List of Papers Browse by Subject Areas Author List
Abstract
Survival analysis using Whole Slide Images (WSIs) is often challenged by histological artifacts and ambiguous tissue morphology within the tumor microenvironment (TME). While Multiple Instance Learning (MIL) has become a standard paradigm for this task, existing methods frequently overlook feature-level reliability, leading to overconfident yet erroneous prognostic stratifications. For advanced prognostic modeling, explicitly managing uncertainty is imperative to reduce out-of-distribution noise and refine ambiguous signals. To this end, we propose UT-MIL, an uncertainty-aware TME-decoupling framework that integrates clinical prognostic logic through two synergistic phases. First, we introduce the Uncertainty-Guided Feature Rectifier (UGFR) to estimate feature-level uncertainty using perturbation sensitivity and a Histomorphological Memory Bank. Unlike passive estimation, UGFR performs active feature intervention via a tri-branch routing mechanism: removing highly uncertain features, preserving reliable features, and refining ambiguous representations via prototype retrieval. Second, to aggregate the resulting purified manifold, we design the TME-Decoupling Contextual Aggregator (TDCA). TDCA employs a dual-stream architecture comprising a Critical Stream to highlight high-risk malignant foci and a Context Stream that utilizes learnable latent queries to decouple multifaceted TME patterns. Comprehensive evaluations across five TCGA benchmarks demonstrate that UT-MIL achieves superior prognostic performance over representative MIL baselines. Beyond empirical gains, our framework yields high-fidelity, noise-resilient spatial interpretability that integrates localized risk determinants with global histological contexts. Our code is available at https://github.com/hutaiyuan/UTMIL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4829_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hutaiyuan/UTMIL
Link to the Dataset(s)
TCGA: https://portal.gdc.cancer.gov
BibTex
@InProceedings{HuTai_UTMIL_MICCAI2026,
author = { Hu, Taiyuan AND Luan, Haijing AND Hu, Jifang AND Li, Han AND Pei, Yue AND Yang, Kaixing AND Li, Ruilin AND Niu, Beifang AND Chi, Xuebin AND Yan, Rui AND Jiang, Jinrong},
title = { { UT-MIL: Uncertainty-Rectified and TME-Decoupling Dual-Stream Aggregation for Robust WSI Survival Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors proposed an uncertainty-rectified multi-instance learning framework, which models the noise injection-based aleatoric uncertainty term, along with a clustering-based epistemic uncertainty term. Moreover, instead of traditional soft attention in ABMIL, a tri-branch hard routing mechanism was introduced to handle the high, intermediate, and low uncertainty patches.
- 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 idea to decouple and explicitly formulate the aleatoric and epistemic terms with noise injection and clustering is intriguing and seems to be effective. The authors also performed extensive evaluation on the pancancer datasets, showing superior performance against existing SotA ABMIL approaches in survival prediction in both c-index and KM-curve/logrank test for risk prediction and stratification. The authors also conducted ablation studies to identify the impact of handling aleatoric and epistemic uncertainty individually.
- 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.
Rather than showing the mean/std of c-index across k-fold cross-validation, it may be more concise to simply compute the “global” c-index of all out-of-fold predictions. Moreover, the authors should perform statistical tests such as paired permutation tests, or at least bootstrap CI of the c-index difference, to demonstrate that the c-index improvement is consistent/significant. It is also easier for readers if the authors directly address the relative c-index improvement in percentage.
Moreover, while the authors already showcase the clinical interpretation of regions with various uncertainty identified by the model, it is important to know whether the uncertainty modeling is inflated because of the existence of abundant of artifacts, or that the model can still identify the most aggressive patterns only within high image quality tissue patches.
- 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
(1) As author stated, the signal of survival trajectory may reside beyond the most aggressive tumor pattern. So to this end, if we limit the bag of patches to only those without artifacts through quality control, do these regions discarded during the routing phase really not informative to survival prediction? This may need an explicit quantitative answer. (2) It can be interesting to know the impact of # of histomorphological phenotypes as well as the threshold of uncertainty.
- 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 study is shipped with relatively comprehensive experimental design and sufficient size of cohort in the scope of a conference publication. However, the authors should at least provide the statistical test or bootstrap CI of c-index as the evidence to show the improvement is significant. The relevant tests involve no model training and can be trivially added to Table 1 and Table 2 so it poses no time/space constraint.
- 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
The main contribution is an uncertainty-aware MIL framework that explicitly disentangles and rectifies aleatoric and epistemic uncertainties at the feature level. It further introduces a TME-decoupling aggregation mechanism to jointly model local risk regions and global histological context, achieving state-of-the-art performance.
- 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 presents a relatively comprehensive experimental evaluation under multiple settings, which helps to demonstrate the effectiveness of the proposed method to some extent.
- 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 Introduction does not sufficiently summarize or analyze the current research landscape. While several related works are mentioned, there is a lack of structured discussion on existing limitations and how the proposed method is fundamentally different from prior MIL-based survival models. Strengthening this part would help clarify the motivation and novelty of the work. 2.In Section 2.1 (Tri-Branch Dynamic Routing), the paper introduces a hard-routing mechanism based on uncertainty quantiles. However, the rationale behind choosing hard routing over conventional soft-attention-based masking is not clearly justified. Given that soft attention is widely adopted in MIL frameworks, a more detailed explanation (e.g., theoretical intuition or empirical motivation) is necessary to support this design choice. 3.The paper lacks sufficient theoretical grounding in the Dual-Uncertainty Estimation module. The two proposed uncertainty measures appear to be heuristic constructs rather than principled probabilistic formulations. Specifically, the so-called aleatoric uncertainty is derived from feature perturbations induced by externally injected noise 𝜖, which is independent of the true data-generating process. This formulation seems to reflect perturbation sensitivity or representation instability, rather than intrinsic data uncertainty. Likewise, the epistemic uncertainty, defined as the distance to the nearest prototype, is more aligned with an OOD or novelty measure. Such a distance-based metric does not directly quantify uncertainty in the model’s predictive function 𝑓, and therefore may not faithfully capture epistemic uncertainty.
- 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?
While the paper demonstrates novelty, the overall presentation and technical explanations need to be more thorough. Certain essential components are insufficiently described and should be further clarified.
- 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 #3
- Please describe the contribution of the paper
The authors propose UT-MIL, an uncertainty-aware TME-decoupling MIL framework designed to address morphological ambiguity in the Tumor Microenvironment (TME) for WSI-based survival analysis. By explicitly modeling uncertainty, the framework prevents over-confident predictions in ambiguous regions, thereby enhancing the clinical reliability and robustness of AI-driven prognostic decision-making.
- 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 introduces an Uncertainty-Rectified Group Aggregator (URGA) that calculates and sums aleatoric and epistemic uncertainties. This allows for a sophisticated tri-branch routing system (keep, refine, drop), ensuring that only high-quality or appropriately adjusted features contribute to the final prediction.
- The proposed TME-Decoupling Contextual Aggregator effectively separates features into a Critical Stream for isolating key individual patch features and a Latent Query Context Stream for capturing diverse histological patterns. This dual-stream approach provides a more nuanced representation of the complex TME landscape.
- The use of t-SNE visualizations to demonstrate the separation of ambiguous and reliable features in the manifold provides strong qualitative support for the effectiveness of the uncertainty-based filtering.
- The study successfully demonstrates that high-level feature engineering—leveraging uncertainty-based routing and contextual decoupling—leads to meaningful performance improvements in WSI survival analysis across multiple benchmarks.
- 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.
- While the proposed modules contribute to performance, the framework relies on a highly explicit and hand-crafted architecture. This design choice deviates from the current trend in representation learning, which favors end-to-end, data-driven architectures over manually engineered routing logic.
- The current implementation requires manual tuning of uncertainty routing thresholds for each specific task. The framework would be more robust and scalable if these thresholds were implemented as trainable parameters rather than fixed hyperparameters requiring manual intervention.
- The paper lacks a thorough sensitivity analysis regarding the uncertainty thresholds. There is insufficient justification for the chosen values, and it remains unclear how sensitive the final survival predictions are to fluctuations in these parameters.
- Despite the claim of “TME-decoupling,” the qualitative evidence in Figure 4 primarily highlights the severe tumor core. There is a lack of detailed clinical explanation or evidence demonstrating how the model specifically distinguishes and decouples the broader TME components.
- In Figure 2, the Kaplan-Meier survival curves are presented without a direct comparison to existing baseline methods. To demonstrate superior prognostic stratification, it is necessary to show that the proposed method yields more distinct and statistically significant survival separations compared to previous models.
- The manuscript contains inconsistent terminology, such as the use of “Uncertainty-Rectified Feature Frontend (URGA).” The application of terms like “Frontend” and “Backend” feels somewhat unconventional in this architectural context and may lead to confusion for the reader.
- 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 paper presents a technically sound framework that enhances clinical reliability by integrating uncertainty-aware routing with TME-decoupling. By filtering for reliable features and diversifying task-specific branches into critical and contextual streams, the authors demonstrate meaningful performance gains in WSI survival prediction.
- 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
N/A
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 paper introduces a new framework for WSI survival prediction by rectifying uncertainty. The proposed mehtod is evaluated on five public datasets and show performance improvements over a series strong baselines. All three reviewers give positive scores. Though quite a few weakness points are listed by the reviewers, most issues can be resolved given enough space and time.