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

Abstract

In precision oncology, multimodal survival analysis is challenged by overfitting from high-dimensional vision foundation model (VFM) features and limited cross-modal interaction in unidirectional fusion. We propose Di-PACT, a multimodal discrete survival framework that integrates whole-slide images with pathway-level transcriptomic representations. Di-PACT combines distance-aware stratified sampling, direct-bottleneck encoding, bidirectional pathway co-attention, and three-branch residual fusion to reduce VFM redundancy, preserve tumor microenvironment heterogeneity, and jointly model WSI-specific, pathway-specific, and cross-modal prognostic information. This design facilitates robust multimodal coordination under modality imbalance and provides multi-level interpretability cues linking prognostic pathways with morphological patterns. A discrete-time survival objective captures non-linear survival dynamics beyond proportional hazards assumptions. Experiments on three TCGA cohorts and two independent CPTAC cohorts demonstrate stable prognostic performance and favorable zero-shot cross-domain generalization compared with evaluated multimodal baselines. The source code is available at https://github.com/Cai-Zhiheng/Di-PACT.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5240_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Cai-Zhiheng/Di-PACT

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CaiZhi_DiPACT_MICCAI2026,
        author = { Cai, Zhiheng AND Li, Fulin AND Cai, Qing AND Zhang, Shugang},
        title = { { Di-PACT: Discrete Prognosis via Aligned Co-Topology of Pathways and Tumor Microenvironment } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes Di-PACT, a multimodal method for cancer survival prediction from whole-slide pathology images and RNA-seq data. The main idea is to improve both feature representation and cross-modal fusion. On the pathology side, the model uses distance-aware sampling and a bottleneck layer to compress high-dimensional UNI2 features while preserving tissue heterogeneity. On the genomic side, it builds pathway level features and uses bidirectional co-attention between pathways and WSI features. The model also includes a residual fusion design and uses a discrete-time survival objective instead of a standard Cox model. Experiments on three TCGA cohorts and two external CPTAC cohorts show improved performance over prior multimodal baselines, with additional support from ablation and interpretability analyses.

  • 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 paper addresses an important problem in computational pathology, how to use high-dimensional foundation model features for survival prediction in relatively small cohorts without strong overfitting. The motivation is clear and relevant.
    2. The overall method is well structured. The distance aware sampling and bottleneck compression are reasonable design choices to reduce feature dimension while keeping morphological heterogeneity. The bidirectional pathway-WSI interaction is also a meaningful extension beyond one way fusion.
    3. The three branch residual fusion design is a nice part of the method. It keeps WSI-specific, RNA-specific, and cross-modal information, which may help reduce modality dominance.
    4. The experimental evaluation is fairly comprehensive. The paper reports results on three TCGA cohorts and two external CPTAC cohorts, and compares against unimodal and multimodal baselines.
    5. The ablation study is useful. It shows that the performance gain does not come from only one module, but from several parts of the framework working together.
    6. The interpretability analysis is a positive aspect. The paper tries to link risk regions on the slide, pathway-level importance, and cross-modal attention, which is more informative than giving only a risk score.
  • 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 performance improvement is positive, but the margin over the strongest baseline is fairly small in several settings. In particular, some gains are very limited, so it is hard to tell how strong the advantage really is without stronger statistical support.
    2. Some important method details are still unclear. For example, the paper mentions adaptive fusion weights and a discrete-time survival setup, but does not explain these choices in enough detail for the reader to fully understand or reproduce the full pipeline.
    3. The external validation is a good addition, but it is not described in enough depth. The paper would be stronger if it reported more details about the external cohorts and preprocessing, so the generalization claim is easier to assess.
  • Please rate the clarity and organization of this paper

    Good

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

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

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

    N/A

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

    N/A

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

    (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 found the paper technically sound and well motivated. It addresses an important problem in multimodal survival prediction, and the overall method is well organized. The evaluation is fairly comprehensive, including multiple internal cohorts, external testing, and ablation studies, which strengthens the paper. My main concerns are that some performance gains are relatively small, and several method details are still unclear, especially in the fusion setup and external evaluation. In addition, the interpretability analysis is interesting but remains mostly qualitative. Overall, I view this paper as slightly above the acceptance threshold. The problem is important, the method is reasonable, and the empirical study is solid, but the paper would be stronger with clearer methodological details and stronger support for some of its claims.

  • 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 #2

  • Please describe the contribution of the paper

    The authors propose a distance-aware hierarchical sampling method and a direct bottleneck architecture to compress high-dimensional VFM features while preserving the spatial topological heterogeneity of the tumor microenvironment. They design a bidirectional pathway co-attention mechanism and a three-branch residual fusion strategy to address the limitations of unidirectional fusion, achieve dynamic cross-modal synergy, and prevent modality collapse. Furthermore, they develop a triple-view interpretability paradigm that provides explainable outputs at three levels: histopathological, transcriptomic, and cross-modal interaction. Additionally, discrete-time survival prediction is employed to overcome the limitations of the conventional Cox proportional hazards model.

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

    The problem orientation is clear, as the authors precisely identify two major challenges of VFM: overfitting and modality imbalance. The proposed technical solutions are highly targeted. The experimental design is comprehensive, encompassing three internal TCGA cohorts and two external CPTAC zero-shot validation sets, providing strong evidence of generalizability. The performance advantages are evident, with the model achieving the highest C-index across all internal and external cohorts, accompanied by smaller standard deviations, demonstrating notable stability. The interpretability is robust: the “Triple Visualization” paradigm constructs a complete chain of attribution evidence, ranging from macroscopic heatmaps to microscopic pathway contributions, offering significant clinical reference value. Finally, the ablation study is thorough, validating the independent contribution of each core component individually.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The core innovations are somewhat dispersed, as the method simultaneously incorporates multiple components, including sampling, bottleneck, masking, bidirectional attention, residual fusion, and discrete loss, making it difficult to identify a single most critical technical breakthrough. Additionally, some performance gains are marginal; for example, on TCGA-LUSC, the improvement over PAMT is only 0.003, which provides limited support for the strong claim of “significantly overcoming modality imbalance.” The external validation baseline comparison is also limited, as only three methods are compared on CPTAC, which is insufficient to fully justify the strong conclusion of “broad cross-domain generalization leadership.” Furthermore, the visualization lacks quantitative validation: the Triple Visualization paradigm is largely qualitative and does not assess the stability of its highlighted regions and pathway attributions or their consistency with pathological priors. Finally, the justification for hyperparameter choices is insufficient; key settings such as K = 40 and sampling numbers 15/5 lack sensitivity analysis or a clear rationale for their selection.

  • Please rate the clarity and organization of this paper

    Good

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

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

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

    N/A

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

    N/A

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

    (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 strengths of this work include precise problem identification, a comprehensive experimental design, strong performance, excellent interpretability, and a thorough ablation study. The weaknesses involve somewhat scattered innovations, limited performance gains in certain comparisons, a narrow set of baselines for external validation, a lack of quantitative support for the visualizations, and insufficient analysis of some hyperparameters. Therefore, we conclude that while the work is solid and has clear practical value, there remains room for improvement in distilling the core innovations and enhancing the rigor of certain arguments, placing it near the acceptance threshold.

  • 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’ rebuttal directly addresses the key concerns raised in my original review, particularly those related to dispersed innovations, marginal performance gains in some cohorts, limited external baselines, lack of quantitative validation for visualizations, and insufficient justification of hyperparameters.



Review #3

  • Please describe the contribution of the paper

    The paper proposes Di-PACT, a multimodal survival analysis framework that integrates whole-slide images (WSIs) and bulk RNA-seq data for cancer prognosis. On the histopathology side, the framework uses UNI2 features, combines distance-aware stratified sampling and a direct-bottleneck compression strategy. On the molecular side, transcriptomic data is encoded by using pathway-based masks derived from prior biological knowledge. The two modalities are fused through a bidirectional pathway co-attention mechanism combined with a three-branch residual fusion design. Survival prediction is formulated as a discrete-time classification task. The method is evaluated on three TCGA cohorts and two external CPTAC cohorts.

  • 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.

    Clinical relevance: This paper addresses an important and real problem in precision oncology, that is, using high-dimensional histopathology data and genomic data to perform multimodal survival prediction.

    Architectural design: The combination of pathway-guided molecular encoding, cross-modal co-attention, and residual fusion is conceptually well-structured, and reflects an attempt to address the modality imbalance problem in multimodal learning.

    Relatively broad evaluation: The framework is evaluated on three internal TCGA cohorts and two external CPTAC cohorts, this is a positive aspect and strengthens the empirical scope of the paper.

    Interpretability: The authors, through slide-level risk maps, cross-modal attention visualizations, and transcriptomic attribution, carry out an effort to provide multi-level interpretability, which is valuable in the clinical setting.

  • 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.

    Overstated topology-preservation claim: The paper claims that the proposed distance-aware stratified sampling method can effectively preserve the spatial topology of the tumor microenvironment. However, the method is based on K-means clustering in latent feature space, followed by the selection of nearest/farthest patches. This promotes feature-space diversity, but does not explicitly preserve 2D spatial adjacency or tissue architecture. As written, the topology-preservation claim is stronger than what the provided evidence is able to support.

    Missing baseline for bidirectionality: The bidirectional pathway co-attention mechanism is presented as a core contribution. However, the ablation study only removes the module entirely (w/o Bi-Attention), and does not compare the full bidirectional design against a unidirectional alternative. Without this direct comparison, the specific advantage that bidirectional attention holds over standard unidirectional querying is not convincingly demonstrated.

    Ambiguity in fusion weight calibration: The paper states that the fusion weights (alpha, beta, gamma) are adaptively calibrated per cohort, but does not provide a clear explanation of whether these weights are obtained through end-to-end learning, tuned on a validation set, or manually selected. This ambiguity affects reproducibility of the experiments, and if target-cohort-specific calibration is required, it also weakens the claim regarding zero-shot generalization.

    Insufficient survival evaluation metrics: The evaluation relies on C-index, time-dependent AUC, and Kaplan-Meier stratification as its metrics. Although these metrics hold certain practical value, if the paper is able to incorporate calibration-oriented survival evaluation metrics, e.g., Brier Score (BS) or Integrated Brier Score (IBS), its persuasiveness will be further enhanced, especially since the model predicts discrete-time survival probabilities.

    Interpretability claims are stronger than the validation: The proposed “Triple Visualization” paradigm is visually compelling, but the interpretability evidence remains largely qualitative and is based mainly on attention-style visualizations. The paper makes stronger claims about clinical interpretability than are currently supported. If expert pathological validation or a more clear biological validation protocol can be introduced, these claims will get further persuasiveness improvement.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

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

    N/A

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

    The paper addresses an important multimodal problem and presents a conceptually interesting architecture. To strengthen the work, I encourage the authors to add a direct comparison between unidirectional and bidirectional attention, as this is necessary to separately measure the contribution of the proposed co-attention design. I also encourage the authors to clarify exactly how the fusion weights (alpha, beta, gamma) are determined, since this has implications for both reproducibility and the zero-shot generalization claim. In addition, including calibration-oriented survival evaluation metrics, e.g., BS and IBS, will improve the rigor of the evaluation. Finally, I recommend softening the claims regarding spatial topology preservation and clinical interpretability, unless they can be supported by stronger empirical data or expert validation.

  • 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?

    I recommend a Weak Reject for this submission.

    The paper addresses a clinically important problem, and presents a reasonably well-structured multimodal framework with a broader-than-average evaluation across TCGA and CPTAC cohorts. These are meaningful strengths.

    However, I am not yet convinced that the central methodological claims are adequately supported by the current evidence. In particular, the paper frames feature-space K-means sampling as preserving spatial topology, but this is not convincingly demonstrated. The core bidirectional attention claim is also insufficiently isolated, since there is no direct unidirectional baseline. In addition, the cohort-specific calibration of the fusion weights is not clearly described, which raises questions regarding reproducibility and the strength of the zero-shot generalization claim. Besides, the interpretability claims are stronger than the level of validation currently provided.

    Overall, I find the paper promising, but I think the current version falls slightly below the acceptance threshold.

    I hope the authors can address these concerns in the rebuttal.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Author Feedback

[R1] Q1: Marginal gains lacking statistical support. A1: For limited gains(TCGA-LUSC), paired t-tests on 5-fold cross-validation prove Di-PACT’s improvement over PAMT is significant(p=0.034<0.05), confirming stable superiority. Q2: Missing details on fusion weights and discrete survival. A2: 1)Fusion Weights: Found via grid search(step 0.1)on TCGA validation set. Optimum α=0.2, β=0.4, γ=0.4 assigns dominance to genomic/cross-modal features, keeping minor pathology weight preventing modality collapse. During CPTAC testing, weights strictly frozen without target data, ensuring rigorous zero-shot evaluation. 2)Discrete Survival: Based on train-set quantiles(balancing interval sizes), survival partitioned into 5 non-overlapping intervals optimized via NLL loss. Bypasses Cox’s proportional hazards assumption, capturing non-linear dynamics. Details appended. Q3: Insufficient external validation and preprocessing details. A3: CPTAC zero-shot: 1)Independence: Filtered complete WSI/RNA-seq/survival intersections; LUAD&LSCC included 211&208 cases. 2)Preprocessing: External WSIs followed TCGA pipeline(Level3 HSV mask, Level0 256x256 patches to UNI2). Post-log2(x+1)transformation, external RNA-seq strictly used TCGA train mean/std for Z-score normalization. Test sets never calculated internal statistics, eliminating leakage. Q4: Interpretability analysis lacks quantitative validation. A4: Agreed heatmaps lack quantitative persuasion. Though early physician feedback confirmed risk/lesion consistency, pixel-level annotation is labor-intensive. Deleted overconfident wording, repositioning model as exploratory tool. Transcriptomic high-weight pathways(e.g.EMT, apoptosis)highly align with literature, validating molecular reliability. Large-scale prospective quantitative evaluation(e.g.IoU)is core Future Work.

[R2] Q1: Dispersed innovations. A1: Agreed; revision clarifies primary-secondary logic: 1)Mechanistic Breakthrough(Core): Bidirectional co-attention breaks unidirectional blind spots, solving modality dominance. 2)Engineering Infrastructure(Auxiliary): Clustering sampling, bottleneck layers&residual fusion are necessary infrastructure suppressing overfitting&modality collapse when applying VFM features to small cohorts. Q2: Marginal performance gains on specific datasets. A2: See R1-A1.Q3: Limited external baselines. A3: Completed more SOTA evaluations, omitted due to”no new experiments”rule. Claims softened to”showing robust cross-domain generalization potential”. Full tables open-sourced/appended. Q4: Visualizations lack quantitative validation. A4: See R1-A4.Q5: Hyperparameter rationale. A5: Clustering parameters(K=40, 15/5 ratio)found via early systematic experiments testing multiple combinations, optimally balancing TME heterogeneity preservation&efficiency.

[R3] Q1: Overclaimed spatial topology. A1: Agreed: latent K-Means promotes feature diversity, not explicit 2D adjacency. Claim downgraded to”preserving TME phenotypic heterogeneity&feature diversity”. Method clarifies: clustering sampling prevents discarding prognostic rare phenotypes during VFM compression, building representative feature set, not physical topology reconstruction. Q2: Missing unidirectional baseline. A2: Obeying”no new experiments”rule, extracted early unidirectional baseline from original logs. TCGA unidirectional performance: BLCA(0.718±0.044), LUAD(0.707±0.039), LUSC(0.694±0.036). Significantly inferior to Di-PACT, confirming bidirectional mechanism overcomes unidirectional modality dominance. Appended to Table 4.Q3: Fusion weight calibration. A3: See R1-A2.Q4: Lacking calibration metrics(BS/IBS). A4: To gauge clinical value, supplemented Brier Scores(BS/IBS)using original probabilities. Existing models suffer severe miscalibration: MCAT(IBS: 0.311), SurvPath(0.223), PAMT(0.175). Di-PACT achieved optimal calibration, lowering IBS to 0.118(1/3/5-year BS: 0.082/0.125/0.158). Tables appended. Q5: Interpretability validation. A5: See R1-A4.



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.

    All reviewers agree that the paper is technical sound and well motivated. The common concerns are around the limited performance gain and the insufficient clarity in methodological details. The authors should prioritize: (1) providing more details on important methodologies pointed out by R1 and R3.(2) discuss the marginal performance gain concern and core innovations concern by R1 and R2, respectively

  • 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 precision oncology survival analysis through a targeted multimodal framework that efficiently balances WSI and bulk RNA-seq processing, showing stable performance across three TCGA and two external validation datasets. The authors delivered an strong rebuttal that with addressed most concerns in parameter definitions, baseline ablations, and corrected overclaiming in the original version.



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 paper proposes a multimodal survival prediction framework that integrates WSI features and pathway-level RNA-seq representations through bidirectional co-attention and residual fusion. Reviewers appreciate the clinical relevance of the task. The rebuttal addresses several concerns such as methodological details, marginal performance gains, external evaluation setup, etc. The revision should incorporate these clarifications into the manuscript and carefully moderate claims that are not fully supported by the current evidence.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    After reviewing the comments and the authors’ responses, I find that the main concerns have been addressed. I therefore recommend accepting the submission, while encouraging the authors to carefully incorporate the reviewers’ suggestions in the final version, particularly those related to reproducibility and the strength of methodological claims .



back to top