List of Papers Browse by Subject Areas Author List
Abstract
Cancer survival prognostication using whole-slide images (WSI) remains challenging, as WSI contain rich yet complex prognostic signals embedded within a broader tissue architectural and tumor microenvironmental context. The gigapixel resolution of WSI also imposes substantial computational constraints on deep learning models. Multiple instance learning (MIL) mitigates this by aggregating tile-level features, but often treats tiles as an unordered set, discarding architectural context critical for modeling tumor biology and disease progression. Graph-based MIL restores spatial structure but is typically limited to pairwise edges and fails to capture higher-order tissue interactions. To address these limitations, we propose TRIAGE-MIL, a two-stage framework that couples unsupervised instance selection with semantic hypergraph modeling. First, we introduce multi-axis stratified sampling (MASS), which selects a fixed-size subset of tiles using statistical proxies along four phenotype-inspired axes: morphologic heterogeneity, tissue regularity, microenvironmental interface, and tissue diversity. This yields a proxy-enriched slide representation without requiring region-level ground-truth labels. Second, we introduce a semantic hierarchical hypergraph network that models intra- and inter-tissue relationships via constrained hyperedges induced by MASS groupings, enabling higher-order relational reasoning beyond pairwise neighborhood. Across six different cancer cohorts, TRIAGE-MIL achieved the highest mean C-index among 13 state-of-the-art baselines and showed significant Kaplan-Meier risk-group separation in all cohorts (log-rank p<0.05), indicating consistent patient risk stratification. https://github.com/barathi-1993/TRIAGE-MIL
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4151_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/barathi-1993/TRIAGE-MIL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{SubBar_TRIAGEMIL_MICCAI2026,
author = { Subramanian, Barathi AND Jeyaraj, Rathinaraja AND Noh, Songmi AND Fisher, George A. AND Shen, Jeanne},
title = { { TRIAGE-MIL: Multi-axis Instance Selection and Semantic Hypergraph Modeling for Survival Prediction from Whole-Slide Images } },
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 proposes TRIAGE-MIL, a two-stage framework for WSI-based survival prediction. The method combines unsupervised multi-axis tile selection (MASS) with a semantic hierarchical hypergraph MIL model to capture higher-order tissue interactions. The paper evaluates the approach across six cohorts against 13 survival-prediction 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 hypergraph design is the most interesting part of the paper for me. Moving past pairwise edges to model higher-order tissue context is well motivated, and the semantic hierarchy isn’t a gimmick — it gives the model something to do with those edges that unordered pooling and standard GNN baselines genuinely cannot.
- Pairing unsupervised tile selection with the relational model is a real methodological contribution, not a reshuffle of existing MIL pieces.
- The evaluation is broader than most comparable submissions in this space: six cohorts, 13 baselines, and ablations that actually isolate the components.
- The ablations support the story. Both MASS and the relational modeling contribute, and the gains aren’t carried by a single 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 issue is the survival evaluation protocol. The 5-fold cross-validation (4:1 train:validation split per fold) description, combined with early stopping on validation performance, leaves me genuinely unable to tell whether the reported C-indices come from a held-out test fold or from the same fold used for model selection. Those are very different numbers, and this is what’s keeping my score where it is.
- The “outperforms 13 state-of-the-art baselines” framing is doing rhetorical work the statistics don’t quite support, given the fold-to-fold variability in the tables.
- Because instance selection is central to the contribution, the baseline setup needs to be pinned down — specifically whether all baselines received a comparable tile budget.
- No code, and one cohort is private and underdescribed. For a method with this many interacting components, that’s a meaningful reproducibility gap.
- 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 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
The two things driving my score:
1.CV protocol (this is the big one). Please state explicitly, in each fold, what data trains the model, what data early-stops it, and what data the reported C-index is computed on. If there’s a held-out test fold distinct from the validation fold used for early stopping, say so directly. Point 4 below is really the same question in a different form — please treat them as one concern.
2.Baseline tile budget. Did all baselines receive the same number of tiles, or the same input budget in some other sense? Since MASS is a central contribution, this matters a lot for how to read Table 1.
Other comments, in rough order of importance:
3.The internal CRC cohort needs more description — survival endpoint, slide source, and how it differs from the TCGA cohorts in ways that might matter.
4.For Fig. 3: please clarify how the KM curves and risk groups were constructed under CV. Did each patient contribute a single out-of-fold risk prediction? Were those pooled to form the cohort-level curve? Was the high/low split defined on the pooled out-of-fold predictions or fold-by-fold? And were the log-rank p-values computed on the pooled held-out groups? This is the other half of point 1.
5.If formal statistical comparisons weren’t run, I’d soften the “outperforms 13 state-of-the-art baselines” language somewhere in the intro or conclusion. The gains look real; the phrasing just overclaims relative to what’s shown.
6.The MASS ablation shows the full design beats a simpler internal variant, which is useful. What it doesn’t show is how MASS compares against plain random or uniform sampling at a matched tile budget. That’s the comparison I actually want to see.
7.Code release would help. A lot.
8.Minor: Fig. 3 is hard to read at the current size.
One side thought, not an ask: I’d be curious about how MASS degrades in smaller cohorts where multi-axis statistics are estimated from fewer slides. Not something to address in rebuttal, just something the design made me wonder about.
- 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?
Promising paper. Important problem, genuinely interesting method, broader evaluation than most. The thing keeping me at weak accept and not higher is that I cannot tell from the current text whether the headline numbers are on a held-out test fold or on the validation fold used for model selection. If that’s clarified in rebuttal — along with the baseline tile budget and the internal cohort description — I’d stay on the accept side.
- Reviewer confidence
Somewhat confident (2)
- [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
This paper presents TRIAGE-MIL, a two-stage framework for WSI-based survival prediction that combines unsupervised multi-axis instance selection with semantic hypergraph modeling. It first selects a fixed subset of informative tiles using an unsupervised multi-axis stratified sampling strategy based on four phenotype-inspired proxy axes, then models relationships among the selected tiles using a semantic hierarchical hypergraph with super-nodes and gated attention pooling.
- 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 method is easy to follow. This paper is a well-motivated paper as discusses two weaknesses in prior WSI survival methods.
- Authors propose a novel unsupervised tile sampling that stratifies tiles using embedding-derived proxies along morphologic heterogeneity, tissue regularity, microenvironmental interface, and tissue diversity.
- This paper also leverages from semantic hierarchical hypergraph with constrained hyperedges and a “super-node” structure that links tiles by spatial proximity and semantic similarity, enabling message passing over higher-order groups beyond pairwise neighborhoods
- The experimental section is comparatively strong. Six cohorts, 13 baselines, and component ablations give the paper credibility.
- 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 baseline set is fairly broad, but it would be stronger to include more recent structured survival models such as ProtoSurv and C2MIL, which are conceptually more relevant. Could the authors compare with them or at least clarify how they expect TRIAGE-MIL to compare against them?
-
The paper provides some ablation evidence that MASS improves performance over naive top-K sampling; however, the module still relies on hand-crafted proxy definitions, a proxy-based Optuna objective, and a fixed sequential assignment rule, without stronger theoretical or biological validation. It would be appreciated if more evidence be provided on How robust is the method to change of some hyperparameters like graph connectivity parameters, number of tiles per super-node, filtering thresholds, and alternative semantic selection orders.
-
It seems that some implementation details need more explaining. For instance, more explicit clarification on the hypergraph constructio, including how super-nodes are determined and how intra-semantic and inter-semantic hyperedges are formed in practice?
-
- 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?
While it seems more clarification and ablations can be added to this paper, I believe that their idea is novel, and it is worth publishing this paper at MICCAI.
- Reviewer confidence
Somewhat confident (2)
- [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
This paper presents a pipeline that combines unsupervised instance selection and semantic hypergraph to predict suvival from whole-slide images. The proposed method outperforms 13 baselines across 6 cohorts, demonstrating remarkable prediction gains. The key contributions include: 1.Using multi-axis stratified sampling (MASS) to stratifies tiles along four phenotype-inspired axes. 2.Employing semantic hierarchical hypergraph with constrained hyperedges to link tiles by spatial closeness and semantic similarity.
- 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.This method integrates relevant features from morphologic heterogeneity, tissue regularity, microenvironmental interface, and tissue diversity within whole-slide images, all these 4 dimensions provide concrete tile stratification for unsupervised tile selector. 2.Extensive experiments and ablation analyses being conducted improves model interpretability.
- 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 paper’s overall contribution is appreciated, but it does not directly compare the practical utility and computational efficiency with conventional graph-based or other baseline methods.
- 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.
(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?
Although whether the overall performance gains justify additional computational burden in clinical scenario remains unclear, the authors provide a thorough suvival prediction analysis.
- 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 proposes TRIAGE-MIL, a two-stage WSI survival prediction framework that couples unsupervised multi-axis stratified sampling (MASS) along four phenotype-inspired axes with a semantic hierarchical hypergraph model that performs higher-order relational reasoning, evaluated across six cohorts against 13 baselines. Reviewers appreciated the originality of the hypergraph design with super-nodes, the complementary MASS sampling strategy, and the broad and credible evaluation. However, the reviewers identified weaknesses including ambiguity around the cross-validation protocol [R1], the framing ‘outperforms 13 SOTA baselines’ that overstates the statistical evidence [R1], unclear baseline tile budget which is central since MASS is a contribution [R1], missing description of the internal CRC cohort and underdescribed Kaplan-Meier construction under CV [R1], absence of code and private cohort hindering reproducibility [R1], missing comparison with recent structured survival models such as ProtoSurv and C2MIL [R2], reliance on hand-crafted proxies and Optuna objective without robustness analysis on graph connectivity, super-node size, filtering thresholds [R2], underdescribed hypergraph construction [R2] and missing comparison of computational efficiency against graph-based baselines [R3]. The AC read the paper and recommends acceptance. The methodological contribution is interesting and the evaluation broad, with concerns centered on protocol clarification that should be addressed in the camera-ready version.