List of Papers Browse by Subject Areas Author List
Abstract
Existing computational pathology methods predominantly operate within whole-slide image (WSI)-level multiple instance learning (MIL) paradigms, while patient-level modeling remains underexplored.
In routine pathological practice, however, pathologists derive diagnostic and prognostic conclusions by integrating evidence across multiple WSIs rather than relying on any single slide.
This discrepancy creates a fundamental misalignment when patient-level supervision is directly imposed on conventional MIL frameworks, often leading to unstable optimization and degraded predictive reliability.
To address this issue, we propose Anchor-Guided Evidence MIL (AGE-MIL), a weakly supervised framework for patient-level prediction. AGE-MIL constructs a patient-level anchor from slide representations to capture global pathological context and guide the retrieval and integration of diagnostically relevant local patches, enabling robust patient-level modeling.
Patient-level risk is further modeled as an evidence accumulation process, promoting stable optimization under weak supervision.
AGE-MIL is evaluated on six clinically relevant patient-level prediction tasks from two independent cohorts. Experimental results show that the proposed framework consistently outperforms eight state-of-the-art MIL methods.
Code is available at https://github.com/wodeniua/AGE-MIL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5366_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wodeniua/AGE-MIL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{NiuJia_AGEMIL_MICCAI2026,
author = { Niu, Jiawei AND Chen, Jian AND Zhang, Di AND Lu, Junbo AND Liao, Zhangcheng AND Liu, Xuhao AND Zhong, Honglin AND Crispin-Ortuzar, Mireia AND Li, Chen AND Gao, Zeyu AND Cai, Yi},
title = { { AGE-MIL: Anchor-Guided Evidence Learning for Patient-Level Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes AGE-MIL, a framework designed for patient-level modeling rather than simple slide-level analysis.
- 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.
By integrating information across multiple slides, just as pathologists do in routine practice. This approach addresses a critical gap in existing MIL research and demonstrates strong clinical feasibility.
- 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 approach of selecting important slides to query key patches is conceptually appealing, its practical advantage over standard Attention-based MIL or simple Mean MIL is unclear. The marginal performance gain over Mean MIL suggests that either the task itself is not challenging enough or the proposed method fails to extract more discriminative representations. The authors should clarify why this complex selection process is necessary if it does not significantly outperform baseline pooling methods.
-
Similar challenges regarding selection bias in multi-slide (or 2.5D/3D) pathology have been addressed in recent literature, such as CARP3D. The authors should provide a detailed theoretical comparison and cite such works to justify the unique contribution of their method.
-
It would be highly informative to see an analysis of how performance scales with the number of slides. This would clarify the conditions under which the proposed selection mechanism becomes most effective.
-
To support the claim of “important slide/patch selection,” visual evidence (e. g. , attention heatmaps) should be provided to verify if the model actually focuses on clinically relevant regions.
-
- 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
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 is well-structured and presents a compelling motivation with a problem definition that is closely aligned with real-world clinical settings.
- 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 #2
- Please describe the contribution of the paper
This paper addresses the problem of patient-level prediction in computational pathology when a single patient is associated with multiple whole slide images (WSIs). To support patient-level conclusions, the authors point out a limitation of direct patch-to-patient aggregation, namely that it may disrupt the natural patch–slide–patient hierarchy. Based on this motivation, they propose AGE-MIL. In the Anchor-Guided Evidence Retrieval module, the method assigns scores to potentially informative evidence patches using both conditional similarity and unconditional similarity. The retrieved evidence is then interacted with slide-level representations through cross-attention, allowing the model to integrate evidence-level and slide-level features. Finally, the model aggregates information across multiple slides using a risk accumulation score based on Log-Mean-Exp to generate the final patient-level prediction. The method is evaluated on a real-world prostate dataset consisting of more than 10,000 WSIs from over 700 patients across six clinical tasks, and the authors report an improvement of approximately 2% 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.
One major strength of the paper is that it proposes a technically meaningful way to preserve diagnostically relevant evidence that may otherwise be weakened or lost when forming slide-level representations in conventional methods. In particular, the use of cross-attention between retrieved evidence and slide-level representations, followed by their integration, is a technically interesting design choice that appears to improve the expressiveness of the final representation.
Another strength is that the authors evaluated the method on a task setting in which the proposed formulation is especially relevant. If the task had been limited to a simpler setting such as cancer detection, one might argue that slide-level MIL followed by a patient-level max-pooling strategy could already be sufficient, and that each slide could potentially be treated independently. In contrast, this paper considers tasks in which risk may accumulate across multiple slides, making patient-level reasoning more appropriate and making the contribution of the proposed method more meaningful.
In addition, the ablation study is a positive aspect of the work. The results suggest that the evidence integration component, especially the evidence concatenation part in AGE-MIL, is one of the most important contributors to performance. This supports the authors’ motivation for the proposed design.
- 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.One weakness is that the dataset appears to come from a single institution (“from the same medical institution”), which limits the ability to assess generalizability. Validation on a public dataset or an external cohort would make the experimental evidence substantially stronger.
2.Another concern is the baseline comparison. The methods used as baselines were originally designed for WSI-level prediction, not specifically for patient-level prediction. Moreover, the manuscript does not clearly describe how these MIL baselines were adapted to the patient level. If the comparison was based on patch-level aggregation, then the paper is effectively evaluating baselines under the very limitation highlighted in Fig. 1.On the other hand, if WSI-level feature extraction was not included as a baseline, then the experiments do not fully demonstrate that the proposed method resolves the core problem raised in the paper’s motivation. For this reason, I think the claim may be somewhat stronger than what is fully supported by the current experiments.
Relatedly, I was left wondering why the methods discussed in the related work section, especially references 10, 3, and 28, were not included in the empirical comparison. Since the main contribution of the paper is closely tied to patient-level aggregation across multiple slides, comparison with methods developed for similar purposes would make the technical contribution more convincing.
3.I also think stronger hierarchical baselines could have been included. For example, one could straightforwardly construct a bag-of-bags style baseline by first applying ABMIL to obtain slide-level representations from patches, and then applying another ABMIL module over the slide representations to obtain a patient-level prediction. Such a baseline seems relatively easy to implement and would provide a stronger and fairer comparison. If the current experiments only compare against patch-level aggregation methods, then the baseline setup may be too weak to fully establish the technical merit of the proposed method.
- 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
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?
My overall recommendation is based on the fact that the paper is methodologically well structured and that both the problem formulation and technical description are clear. The paper also addresses an important setting in pathology, namely patient-level prediction from multiple WSIs, which is clinically meaningful.
However, I think the current experimental evidence is not yet sufficient to fully establish the method’s generalizability or technical advancement. In particular, the data appear limited to in-house data from a single institution, and the baseline comparisons seem somewhat weak. Rather than comparing primarily against models developed for similar patient-level objectives, the paper mainly compares against methods originally designed for WSI-level prediction, which may not be a fully fair comparison. In particular, if methods from ABMIL to ILRA were all adapted using the same direct patch-to-patient aggregation strategy, this may create an overly sparse weak-supervision setting that inherently disadvantages the baselines.
For a fairer and more informative comparison, I think it would be useful to include not only the patch-to-patient aggregation setting used in the paper, but also hierarchical baselines based on patch-to-slide-to-patient aggregation. Such comparisons would make it easier to understand the true technical benefit of the proposed method. Overall, I find the paper promising and well motivated, but I think the current experimental setup is somewhat limited for fully supporting the claimed novelty and generalizability.
- 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
This paper propose AGE-MIL, a weakly supervised framework for patient-level prediction that constructs a patient-level anchor to capture global pathological context and guide the retrieval and integration of diagnostically relevant local patches. Patient-level risk is modeled as an evidence accumulation process to enable stable optimization under weak supervision. Evaluated on six clinically relevant tasks across two independent cohorts, AGE-MIL consistently outperforms eight state-of-the-art MIL 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 method benefits from a rigorous experimental setup with comprehensive validation across diverse tasks and datasets, yielding substantial and consistent performance improvements that strongly support its effectiveness.
- 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.
First, duplicate references. In the reference list, entries [1] and [2] are identical, representing the same work being cited twice. Second, the abbreviation “Cond-Sim” is not explicitly defined in the main text. In the ablation study presented in Table 3, “Cond-Sim” refers to the conditional similarity retrieval mechanism. However, this abbreviation is not explicitly introduced in the method section. Third, certain symbols are not explained at their first occurrence. In Equation (1), the temperature parameter \tau and the function \sigma appear without explicit definition.
- 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?
The method benefits from a rigorous experimental setup with comprehensive validation across diverse tasks and datasets, yielding substantial and consistent performance improvements that strongly support its effectiveness.
- 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
R1.Thank you for the comment. This work studies patient-level prediction from multiple WSIs, which differs from conventional single-WSI MIL. AGE-MIL is designed for this hierarchical setting, whereas directly applying ABMIL or MeanMIL is less suitable.
R2.Thank you for the suggestion. All baselines were compared under a unified patient-level aggregation framework using TITAN-derived slide features. References [10, 3, 28] mainly address slide-level hierarchical MIL rather than our patient-level task.
R3.Thank you for your careful reading. We corrected the duplicated reference, defined Cond-Sim, and clarified τ and σ in Eq. (1).
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 a weakly supervised framework for patient‑level prediction that constructs a patient‑level anchor to guide the retrieval and integration of diagnostically relevant patches across multiple whole slide images, modeling patient risk as an evidence accumulation process. This paper addresses a critical gap in existing MIL research by moving from slide‑level to patient‑level prediction and the method is comprehensively evaluated on six clinically relevant tasks. The reviewers also raised concerns about the practical advantage over simpler baselines and the manuscript does not clearly describe how the MIL baselines were adapted to the patient‑level setting. I think the weakness can be addressed in practice and the problem solved in this paper is valuable to the field.