support the downstream reasoning process. Why not chose medical domian VLMs ?

    5.The risk of shortcut learning from omics is not fully excluded. The omics input is highly related to several prediction targets, especially HER2, subtype, and possibly risk. This raises the concern that the model may rely heavily on molecular priors instead of truly learning to organize image evidence. Although the paper compares different retrieval strategies under the same diagnosis head, stronger control experiments, such as removing or shuffling omics, would be needed to rule out this alternative explanation more convincingly.

    6.The data and label sources are not described clearly enough. The experimental section clearly reports the TCGA-BRCA cohort, including 930 WSIs, 878 patients, 14,115 CpGs, and 315 miRNAs, as well as the prediction targets ER, PR, HER2, subtype, and risk. However, the manuscript does not explain in enough detail how these target labels were obtained or defined. For example, it is unclear whether ER/PR/HER2 labels come from IHC/ISH results or from clinical phenotype tables, how subtype is defined, and how the risk labels are constructed. This weakens the interpretation of the results, especially because omics is used as an input while the targets are also strongly related to molecular status.

    7.The fairness of the baseline comparison needs to be discussed more carefully. The paper states that, for fairness, all multimodal baselines were implemented using the same concatenated DNA methylation and miRNA vectors as the proposed method. While this makes the input setting more uniform, it may also move some baselines away from their original or best-performing configurations. As a result, the comparison may reflect performance under a unified input protocol rather than the true strength of each baseline in its intended setting. This point is particularly important for methods such as MCAT, MOTCat, and SurvPath, which are based on different modeling assumptions.

    8.External validation is missing. All experiments are conducted on TCGA-BRCA using patient-level 5-fold cross-validation, which is a reasonable internal evaluation setup. However, this is still a single-dataset, single-cohort, and single-cancer study. For a method that depends heavily on caption quality, predefined vocabularies, and omics-to-morphology correspondence, the lack of external validation makes it difficult to assess robustness across scanners, staining conditions, institutions, and population shifts. The proposed consistency gate may improve internal reliability, but it does not by itself demonstrate cross-center generalization.

    9.Table 3 provides only limited support for the importance of W. The paper uses W-noise, W-drop, and W-shuffle to argue that the prior W captures meaningful biological structure. However, noise and drop lead to only small performance changes, while shuffle causes a more visible drop. This suggests that the coarse structure of W may be useful, but it does not strongly demonstrate that the precise expert-designed mapping is essential. In addition, the paper does not clearly define how noise and drop are implemented or how strong these perturbations are, which makes the robustness claim harder to evaluate. 10.The ablation study is still incomplete. The paper focuses the main ablation study on HER2 and Risk, which is reasonable because these tasks are likely to depend strongly on retrieval quality. However, since the method is presented as a unified multi-task framework, it would also be useful to show whether similar trends hold for ER, PR, and subtype. In addition, while the paper shows that text-based retrieval outperforms visual-only retrieval and that gate+repair is helpful, it does not provide more direct ablations of the final diagnosis stage, such as using only the final 20× mosaic, only the molecular narrative, only the consistency summary, or removing the narrative input altogether. Without these analyses, the contribution of each Stage 4 input remains unclear.

  • 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.

    (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 meaningful shift from treating omics as a parallel feature stream to using omics as a retrieval-controlling hypothesis. However, the current manuscript still leaves several important methodological and experimental questions unanswered, especially regarding reproducibility, shortcut risk, and the reliability of the text-mediated intermediate representations. I therefore view this work as promising but not yet fully convincing in its current form.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    The main contribution of HERO is a novel hypothesis-driven evidence retrieval framework that uses matched multi-omics data (DNA methylation + miRNA) as a testable morphology hypothesis to actively guide region selection and verify molecular-visual consistency in whole-slide images via a cosine consistency gate and deficit-driven repair, enabling more accurate, interpretable, and endpoint-aware multi-task breast cancer analysis.

  • 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.

    One notable aspect is the concept itself. Most papers merely regard biological data as an additional feature or something similar. However, in this case, it is utilized as a hypothesis that directs the entire retrieval process. I appreciate this methodology. The consistency gate mechanism, where the author(s) assess whether the observed image corresponds with the molecular signal, is quite ingenious. If the score falls below a certain threshold, it automatically seeks different patches. This approach is logical. The results are also exceptionally robust. The author(s) conducted tests on the TCGA dataset, which comprises 930 slide images, and they outperformed all other methods. The enhancement in HER2 is 0.050 AUC, which is a significant figure. Additionally, the improvement in risk prediction is 0.030 C-index. These enhancements are particularly noteworthy for challenging prediction tasks. Another commendable aspect is the transparency of the system’s operations. One can observe which regions it focuses on, the features it identifies, and how these contribute to the final outcome. This transparency is crucial for doctors or clinics, as they need to understand the rationale behind the machine’s decisions. The author(s) employ word matching instead of complex embedding techniques, which ensures that interpretations do not deviate into nonsensical territory. Everything can be traced back to the actual text. The ablation study presented in table 2 is informative. When the author(s) compare text-based retrieval with visual-only embedding, the text-based approach significantly outperforms by a margin of 0.077 on HER2 AUC. This finding is intriguing as it indicates that the significance of pathology terminology outweighs the aesthetic appeal or prominence of the images. Furthermore, the robustness aspect concerning the W matrix is commendable. The author(s) manually constructed W based on medical guidelines, and when they shuffled it, the performance decreased by 0.033.In contrast, when they merely added noise, the performance drop was only 0.009.This demonstrates that the author(s) are not merely fitting random elements but are genuinely encoding authentic biological knowledge.

  • 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.

    Initially, the author(s) conduct tests using only a single dataset. While TCGA is well-known, it is essential to evaluate other datasets to demonstrate the general applicability of the findings. Furthermore, the 930 slides are relatively small compared to contemporary standards. Regarding the axes defined, specifically the 16 morphological features, are this exclusive to breast cancer? Can this methodology be applied to other types of cancer? This remains uncertain. Additionally, the speed of the process raises concerns. They report a duration of 57 seconds per slide using two powerful GPUs, which is quite substantial. In a real hospital setting, such a wait time per case is impractical. Moreover, the VLM model is invoked multiple times at various stages, and the repair process occurs in 37% of the slides, indicating a significant amount of additional computation. The authors should provide more insight into which components consume the most time. Why was K=16 selected in the first place? As shown in Table 3, K=32 yields nearly identical results (0.867 vs. 0.865). Why not opt for 32 or even 8? Although they perform a sensitivity analysis, this suggests that the author(s) may not have thoroughly optimized this parameter. Which specific 16 axes did the author(s) choose, and what was the rationale behind selecting these particular axes? There is also a lack of discussion regarding potential failures (the sad path). They present one successful case in Figure 3, but what happens when the system identifies incorrect regions? What if the genetic data does not align with the morphological data? Such discrepancies can occur in real cancer cases. While the consistency gate provides some assistance, what about these adverse scenarios? The entire system relies on the Qwen2.5-VL model for generating image captions. If this model inaccurately describes the morphological features, it could lead to failures downstream. They claim that a structured prompt enhances reliability, yet the author(s) do not provide any analysis to support this. For instance, how often does the VLM misdescribe features? What percentage of the captions is genuinely accurate? The construction of the W matrix is also ambiguous. The author(s) mention adherence to WHO guidelines, but the specifics are unclear. What happens when different pathways suggest the same axis? To what extent is human judgment involved in this process? Would different individuals arrive at different W constructions? Finally, for actual deployment in a hospital setting, validation on entirely different datasets is necessary. This includes variations in hospitals, scanners, and staining techniques. Five- Five-Fold CV on single dataset is not enough.

  • 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

    TECHNICAL CONCERNS

    In equation 1 they put negative signs but are not clear enough. High methylation means pathway down? Should be clearer in text. The TF-IDF on 300 captions seem small. How stable is the IDF weight? If they pick 300 different representatives instead, does retrieval ranking change lot? The threshold for consistency is 0.5 but why? They show in table 3 it works but they do not really justify why 0.5 is good choice. Seems kind of arbitrary.

    Figure 3 is very messy and hard to read. Too much information, text too small. They generate “molecular narrative” text but then not really use it much afterward? Or maybe use but not explain clearly.

  • 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?

    This paper introduces an innovative and promising concept and utilizing matched multi-omics data as a testable morphology hypothesis to actively facilitate evidence retrieval and confirm consistency in whole-slide image (WSI) analysis. The hypothesis-driven retrieval, combined with a consistency gate and deficit-driven repair, is an intelligent approach that addresses a significant limitation in current MIL, VLM-based, and multimodal fusion methodologies. The findings on TCGA-BRCA are robust (new state-of-the-art across various endpoints, with substantial improvements on HER2 (+0.050 AUC) and risk prediction (+0.030 C-index)), and the ablation studies effectively demonstrate the importance of text-based retrieval and the consistency mechanism. The focus on interpretability (lexical matching, auditable evidence chain) represents a distinct advantage for clinical application.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #3

  • Please describe the contribution of the paper

    Authors introduce HERO, a four-stage pipeline designed for retrieval based on omics data. The framework not only retrieves relevant evidence but also evaluates whether the retrieved results align with expected outcomes. If discrepancies are detected, HERO initiates a repair loop to refine the retrieval process and improve consistency.

  • 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.
    • Novelty: HERO’s main strength is the introduction of a RAG-like pipeline for whole slide images (WSIs), where retrieval is not treated as a one-shot step but is coupled with a verification and repair loop. This verify-retry mechanism ensures that retrieved omics evidence is actively checked against expectations and refined when necessary, leading to more reliable outputs. The approach also represents a shift from standard practices.
  • 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.
    • Abstract’s opening claim: The first sentence of the abstract appears somewhat strong, as there is existing literature supporting the claim that predicting biomarkers and prognosis from WSIs benefits from matched multi-omics data. For example: Hu, Jia, et al. “Prediction of clinicopathological features, multi-omics events and prognosis based on digital pathology and deep learning in HR+/HER2− breast cancer.” Journal of thoracic disease 15.5 (2023): 2528.- Terminology and claims in the introduction: The term “evidence drift” could be reconsidered in favor of more established terminology such as “retrieval bias”, which may better convey the underlying issue. Additionally, the statement that VLM-based WSI readers and multi-agent systems explore without endpoint-specific priors appears somewhat strong. In practice, such priors can often be incorporated through prompting, and acknowledging this could provide a more balanced perspective.
    • Limited clarity in methodology: While the architectural components are presented, the overall description remains quite concise, which may hinder full comprehension. In particular, some variables and acronyms are not explicitly defined. A more structured and detailed exposition of the pipeline, including clearer definitions and intermediate steps, would improve readability and make the method easier to follow.
    • Training and fine-tuning details: It is not entirely clear why only Stage 4 is fine-tuned with LoRA. Clarification on whether tuning other stages was considered, and how this might affect performance, would be valuable. Moreover, additional details on the training process across all models, especially regarding convergence criteria and fine-tuning protocols for the baselines, would help ensure a fair and transparent comparison.
  • 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
    • Formatting issue in Section 2.3: The equation extends beyond the page margins, which affects readability. Adjusting the formatting would improve the presentation.
  • 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 thank the Authors for their work and for addressing an important problem. The paper presents an interesting and conceptually meaningful approach with promising results, but would benefit from clearer methodological exposition and more detailed experimental reporting.

  • Reviewer confidence

    Somewhat confident (2)

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

    N/A

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

    N/A



Author Feedback

We thank the reviewers for their careful reading and constructive feedback. It is encouraging that the main point of HERO was clear: we use omics not as another feature vector, but as a hypothesis that tells the model where to look and then checks whether the retrieved tissue evidence agrees with it. Several parts of the paper were too compressed, and a few claims should be worded more carefully. We do not add new experiments here; instead, we will use the camera-ready revision to add concise clarifications where space permits.

We will add a compact algorithm/table with the 16 morphology axes, example checklist terms, the Stage-1 output schema, the 10x/20x prompts, decoding settings, and the construction and perturbation of W. Eq. (1) will be rewritten to state that the denominator is pathway-wise averaging over selected features, and that the negative sign converts promoter hypermethylation and miRNA-mediated repression into pathway down-regulation scores. We will also state that W is binary/nonnegative before row normalization, fixed across folds, and not learned from endpoint labels.We will also clarify the repair step. The 10x captions are generated for the 300 representative patches. The selected 20x crops are arranged into a coordinate-labeled 8x8 mosaic and processed by the VLM in one call, which returns structured tile descriptions. If c < tau_c, we compute the deficit from m-v, score unselected representatives by score_10 + w’^T a_i, add the top 32 candidates, and rebuild the final 64-patch mosaic from this repaired pool before one more 20x captioning pass. This final selection step will be described more clearly.

We do not claim that omics alone is uninformative; several targets are molecularly related by definition. Our point is that HERO uses omics to organize visual evidence, not to replace it. We will clarify that endpoint labels come from TCGA clinical/molecular annotations and survival endpoints, not from the input omics vectors used by HERO, and we will spell out the definitions for ER/PR/HER2, subtype, and risk. We will also soften wording such as “prevents shortcut learning”: W, lexical retrieval, and the controlled VLM baselines are safeguards and diagnostic checks, not proof that shortcuts are impossible.For baselines and parameters, we will add the missing caveats. Multimodal baselines were run with the same DNA methylation+miRNA input for a controlled comparison, but this may not be each method’s best original setting. K=16 was chosen as a practical middle point: K=8 is less specific, while K=32 gives little extra gain but requires more vocabulary design. tau_c=0.5 was not tuned per endpoint; the sensitivity results show similar behavior from 0.3 to 0.7.We will also define W-noise, W-drop, W-shuffle, and state that TF-IDF is fit locally on the 300 representative captions of each slide. Only Stage 4 is tuned because Stages 1-3 are kept deterministic and frozen, so the effect of the retrieval and verification mechanism can be isolated.

We used Qwen2.5-VL because it gives an open and reproducible captioning/diagnosis pipeline under one setup. However, structured prompting does not guarantee pathologic correctness. We will revise the text to say that checklist matching and the consistency gate make the evidence chain inspectable, while caption errors and possible same-model bias remain limitations.Finally, we will state the other limitations more directly. TCGA-BRCA 5-fold CV is internal validation. We also acknowledge the computational overhead (57s/WSI) raised by Reviewer #3; optimizing the multi-stage VLM calls and inference speed is a critical next step before real-world clinical deployment. We will replace or qualify “evidence drift” as “retrieval bias”, soften the abstract opening, improve Fig. 3 readability, fix the long equation formatting, and add the missing LoRA/training details.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 authors propose HERO, a novel retrievel framework to integrate omics data in computational pathology analysis. While the paper could benefit from providing more methodological details and experimental reporting, all reviewers agree that the method is novel and promising. The paper can create opportunity for discussion and potentially inspire similar work. Therefore I recommend acceptance.



back to top