acceptodds
Withdrawn from ICLR 2027

MSSF-Surv: Multimodal State-Space Fusion for Cancer Survival Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate cancer survival prediction informs patient risk stratification and individualized treatment planning. Histopathology and multi-omics provide complementary prognostic evidence, but their integration can be limited by the loss of tissue context during patch aggregation and insufficient modeling of cross-modal interactions. We propose Multimodal State Space Fusion (MSSF-Surv), a framework that preserves spatial organization before conditioning pathology readout on molecular profiles. Its dual-granularity design captures tissue architecture alongside cellular morphology, while multidirectional state-space modeling handles long sequences of fine-grained whole-slide image patches with linear complexity. Molecular queries then retrieve information from these spatially contextualized representations, rather than from isolated patch features. A shared state-space model combines the retrieved features with omics representations through modality-specific gates, linking tissue context to molecular information for survival prediction. Extensive experiments using five-fold cross-validation on six TCGA cancer cohorts demonstrate the effectiveness of MSSF-Surv. Survival analyses and ablations support its predictive performance and the contributions of dual-granularity encoding and omics-guided fusion.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multimodal state space fusion, Whole-slide images, Mamba, Survival prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_jqPyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Existing methods for multi-modal survival prediction may suffer from the loss of tissue context during patch aggregation and insufficient modeling of cross-modal interactions. To address this problem, the authors proposed Multimodal State Space Fusion (MSSF-Surv), a framework that preserves spatial organization before conditioning pathology readout on molecular profiles. In MSSF-Surv, the dual-granularity design can capture tissue architecture alongside callular morphology, while multidirectional state-space modeling can address long sequences of fine-graned WSI patches with linear complexity. Molecular queries can retrieve information from these spatially contextualized representations, rather than from isolated patch features. A shared statespace model combines the retrieved features with omics representations through modality-specific gates, linking tissue context to molecular information for survival prediction.

Experimental results on 6 TCGA cancer cohorts shows the effectiveness of the proposed methods.

Critical strengths
  1. Instead of just concatenating sequences or computing cross-modal attention directly across tens of thousands of visual tokens, the framework uses the omics profile as queries to compress pathology features. This keeps the sequence length strictly bounded for the final multimodal fusion stage, ensuring high efficiency and a lightweight design.
Critical weaknesses
  1. The scanning method is a direct plug-and-play from standard 2D Vision Mambas. The authors do not incorporate domain-specific histopathological spatial priors—such as tissue-cluster-aware ordering, spiral trajectories, or morphological topology-guided scanning paths. Consequently, the architecture primarily combines pre-existing building blocks with limited foundational novelty for computational pathology.

  2. While motivated by avoiding the O(N2)O(N^2) Transformer bottleneck, the paper lacks hardware profiling metrics—such as peak GPU VRAM, FLOPs, training time, and inference throughput (seconds per slide).

Questions for authors

Please provide concrete hardware profiling metrics comparing MSSF-Surv against standard MIL baselines, specifically reporting: (1) peak GPU VRAM usage during training/inference, (2) FLOPs, and (3) inference latency per WSI (seconds per slide). I am willing to consider raising my score if the authors can convincingly address the following concerns during the rebuttal:

Reviewer_s17pOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents MSSF-Surv, a framework that combines pathology images at two scales with multi-omics data to predict cancer survival risk. Coarse and fine pathology patches are encoded with a Transformer and a multidirectional Mamba, respectively. RNA expression, miRNA expression, and DNA methylation are each projected into one token. An Omics-Transformer processes these tokens, which then serve as queries to aggregate features from both pathology branches. These tokens are processed by an Omics-Transformer and used as queries to aggregate pathology features, followed by shared state-space fusion with separate gates. The authors evaluate the model using five-fold cross-validation on six TCGA cohorts and report the highest mean C-index among the compared methods on five cohorts.

Critical strengths
  1. Using Mamba for the long sequence of fine-grained pathology patches is a reasonable choice, given the computational cost of full self-attention.
  2. Combining coarse and fine pathology features is well motivated. The dual-scale model performs better than either scale alone across all six cohorts, supporting the use of information at both scales.
Critical weaknesses
  1. The survival endpoint is not specified. Section 3 defines (t_i) as "the observed time to an event or censoring," but does not explain whether the event is death, progression, or something else. The clinical label source, time origin, and censoring rules are also missing.
  2. The experiments do not separate the value of additional omics data from that of omics-guided pathology aggregation. Table 2 tests input combinations, while Figure 6 shows gains of only 0.001 - 0.006 over fine-grained pathology queries on five cohorts and worse performance on STAD. A comparison with independent pathology pooling followed by simple fusion, using the same inputs and encoders, is missing.
  3. The final fusion module operates on only three tokens (§3.3). The long-sequence motivation for Mamba in the fine pathology branch does not explain its use here. Without comparisons against a gated MLP or attention, the added value of shared state-space fusion remains unclear.
  4. The claim that the framework "preserves spatial organization" (Abstract) is not directly tested. Combining two scales improves prediction, but this alone does not show that spatial arrangement drives the gain. The paper does not describe positional encoding in the coarse Transformer or include a control that disrupts spatial ordering while retaining the same patch features.
Questions for authors
  1. Which survival endpoint is used in each cohort? Please specify the clinical label source, time origin, and censoring rules.
  2. How does omics-guided aggregation compare with independent pathology pooling followed by simple fusion, using the same inputs and encoders?
  3. How does the three-token SSM fusion compare with a gated MLP or attention-based fusion?
  4. How is positional information incorporated into the coarse pathology Transformer? Does disrupting the spatial ordering of fine patches reduce performance?
Reviewer_SA1sOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MSSF-Surv, a multimodal framework for cancer survival prediction using whole-slide pathology images and multi-omics data. The method combines dual-granularity pathology encoding, four-directional Mamba scanning, omics-guided cross-attention, and a shared Mamba-based fusion module. Experiments on six TCGA cohorts show competitive C-index results, supported by several ablation studies.

Critical strengths

The paper addresses a meaningful problem in multimodal cancer survival prediction, and the experimental results are competitive across six cancer cohorts

Critical weaknesses
  1. The proposed architecture appears to combine several established components, with rather limited incremental contributions. Four-directional scanning is already widely used in Vision Mamba [1], while closely related methods such as SurvMamba [2] and SAMamba [3] have explored Mamba-based multimodal survival prediction. Although the authors also discuss other Mamba-based approaches in Related Work, they neither clearly distinguish their contributions from these existing methods nor include these Mamba-based baselines in their experiments.
  2. Cross-attention has already established interactions between omics and pathology features, after which the model simply applies a shared Mamba block to a sequence of only three tokens. What additional cross-modal modeling capability does this Mamba block provide? Why is a state-space model preferable to simpler MLP- or attention-based fusion for such a short sequence? The paper lacks direct comparisons to justify this design.
  3. Section 4.1 describes differential expression analysis and Random Survival Forest-based feature selection, but does not clarify whether these steps are performed on the entire dataset or independently within the training split of each cross-validation fold. Since feature selection uses survival outcomes, performing it before data splitting would introduce label leakage and compromise the validity of the reported results.
  4. Interestingly, all six overlapping baseline results in Table 1, including both means and standard deviations, are identical to those reported in Table 1 of POMP[4]. However, the authors do not include POMP itself in the comparison, despite citing it in the references. POMP reports better results than MSSF-Surv on STAD, LUAD, and LGG, which substantially weakens the claimed state-of-the-art performance. The authors should clarify the source of the baseline results and explain why POMP was excluded.

[1] Liu Y, Tian Y, Zhao Y, et al. Vmamba: Visual state space model[J]. Advances in neural information processing systems, 2024, 37: 103031-103063.

[2] Chen, Ying, et al. "Survmamba: State space model with multi-grained multi-modal interaction for survival prediction." 2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2025.

[3] Zhang, Wei, et al. "Samamba: Integrating state space model for enhanced multi-modal survival analysis." 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2024.

[4] Suixue Wang, Shilin Zhang, Huiyuan Lai, Weiliang Huo, and Qingchen Zhang. Pomp: Pathologyomics multimodal pre-training framework for cancer survival prediction. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pp. 7813–7821, 2025.

Questions for authors
  1. What is the key methodological novelty over existing Mamba-based multimodal survival prediction methods?
  2. Can the authors compare the proposed Cross-Mamba Fusion against simpler alternatives, such as MLP, gated fusion, and Transformer-based fusion, under comparable settings?
  3. Is the entire preprocessing pipeline, especially survival-based feature selection, fitted exclusively on training data within each cross-validation fold?
  4. Were the baseline results in Table 1 reproduced by the authors or taken from POMP? Why was POMP itself not included in the comparison?

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

Related papers

Discussion (0)

Sign in to comment.