acceptodds
Withdrawn from ICLR 2027

DualRE-Surv: Dual-Path Complementary Fusion of Representation and Evidence for Multimodal Cancer Survival Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate survival prediction is vital for clinical decision. Fusing whole slide images (WSIs) and genomics data has become a key research direction, as they respectively capture the histomorphological and molecular characteristics of tumors. Existing methods typically employ feature concatenation, attention mechanisms, or cross-modal interactive learning for joint representation; however, they rarely explicitly preserve the prognostic evidence provided by different modalities at the decision level. When pathology and genomics yield inconsistent judgments regarding a patient's survival risk, modal conflicts may be propagated to the final prediction via the joint representation. To alleviate this issue, we propose DualRE-Surv, a dual-path complementary fusion of representation and evidence for multimodal cancer survival prediction. In addition to the representation path for learning pathology-genomics joint representations, we introduce a parallel evidence-aware path. This path employs Dirichlet Vacuity Estimation (DVE) to measure the sufficiency of modal evidence, utilizes Consensus-Complementarity Evidence Decomposition (CCED) to decouple cross-modal consensus evidence and complementary evidence, and enhances the discriminability of consensus evidence across different time intervals through Consensus Evidence Enhancement (CEE). Subsequently, Joint Evidential Survival (JES) integrates the two types of evidence and constructs the survival curve of the evidence path. The Representation-Evidence Survival Mixture Module further integrates the prediction results of the two paths, enabling the model to simultaneously utilize the cross-modal joint representation and the conflict-adjusted modal decision evidence. We evaluate the proposed method on eight TCGA cancer cohorts using five-fold cross-validation. DualRE-Surv achieves C-indices of 0.6477, 0.7192, 0.7318, 0.8114, 0.8120, 0.6104, 0.6830, and 0.6570 on HNSC, BLCA, BRCA, UCEC, KIRC, LUSC, SKCM, and STAD, respectively, outperforming existing methods. Furthermore, extensive validation experiments further verify the effectiveness of the overall DualRE-Surv framework and its key modules.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Survival Prediction, Computational Pathology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8WZTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DualRE-Surv for multimodal cancer survival prediction. It combines representation-level fusion with an evidence-aware path that decomposes modality evidence into consensus and complementary components. Experiments on eight TCGA cohorts show improved C-index over existing methods.

Critical strengths
  1. The paper addresses an interesting problem of conflicting predictions across modalities.
  2. The experiments are relatively comprehensive, including eight datasets, ablations, and robustness analyses.
Critical weaknesses

My main concern is the methodological soundness. The method directly converts survival logits into Dirichlet evidence and interprets the resulting vacuity as predictive uncertainty. However, survival risk and predictive confidence are fundamentally different concepts. Unlike standard EDL, the method does not use an explicit evidential learning objective or validate uncertainty calibration. This questionable assumption undermines the foundation of the entire evidence-aware fusion mechanism.

Questions for authors
  1. How can survival logits be interpreted as evidence of predictive confidence? If the unimodal logits represent conditional hazards, then the proposed vacuity satisfies: Vm=KK−log⁡Sm(K)V_m=\frac{K}{K-\log S_m(K)}. This makes vacuity a deterministic function of predicted survival probability rather than an independent uncertainty estimate. Could the authors clarify the unimodal logits and justify this design, preferably with uncertainty calibration experiments?
  2. Related prior work (M2EF-NNs[1]) has already explored Dirichlet-based uncertainty estimation and decision-level evidence fusion for multimodal cancer survival prediction. Although it is briefly mentioned in Related Work, the paper does not clearly explain how the proposed evidence fusion differs from or improves upon this closely related method. A direct experimental comparison is also missing.

[1]Luo, Hui, et al. "Multimodal multi-instance evidence fusion neural networks for cancer survival prediction." Scientific Reports 15.1 (2025): 10470.

Reviewer_zghFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DualRE-Surv adds a decision-level evidence path to a slot-based pathology-genomics representation model. It separates elementwise consensus from residual evidence, sharpens consensus, scales residuals using Dirichlet vacuity, and mixes the resulting survival curve with the representation-based curve through a learned scalar. Eight-cohort TCGA experiments report the highest mean C-index among listed methods, with component ablations and controlled perturbation analyses.

Critical strengths

S1. Table 1 reports the highest mean C-index on all eight cohorts, with identical cross-validation splits for the compared methods. Appendices G-H provide fold variability and component ablations, making the scope of the empirical gains inspectable.

Critical weaknesses

W1. Smaller perturbation-induced drift alone can reflect insensitivity, so predictive accuracy under corruption should also be measured.

W2. The strongest-baseline C-index margins range from 0.0028 to 0.0139 in Tables 4-5, against appreciable fold variability. Paired differences, repeated splits, and calibration metrics are needed before these rankings support reliable superiority in survival prediction.

Questions for authors

Q1. Does K include a tail interval beyond the prediction horizon? How are the last-interval censored observations handled when evidence-path survival is exactly zero, and what values are used for K, interval boundaries, and the censoring-loss weight?

Q2. How are the unimodal survival curves Sw and So computed from their logits? Please reconcile their training likelihood with the event-time interpretation used by the evidence path, ideally with a concrete two-interval example.

Q3. Does the full model retain C-index and calibrated survival probabilities under corruption, rather than merely changing its predictions less? How does it compare with matched reliability-aware fusion and a simple learned mixture of unimodal survival curves?

Reviewer_XkZROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DualRE-Surv for survival prediction from pathology whole-slide images and genomics. Building on a SlotSPE-inspired representation pathway, it introduces an Evidence-Aware Path that separates consensus and complementary evidence, modifies their contributions, and constructs an additional survival curve. The two curves are combined through a learned mixture. Experiments on eight TCGA cohorts show consistently higher mean C-indices than the evaluated baselines, with limited additional computational cost.

Critical strengths
  1. Addressing disagreement between pathology and genomics is a worthwhile problem. Explicit decision-level modeling offers a plausible complement to learning joint representations.

  2. The paper evaluted 8 cancer cohorts against 17 unimodal and multimodal baselines, including recent methods such as LD-CVAE and SlotSPE. The consistent improvement in mean C-index across all eight cohorts is encouraging. Common five-fold splits and the reported standard deviations improve the usefulness of these comparisons.

3.Positive ablation results and low overhead. The reported training and inference costs remain close to SlotSPE, suggesting that the proposed addition is computationally economical.

Critical weaknesses
  1. The novelty over existing representation and evidential fusion approaches is insufficiently established. The representation pathway explicitly draws on SlotSPE-style slot attention, selection, and decoding. Meanwhile, evidential learning for pathology–genomics survival prediction is already established. The potentially novel contribution is narrower than the general introduction of a representation-and-evidence framework.

  2. The empirical advantage over strong baselines is modest and its reliability and attribution remain unclear. The absolute C-index improvements over the strongest competing method in each cohort were about ~1%, paired uncertainty estimates are needed to determine how reliably these differences distinguish the proposed method from strong alternatives.

  3. The probabilistic formulation requires clarification at important points, as in eq 26 etc.

Questions for authors

1.What is the precise advance over existing evidential survival fusion? In particular, which limitation of M2EF-NNs, UMSA, or related fusion rules motivates the proposed consensus–complementarity construction? Which claimed benefit specifically requires this construction rather than an established evidence-fusion rule?

2.Do the existing ablations isolate that advance? Please define the replacement computations in the “w/o DVE,” “w/o CCED,” and “w/o CEE” settings. Can the existing comparisons establish an advantage over a simple fusion alternative with the same backbone, auxiliary supervision, and final mixture?

3.How is the survival formulation implemented? Please define the mappings from unimodal logits to Sw,So, the time intervals, and the treatment of final-interval censoring when computing evidence-path NLL.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.