List of Papers Browse by Subject Areas Author List
Abstract
Multimodal survival prediction using whole-slide images (WSIs) and transcriptomics has shown promise in cancer prognostication, yet remains fundamentally limited by two challenges: (1) substantial disease-irrelevant redundancy within each modality, and (2) unstable discriminative signals caused by intratumoral heterogeneity. Existing multimodal methods primarily focus on cross-modal interaction modeling but lack mechanisms to explicitly suppress redundancy and stabilize clinically meaningful representations. To address these challenges, we propose KERA, a Knowledge-Enhanced Retrieval-Augmented multimodal framework for survival prediction. KERA introduces a two-stage training strategy. First, a pan-cancer pretraining paradigm (ARPnet) performs modality alignment while explicitly retaining disease-relevant information under cancer-type supervision, enhancing informative representations while alleviating redundancy. Second, a retrieval-augmented fusion module (RAFnet) enriches patient representations by incorporating contextually similar intra- and inter-modality features, mitigating representation instability caused by sparse malignant patterns. Across five independent TCGA cohorts, KERA achieves a mean C-index of 0.716, yielding a 3.17% relative improvement over the strongest multimodal baseline, with statistically significant gains (Wilcoxon signed-rank test, p< 0.05). These results demonstrate that structured knowledge-enhanced pretraining combined with retrieval augmentation provides a principled solution to redundancy and heterogeneity in multimodal survival modeling. The code is available at https://github.com/AIMLab-UBC/KERA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1405_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AIMLab-UBC/KERA
Link to the Dataset(s)
TCGA dataset (WSIs and RNA-seq): https://portal.gdc.cancer.gov/
Disease-specific survival (DSS) data: https://xenabrowser.net/
BibTex
@InProceedings{ZhaZey_KnowledgeEnhanced_MICCAI2026,
author = { Zhang, Zeyu AND Azadi, Puria AND Maneshgar, Behnam AND Khajegili Mirabadi, Ali AND Alabi, Nikolay AND Farahani, Hossein AND Bashashati, Ali},
title = { { Knowledge-Enhanced Representation Learning with Retrieval-Augmented Multimodal Fusion for Survival Prediction } },
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 presents KERA, a two-stage multimodal framework that addresses feature redundancy and representation instability in cancer survival prediction.
- 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.
Principled Problem Framing: Explicitly targets redundancy and heterogeneity via disease-relevant retention loss and retrieval augmentation. Technically Novel Components: Combines pan-cancer pretraining with AdaSP loss and a computationally efficient retrieval-augmented fusion module.
- 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.Ambiguous Cross-Validation Isolation: The construction of the retrieval database during fine-tuning is not clearly described. The authors must explicitly confirm that retrieval indices were computed within each cross-validation fold using only training split samples to prevent data leakage. 2.Unclear Pretraining Data Split: The relationship between the 24 pretraining cancer types and the 5 downstream cohorts is ambiguous. The authors should clearly state that downstream cancer types were strictly excluded from the pan-cancer pretraining set. 3.Overly Complex Figure 1: The overview figure attempts to convey both the two-stage training paradigm and the retrieval index precomputation workflow, resulting in visual clutter that obscures the core pipeline. Simplifying the diagram or separating it into two focused subfigures would improve clarity. 4.Limited Qualitative Analysis of Retrieved Samples: The paper lacks any qualitative examination of the retrieved multimodal features. Providing a case study showing representative retrieved patches or gene profiles would strengthen confidence that the retrieval mechanism captures clinically meaningful similarities rather than superficial feature correlations.
- 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
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?
The approach is well-motivated and demonstrates consistent improvements over strong baselines across five TCGA cohorts. However, the manuscript lacks critical details regarding data isolation protocols and would benefit from improved presentation clarity.
- 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
KERA utilizes a two-tiered pipeline: ARPnet to learn modality-universal and cancer-related representations through alignment and retention losses during pan-cancer pre-training and RAFnet to augment patient representation through similar features stored in an external pre-computed data collection. The problem statement and rationale for their work are obvious, and the experiment design is robust.
- 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 two-stage design maps cleanly onto two well-identified problems: ARPnet addresses feature redundancy via structured pan-cancer pretraining, and RAFnet addresses representation instability via retrieval augmentation.
2.Applying retrieval augmentation to survival prediction is a relatively fresh idea. Retrieving similar patients from a database to enrich weak or noisy signals is a practical and well-motivated design.
3.The evaluation covers five TCGA cancer cohorts provides reasonable evidence of generalizability. The baseline comparison is also comprehensive.
- 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 paper claims survival prediction fails not because cross-modal interaction is insufficiently modeled, but because the underlying representations are unstable and overly redundant. However this is not been proven. There is no direct measurement of redundancy or instability and the ablation only shows modules help, not why they help.
2.In Table 2, removing pretraining actually improves BRCA performance (0.723 vs. 0.701), which suggests negative transfer on some datasets. This is not discussed in the paper.
3.The pretraining uses 24 cancer types, but I am curious how the number of cancer types affects performance, and what criteria were used to select these 24 types over others.
- 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
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?
Well-motivated two-stage design, rigorous experimental protocol with statistical tests, and thorough ablations. Main concerns are the unverified core claim about redundancy/instability
- 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
The paper presents a two-stage multimodal framework for cancer survival prediction that combines representation pretraining and retrieval-augmented fusion. It first learns modality-common and disease-relevant features to reduce redundancy and improve alignment between pathology and genomic data, and then enhances patient representations by retrieving and integrating similar samples across and within modalities. Experiments on TCGA cohorts show consistent performance improvements over existing methods.
- 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 main strengths of the paper lie in its clear motivation and well-structured design, addressing practical challenges such as feature redundancy and data heterogeneity in multimodal survival prediction. The proposed framework is coherent and reasonably justified, integrating pretraining and retrieval in a unified pipeline. In addition, the experimental evaluation is fairly comprehensive, with comparisons across multiple cohorts and ablation studies that help validate the contribution of each component.
- 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 main weaknesses lie in the limited methodological novelty and the incremental nature of the improvements. Most components, including contrastive alignment, metric learning, and retrieval-based fusion, are adapted from existing techniques rather than fundamentally new contributions, making the overall framework closer to a structured combination of known modules.
- 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.
(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?
Despite the limited methodological novelty, the paper addresses a relevant clinical problem with a clear and well-motivated framework. The proposed design is coherent and integrates representation learning and retrieval in a principled way, rather than as an ad hoc combination. In addition, the method demonstrates consistent, statistically significant improvements over multiple strong baselines across several cohorts (e.g., achieving the best overall C-index of 0.716 ), supported by comprehensive experiments and ablation studies. Given its practical relevance, solid empirical validation, and reasonable technical soundness, the work merits a weak accept.
- 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
Author Feedback
We sincerely appreciate the reviewers and the Area Chair for their valuable feedback and constructive suggestions. We will carefully revise the manuscript to address the reviewers’ comments prior to publication. Below, we provide detailed responses to the concerns raised.
Reviewer#1
(Q1: Cross-Validation Isolation.) A1: We appreciate the reviewer’s observation regarding retrieval database construction. To clarify: the retrieval database was constructed independently within each cross-validation fold, using only samples from that training split, to prevent data leakage. We will add this clarification to the revised manuscript (Section 2.2).
(Q2: Pretraining Data Split.) A2: The five cancer types used for downstream fine-tuning were strictly excluded from the 24 cancer types for pan-cancer pretraining, ensuring the model had not seen any downstream task data before fine-tuning. We will clarify this in the paper.
(Q3: Overly Complex Diagram.) A3: We have simplified Figure 1 to better highlight the core pipeline and improve overall clarity.
(Q4: Limited Qualitative Analysis of Retrieved Samples.) A4: We thank the reviewer for this insightful suggestion. We agree that qualitative analysis of retrieved samples would further strengthen the interpretability of the retrieval mechanism. Due to the page limit, we will conduct a more detailed qualitative analysis in future work.
Reviewer#2
(Q1: Empirical Evidence for Redundancy and Instability.) A1: We thank the reviewer for this important point. We agree that our current framing overstates what is empirically demonstrated, and we will soften the corresponding statement in the introduction. While our ablations do not directly measure redundancy or instability, they offer indirect support: removing L_retent drops the mean C-index from 0.716 to 0.647 and removing the retrieval module drops it to 0.677, consistent with our hypothesis that explicit suppression of redundancy and stabilization of representations each contribute independently. We will explore direct quantification schemes in future work.
(Q2: Performance on BRCA.) A2: We thank the reviewer for raising this. TCGA-BRCA is the largest downstream cohort (n=983) and was strictly excluded from pan-cancer pretraining to prevent data leakage; representations learned from the other cancer types may therefore not fully generalize to BRCA-specific characteristics, which we believe explains the marginal w/o Pre-T result. Two observations argue against negative transfer in any clinically meaningful sense: (i) the full KERA model achieves lower variance on BRCA (±0.043 vs. ±0.064 for w/o Pre-T), suggesting more stable estimation, and (ii) on BRCA, KERA still outperforms every multimodal baseline (0.701 vs. MCAT 0.621, MOTcat 0.675, PathOmics 0.667). We will add this discussion to the revised manuscript.
(Q3: Cancer Types for Pretraining.) A3: Pan-cancer pretraining aims to learn contextual representations from diverse cancer types, therefore incorporating more cancer types could generally improve model generalizability. Additionally, TCGA contains 33 cancer types in total. We selected GBM/LGG, BLCA, BRCA, UCEC, and LUAD (also used in MOTCat) for fine-tuning, while 24 remaining cohorts were used for pan-cancer pretraining. Other cohorts were excluded due to modality availability or data quality constraints.
Reviewer#3
(Q1: Limited methodological novelty.) A1: We thank the reviewer for recognizing KERA as a clear, well-motivated framework. We agree that several of KERA’s building blocks draw on established techniques. However, we respectfully suggest that our contribution is not the components themselves but their integration into a unified framework explicitly designed to address two underexplored failure modes in multimodal survival modeling: information redundancy and representation instability.
We once again sincerely thank the reviewers and the Area Chair for their meticulous work and constructive suggestions.
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 reviewers are generally positive, with one Accept and two Weak Accept recommendations, and they recognize the paper’s clear motivation for addressing redundancy, heterogeneity, and representation instability in multimodal survival prediction. The two-stage design is coherent: ARPnet targets disease-relevant representation learning through pan-cancer pretraining, while RAFnet introduces retrieval-augmented fusion to improve patient-level representation stability. The experiments cover five TCGA cohorts, include comparisons with strong baselines, and report consistent improvements supported by ablation studies and statistical testing. The main issues to address in the final version are the data isolation protocol for retrieval database construction within cross-validation folds, whether downstream cancer types were excluded from pan-cancer pretraining, and qualitative analysis of retrieved samples. Overall, despite moderate methodological novelty, the practical relevance, coherent design, and solid empirical validation support acceptance.