List of Papers Browse by Subject Areas Author List
Abstract
In clinical oncology, biomarker status prediction and survival risk estimation are closely related but distinct objectives. While joint multi-task learning benefits from exploiting correlations among tasks, most existing methods remain unimodal and neglect task-specific heterogeneity, limiting joint modeling of local biomarker cues and global survival patterns using multimodal clinical context. To address these limitations, we propose PathoSlot, a concept-conditioned slot factorization framework that factorizes shared multimodal evidence into clinically anchored slot representations for joint biomarker and survival prediction. Specifically, PathoSlot factorizes fused features from whole slide images (WSIs) and pathology reports into a set of task-aligned slots through structured slot factorization. The factorized slots enable hierarchical prediction, with biomarker status predicted from slots encoding local diagnostic morphology and survival risk predicted from the slot capturing global prognostic patterns. To stabilize slot–task alignment, we further anchor each slot to its corresponding clinical prior via concept conditioning, constraining inter-task interference and enhancing interpretability and robustness across centers. Extensive experiments on internal and external cohorts demonstrate that PathoSlot consistently outperforms state-of-the-art methods on both biomarker and survival prediction tasks. Code is available at https://github.com/jtneuron/PathoSlot.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1567_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jtneuron/PathoSlot
Link to the Dataset(s)
TCGA-BRCA: https://portal.gdc.cancer.gov/projects/TCGA-BRCA
BibTex
@InProceedings{HuaZuq_Joint_MICCAI2026,
author = { Huang, Zuqi AND Mao, Liang AND Li, Wentao AND Dong, Tianxiang AND Tian, Mengxin AND Tang, Zhaoqing AND Li, Zhongyu},
title = { { Joint Biomarker and Survival Prediction via Concept-Conditioned Multimodal Slot Factorization } },
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
PathoSlot introduces a concept-conditioned slot factorization framework that jointly models whole-slide images and pathology reports for biomarker and survival prediction. Its core innovations include: a competitive reweighting mechanism to emphasize information-dense cues from both modalities; a soft factorization into task-aligned latent slots that enforce biomarker-oriented slots to focus on local diagnostic morphology and survival-oriented slots to capture global prognostic risk patterns; concept conditioning that anchors each slot to clinical priors for improved interpretability and cross-center generalization; and a context-gated residual update that treats the survival slot as a control variable, selectively injecting global multimodal context under patient-level survival supervision. Together, these design elements enable superior performance over state-of-the-art methods in joint biomarker and 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.
a novel application.
- 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.
Application is not novel.
- 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.
(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?
Lack of methodological innovation
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes PathoSlot, a multimodal framework for joint biomarker prediction and survival modeling from WSIs and pathology reports. The method first extracts multimodal tokens from TITAN, applies token reweighting, and then factorizes the fused representation into concept-conditioned, task-aligned slots. Each biomarker is predicted from its corresponding slot, while survival is predicted from a dedicated survival slot augmented with a context-gated residual. Experiments are conducted on TCGA-BRCA, GC-Internal, and GC-External, and the paper reports strong internal performance and competitive external results.
- 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 is important and clinically relevant. The paper is well motivated by the different granularity of biomarker prediction and survival modeling. Empirically, the method performs strongly on internal datasets and remains competitive on the external cohort.
- 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.
My main concern is conceptual. The method uses one slot per task, explicitly treats each clinical target as an entity, and sets the number of slots equal to the number of tasks. At the same time, it uses slot-wise softmax so that slots “competitively explain” the tokens. However, ER/PR/HER2/MMR and survival are not naturally distinct latent factors, and the evidence supporting these predictions is likely to be substantially shared. In this setting, slot-wise softmax is a strong inductive bias that may force shared evidence into artificial task-specific partitions. As a result, the method looks closer to task routing/bottlenecking than genuine factorization.
This issue is clearer when compared with recent slot-based work[1], which uses slots to model sparse, patient-specific latent prognostic events, rather than assigning one predefined slot to each supervised endpoint. The authors should explicitly discuss and compare the difference between their task-aligned slot design and event-centric slot designs such as SlotSPE[1], and justify why competitive one-slot-per-task decomposition is the right inductive bias here.
Besides, survival is not inherently a peer concept to biomarkers; it is a downstream prognostic outcome that integrates multiple shared factors. The fact that the survival slot is further given a special global residual branch in the paper also suggests that it does not fit the same competition-based slot assumption as the biomarker slots, which is internally inconsistent.
[1]Zhang, Y., Nanbo, L., Yang, C., Schmidhuber, J., & Gao, X. (2025). Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis. arXiv preprint arXiv:2512.01116.
- 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
Other questions: 1.The paper uses concept-conditioned initialization and also a cosine loss to keep slots aligned with concepts. This is plausible in practice, but the paper does not show whether both are necessary. An ablation with init only / cosine only / both is needed. Otherwise, it is unclear what actually stabilizes slot semantics. 2.The proposed method starts from TITAN pre-pooling multimodal tokens and then adds token reweighting and slot attention. However, several baselines use patch features + ABMIL: (a) the proposed slot representations, or (b) the use of richer pre-pooling multimodal tokens. Please include stronger matched-backbone comparisons, e.g., TITAN tokens + simple concatenation, and Patch features + slots attention.
- 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?
See the major weaknesses of the paper.
- 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 rebuttal addresses several experimental concerns and partially clarifies the conceptual issue. I remain unconvinced that the method performs genuine factorization rather than task-conditioned routing, but this concern can likely be addressed by revising the claims and explicitly acknowledging the methodological limitation. Given the strong empirical results, external validation, and the authors’ commitment to adding the requested ablations and clarifications, I lean toward borderline accept.
Review #3
- Please describe the contribution of the paper
This paper presents PathoSlot, a multimodal framework for joint biomarker prediction and survival prediction from paired whole-slide images (WSIs) and pathology reports. The method factorizes reweighted multimodal tokens into task-aligned slots, with separate slots intended to capture biomarker-related and survival-related information within a shared latent space. To encourage stable slot identity, the authors introduce concept-conditioned slot initialization, where each slot is anchored by a short task-specific clinical description. A context-gated residual is further used to inject global multimodal context into the survival slot. Experiments on TCGA-BRCA, GC-Internal, and GC-External show improved performance over a broad range of WSI-only and WSI-report baselines for both biomarker classification and survival prediction, and the ablation studies support the contribution of the main components.
- 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 paper studies an important and clinically relevant joint setting. The motivation that biomarker prediction and survival prediction are related but require different levels of information granularity is well explained and provides a good rationale for the proposed design. The overall method is clear and reasonably well motivated. The idea of using task-aligned slots to separate local biomarker-related cues from more global prognostic information is intuitive, and the concept conditioning is a sensible way to stabilize slot semantics. The multimodal formulation is practical. Combining WSIs and pathology reports is a natural choice for this setting, especially for survival prediction where report-derived context can be helpful beyond image-only modeling. The empirical evaluation is fairly comprehensive for a conference submission. The paper includes internal and external validation, compares against a broad set of baselines, and provides ablations for the main architectural components. The external evaluation on GC-External is particularly valuable. The paper also provides some qualitative support for the intended slot specialization through the reported attention examples, which is a useful addition.
- 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 claim that slot factorization explicitly disentangles task-specific representations is plausible, but the current evidence is still somewhat indirect. The experimental results support the usefulness of the design, though the paper provides limited direct analysis of how consistently the learned slots correspond to the intended task semantics across settings. The performance gains are encouraging and generally consistent, but they are not always large across all tasks and datasets. In addition, the paper does not report statistical significance analysis, which would make the empirical claims easier to assess. The baseline comparison is broad and valuable, but a little more detail on how the different baselines were adapted and tuned within the unified evaluation setup would further strengthen confidence in the fairness of the comparison. The pathology-report rewriting step is interesting and potentially useful, but its reproducibility is not fully clear from the current description. Additional detail on the prompting and generation procedure would make this part of the pipeline easier to reproduce. Overall, the strongest contribution of the work is the complete formulation and its application to this joint multimodal setting. Some of the individual components are relatively standard, so the novelty appears more in the overall integration than in a single sharply distinct methodological advance.
- 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?
Potential concern: the manuscript uses private patient cohorts and pathology-report processing, but I could not identify an explicit IRB/ethics statement in the paper. This should be clarified.
- 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 gave this paper a weak accept because I found the overall idea interesting, relevant, and reasonably well executed. The paper tackles a clinically meaningful joint setting, namely biomarker prediction together with survival prediction, and the motivation that these two tasks require different levels of information is well presented. The proposed slot-based formulation is clear, and the use of concept-conditioned slots to encourage task-specific representations is a sensible design choice. I also appreciated that the work goes beyond image-only modeling by incorporating pathology reports, which seems particularly relevant for survival prediction.
Another positive factor is the experimental section. The paper includes comparisons against a broad set of baselines, reports both internal and external validation, and provides ablation studies for the main components. The external validation on an independent cohort is especially valuable and strengthens the practical relevance of the work. Overall, I found the paper technically sound and appropriate for the MICCAI audience.
My score is not higher because I still have a few reservations. While the results are encouraging, the improvements are not always large across all tasks and datasets, and the paper does not include statistical significance analysis. In addition, the central claim that the proposed slot factorization meaningfully decouples task-specific information is plausible, but the current evidence is still somewhat indirect. I also think the reproducibility of the report rewriting step could be described more clearly. Overall, I found this to be a good paper with solid motivation and evaluation, but with a few remaining questions that keep me at weak accept rather than accept.
- 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
We thank the Reviewers for their valuable feedback. They highlighted PathoSlot’s strong motivation (Meta-reviewer#1), clinical relevance (R2, R3), and comprehensive evaluation with external validation (R3). We address all concerns below. Ethics/IRB (Meta-reviewer#1, R3) All private cohorts had IRB approval. Data were de-identified. We will add an IRB/ethics statement in camera-ready. Reproducibility (R1, R3) Code will be released, including report rewriting prompts, target format, retained/removed fields, decoding/filtering details. Slot factorization (R2) 1.Following DyMON (NeurIPS’21) and Slot-VAE (ICML’23), “factorization” denotes structured evidence summarization, not causal disentanglement or independent clinical factors. 2.Slot-wise softmax gives soft, non-exclusive competition: it avoids collapse by making endpoints weight shared WSI-report tokens differently, while all slots access the same token pool. 3.Routing/bottlenecking selects or compresses task-specific information paths, whereas PathoSlot keeps evidence shared and only induces endpoint-specific soft weighting. SlotSPE (R2) 1.Event discovery vs. endpoint prediction. SlotSPE uses slots to discover sparse patient-specific prognostic events for survival modeling, whereas PathoSlot uses endpoint-conditioned slots for supervised prediction of observed biomarkers and survival. We will add this distinction in the revision. 2.Why one-slot-per-task fits. Since endpoints are predefined and directly supervised, one slot per task provides a stable endpoint readout over shared WSI-report evidence; soft competition and concept priors prevent collapse and stabilize semantics without event-to-endpoint matching ambiguity. Survival slot (R2) 1.Diagnostic-to-prognostic hierarchy. Survival is not treated as a peer biomarker. PathoSlot follows a hierarchy: biomarker slots capture local diagnostic evidence, while the survival slot models downstream prognostic risk and controls global context injection. 2.Residual is intentional, not inconsistent. The context-gated residual lets the survival slot gate patient-level WSI-report context needed for prognosis, while biomarker slots remain local; adding this residual to biomarker slots degraded performance in Fig. 3.Slot semantics (R3) 1.Clarified claim. We will revise “decouples/factors” and clarify that PathoSlot learns task-aligned slots, not independent latent factors. 2.Current evidence. The last two rows of Table 2 show concept conditioning stabilizes slot semantics and improves external robustness; Fig. 2 shows clinically plausible MMR/HER2 slot attention. 3.Added evidence. We will add slot–concept similarity analysis to directly assess slot-semantics consistency. Ablation/baselines (R2) On TCGA-BRCA, from Table 3 row 2, adding init-only gives ER/PR/HER2 AUC=0.906/0.820/0.703, C-index=0.640; cosine-only gives 0.907/0.816/0.706, 0.646; using both is best, row 3: 0.911/0.825/0.706, 0.648.Thus, both help slot alignment. For matched backbones, Table 3 rows 1–4 are TITAN tokens + simple concatenation; row 5 shows Token Reweighting improves all metrics. Patch-feature PathoSlot gives 0.905/0.826/0.691, C-index=0.654, below WSI-report pre-pooling tokens, 0.923/0.859/0.737, 0.703.Significance (R3) We performed paired tests against the best competitor per endpoint and will report Δ, 95% CI, and p-values, e.g., TCGA-BRCA: ΔER-AUC=+0.013 [0.004,0.023], p=0.019; ΔHER2-bACC=+0.030 [0.017,0.051], p=0.007; ΔC-index=+0.085 [0.055,0.124], p<0.001.Novelty (R1) The novelty is not the application alone, but concept-conditioned multimodal slot factorization for joint biomarker-survival prediction. Unlike prior unimodal methods or ones ignoring task heterogeneity, PathoSlot learns task-aligned slots over shared WSI-report evidence and uses concept conditioning to stabilize slot semantic. Comparison details (R3) We kept each baseline’s training setup, modifying only the prediction heads for biomarker-survival endpoints. We will clarify this.
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.
This paper proposes PathoSlot, which jointly optimises biomarker prediction and surival prediction based on whole-slide-images and pathology reports. The algorithm is well motivated and relevant. R2 and R3 posed several major concerns for the authors to clarify. The manuscript uses private patient cohorts and pathology-report processing, but I could not identify an explicit IRB/ethics statement in the paper. This should be clarified.
- 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.
The paper addresses the challenge of balancing biomarker prediction with global survival modeling using multi-modal digital pathology data. To resolve this, the authors propose an pipeline that maps whole-slide images and text reports into task-aligned, concept-conditioned latent slots, demonstrating strong performance on independent external cohorts. The rebuttal effectively addressed the most of the reviwer concerns for example regarding evidence partitioning by clarifying the methodological design, resolved the omission of the IRB statement, and supplied statistical testing.
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 concept-conditioned multimodal slot framework for joint biomarker and survival prediction from WSIs and pathology reports. Reviewers appreciate the clinical relevance of the task and the promising empirical results. The rebuttal addresses several concerns regarding slot interpretation, report-processing reproducibility, statistical support, and ethics clarification. Some concerns remain regarding the strength and wording of the claims, but they should be addressable in revision.
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.
The reviewer raising negative concerns did not provide specific comments, and the conclusions appear inconsistent.