List of Papers Browse by Subject Areas Author List
Abstract
Whole Slide Images (WSIs) pose significant challenges for survival modeling due to their complex tissue heterogeneity. While Multiple Instance Learning (MIL) shows promise, it lacks semantic transparency, failing to identify the underlying pathological concepts driving predictions. Moreover, although vision-language models introduce semantic priors, they rely on static text prompts that remain invariant during inference across all patients, failing to capture the morphological variations of phenotypes. To bridge this gap, we propose ConVL, a novel, interpretable concept-guided vision-language MIL framework that explicitly grounds survival prediction in explicit pathological knowledge. ConVL introduces three novel mechanisms: (1) Instance Conditioned Prompt Learning, which dynamically adapts concept prototypes to slide-specific visual contexts, enhancing robustness to tissue heterogeneity; (2) a Synergistic Dual-Stream Mechanism combining data-driven visual features with a knowledge-driven Concept Stream, utilizing Concept-Guided Cross Attention to quantify pathological attributes; and (3) a Distribution Alignment Loss that enforces consistency between the visual feature manifold and the concept-based prediction manifold. We evaluate ConVL across nine real-world cancer cohorts from TCGA. ConVL not only achieves state-of-the-art survival prediction performance but also provides fine-grained, concept-driven interpretability. Code is available at https://github.com/junjianli106/ConVL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3030_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/junjianli106/ConVL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiJun_ConVL_MICCAI2026,
author = { Li, Junjian AND Kuang, Hulin AND Liu, Jin AND Yue, Hailin AND He, Mengshen AND Li, Xinyu AND Wang, Jianxin},
title = { { ConVL: Interpretable Concept-Guided Vision-Language MIL for Survival Analysis in Whole Slide Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, the authors propose an interpretable concept-guided vision-language MIL framework that explicitly grounds survival prediction in explicit pathological knowledge. Extensive experiments on nine TCGA cohorts demonstrate the effectiveness of the proposed 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 proposed methodology demonstrates a certain degree of novelty, and the model achieves good performance.
- 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.
See comments.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 proposed methods focus on integrating clinical concepts to improve survival prediction accuracy. Overall, the paper is fine, but several concerns remain, as outlined below: 1: I think the description about “Instance Conditioned Prompt Learner” is unclear. For example, how does the “Concepts with Description” been utilized in the Learner? What is the correlation between concepts (text) and conditioned prompt (feature)? 2.Since the model generates two survival predictions, P_vis and P_con, how to obtain the final prediction used in experiments? 3.The experiments are limited to survival analysis tasks, whereas the evaluation of multiple instance learning models on cancer diagnosis tasks is more convincing. 4.A minor suggestion for Fig. 3: In the Concept Distribution subfigures, it would be better to assign the same color to the same concept across subfigures, while using distinct colors for different concepts. For instance, ‘Well-differentiated tumor cells’ appears in different colors between the two subfigures, which may cause confusion. In summary, I think the current manuscript is fine but requires further refinement.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
While not fully satisfied with the response, but I think this work overall acceptable.
Review #2
- Please describe the contribution of the paper
This paper proposes ConVL, a concept-guided vision-language MIL framework for survival prediction from whole-slide images. The method introduces instance-conditioned prompt learning, a dual-stream architecture combining visual and concept-based reasoning, and a distribution alignment loss to enforce consistency. Experiments across nine TCGA cohorts demonstrate consistent improvements over prior 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.
Strong empirical evaluation across multiple datasets, showing consistent improvements. The framework is well-motivated and integrates semantic and visual reasoning in a coherent manner. Provides interpretable outputs linking predictions to pathological concepts.
- 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.
- Insufficient validation of instance-conditioned prompt learning: although the proposed prompts are adapted per slide, the paper does not provide a clear comparison against standard static prompt approaches or analyze when and why dynamic prompts are beneficial.
- While an ablation study is provided, the overall framework remains relatively complex, and it is unclear whether a simpler formulation could achieve similar performance.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
Strong empirical performance across multiple datasets is the main positive factor. The main concerns are insufficient validation of the dynamic prompt design, and relatively high model complexity compared to the performance gain.
- 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 author’s feed back solve my mainly concern.
Review #3
- Please describe the contribution of the paper
The paper proposes ConVL, a concept-guided vision-language MIL framework for survival prediction on whole slide images. Experiments on nine TCGA cohorts report improvements in C-index over recent MIL and vision-language baselines.
- 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 is well written and easy to follow. The motivation connecting tissue heterogeneity to the limitation of static prompts is clearly stated, and the method section is logically organised. Figure 1 offers a clear overview of the dual-stream pipeline, and Figure 3 communicates the intended interpretability story by placing the original WSI, attention heatmap, concept assignment map, and concept distribution side by side, which makes the claimed behaviour easy to inspect.
- 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 methodological novelty is limited. Instance Conditioned Prompt Learning is a direct adaptation of CoCoOp to the bag-level WSI setting, where the conditioning signal is simply mean-pooled instance features. The Concept-Guided Cross-Attention is a standard cross-modal attention module, and the Distribution Alignment Loss is a conventional symmetric KL consistency term widely used in multi-view and multi-modal learning. The overall design reads as an assembly of known techniques rather than a new formulation.
2.The evaluation relies solely on C-index, which is insufficient for survival analysis. Kaplan Meier curves with log-rank tests on stratified risk groups are standard complementary metrics and should be reported to substantiate the clinical value of the proposed framework.
3.The LLM-generated concept pool, which is the semantic foundation of the method, is not validated. The concepts are produced by a single prompt to Gemini 3 Pro without pathologist review, without reproducibility analysis across repeated LLM runs, and without comparison across different LLMs such as GPT-5 or Claude.
4.The interpretability analysis is weak. The claim rests on two qualitative cases in Figure 3 with a single pathologist comment. No fine-grained feature visualisation is provided, for instance overlaying concept activations on cell-level or gland-level annotations, and no quantitative interpretability metric is reported. Since interpretability is a headline claim of the paper, stronger evidence is expected.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
The paper is clearly written and targets a relevant problem, but the contribution is incremental. The evaluation uses only C-index, and the two central claims of the paper, semantic grounding through LLM-derived concepts and fine-grained interpretability, are not convincingly supported.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 author’s reply addressed my concerns, and I agree to accept it.
Author Feedback
We thank all reviewers for their constructive comments. We are encouraged that they found ConVL well-motivated (R2, R4), effective across 9 datasets (R1, R2, R4), reproducibility (R1, R2, R4),and interpretable (R2, R4). We respond below.
[R1.Q1].ICPL Clarification:In ICPL,each concept description is first tokenized by the VLM tokenizer.Then,learnable prompt tokens are combined with patient-specific tokens generated from the WSI feature.The resulting conditioned prompts are concatenated with each concept token embedding and fed into the text encoder.Thus,concept descriptions provide pathological semantics,while patient-specific tokens inject WSI-level visual context.
[R1.Q2].Final Prediction:We average the two branches for all experiments,i.e.,P=(P_vis+P_con)/2. [R1.Q3].Diagnosis Task:We agree that diagnosis tasks would further validate the generality of MIL. We will add this task in the journal version.
[R1.Q4].Figure Color:We will revise Fig. 3 accordingly.
[R2.Q1].Dynamic Prompt Validation:In ablation(Tab.2),replacing Dynamic Prompt(w/o ICPL) with static prompts decreases the avg C-index from 0.658 to 0.631.In addition,static-prompt VLM baselines,such as TopMIL(0.634), FOCUS(0.634), and VLSA(0.627),also underperform ConVL.These results confirm the benefit of ICPL.
[R2.Q2].Framework Complexity:In the ablation study(Tab.2),simplified variants reduce performance:static prompts instead of ICPL,gated attention instead of CGA,and removing DAL.Their avg C-index drops to 0.631,0.639,and 0.641,compared with 0.658 for full ConVL,suggesting that these simplified variants cannot achieve comparable performance.We also added CO&RE(625 WSIs) computational profile results,ConVL uses fewer FLOPs and infer time(131.0G,15.1ms/WSI) than VLSA(188.5G,21.3ms) and FOCUS(160.2G,106.2ms).
[R4.Q1].Novelty:ConVL’s main novelty is patient-specific concept adaptation for WSI survival MIL.First,unlike TopMIL(23NeurIPS),FOCUS(25CVPR),and VLSA(25ICLR),which mainly use static prompts shared by all WSIs,ConVL adapts each concept prototype to the current WSI via ICPL.The slide-level instance summary generates patient-specific prompt offsets before text encoding,so the same concept can have different WSI-specific features under different morphological contexts.Second,CoCoOp conditions image-level class prompts on a single image features,whereas ICPL conditions pathological concept prompts on a bag-level WSI features derived from thousands of instances.ConVL therefore addresses WSI survival prediction with patient-specific concept prototypes,rather than image-level class prompts.Third,ICPL,CGA,and DAL form a WSI-specific framework for concept adaptation,evidence localization,and dual-stream consistency.The ICPL ablation drop from 0.658 to 0.631 further supports its contribution.
[R4.Q2].Survival Evaluation:Based on the same predicted risk scores used for C-index evaluation,we performed KM analysis by stratifying patients into high- and low-risk groups. Log-rank tests are significant on all nine cohorts:BLCA p=0.0122,BRCA p=0.0022,CO&RE p<0.0001,LUSC p=0.0018,KIRC p<0.0001,SKCM p<0.0001,LIHC p<0.0001,SARC p=0.0016,and UCEC p<0.0001. [R4.Q3].Concept Pool Validation:Repeated LLM (5 times,Gemini) generations are stable, with an avg C-index of 0.657±0.003.Cross-LLM pools also give comparable results: Gemini(0.658), GPT(0.654), and Claude(0.660), showing that ConVL generalizes across different LLMs.. We will report concept stability,and include pathologist verification.
[R4.Q4].Interpretability:We strengthened interpretability with CRC-100K predicted 9-class tissue maps.The tissue map is a patch-level tissue-type reference from a classifier trained on 9 colorectal tissue categories,used to assess whether ConVL’s concept assignments align with recognizable tissue structures,and this consistency was further confirmed by an expert pathologist.The visualization shows consistency between concept assignments and tumor-related phenotypes.We will update Fig.3.
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 a concept-guided vision-language MIL framework for WSI-based survival prediction. The reviewers generally appreciate the integration of pathological concepts into survival modeling and the empirical evaluation across multiple TCGA cohorts. Several concerns remain regarding the strength of the methodological contribution and the support for the main claims. For example, the reviews question whether the dynamic prompt design and the overall model complexity are sufficiently justified compared with simpler alternatives. The authors are encouraged to address the detailed concerns in the rebuttal.
- 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 rebuttal addresses most concerns such as the method design, prediction strategy, model complexity, survival evaluation, etc. Some limitations remain regarding the scope of evaluation and the strength of interpretability claims, but these seem addressable in revision.
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.
After rebuttal, this paper receives three accept recommendations, including one reviewer who changed from weak reject to accept. The AC acknowledges that ConVL addresses an important problem in WSI-based survival analysis by integrating pathological concepts into a vision-language MIL framework, with consistent empirical improvements across nine TCGA cohorts. The rebuttal reasonably addresses the main concerns by clarifying the instance-conditioned prompt learner and final prediction strategy, providing stronger evidence for dynamic prompts and model components through ablations, adding survival-specific evaluation with KM/log-rank analysis, and discussing concept-pool stability and interpretability validation. Although some concerns remain regarding methodological complexity, limited diagnosis-task evaluation, and the need for stronger concept/pathologist validation in the final version, the reviewers generally agree that the work is well motivated, reproducible, empirically solid, and relevant to MICCAI. Therefore, the AC recommends acceptance.
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 authors successfully addressed most of the concerns raised by the reviewers. All reviewers agree to accept this manuscript after rebuttal.