List of Papers Browse by Subject Areas Author List
Abstract
In text-guided few-shot whole-slide image (WSI) classification, multiple instance learning (MIL) methods can leverage vision–language priors with very limited annotations, yet they still lag behind strongly supervised baselines. We empirically show that this gap is largely caused by slide-to-slide domain shifts induced by tissue-composition and appearance heterogeneity. Under few-shot supervision, these shifts lead to unstable key-instance selection and feature aggregation, thereby degrading the reliability of slide-level predictions. This problem is particularly pronounced at low magnification (5x): although 5x features provide broader global context, they are more sensitive to staining variation, artifacts, and background content than 20x features, making low-magnification cues less reliable across slides. To address this issue, we propose a Heterogeneity-Aware Text-guided MIL with two modules: the Tissue-Composition Balancing Module (TCBM), which performs tissue-balanced patch selection across tissue-related clusters to mitigate composition bias, and the Appearance-Robust Aggregation Module (ARAM), which constructs low-magnification contextual representations from stable 20x local structural patterns without using raw 5× appearance features, which are sensitive to staining variation and slide-to-slide domain shifts, thereby improving robustness to appearance heterogeneity. Code is available at: https://github.com/noviq123/HAT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3002_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiYon_Addressing_MICCAI2026,
author = { Li, Yongcen AND Xu, Zelin AND Shi, Xiaoyu AND Huang, Zhiyi AND Ye, Xinchen AND Zhang, Miao AND Wang, Zhihui AND Xu, Rui AND Chen, Yen-Wei AND Xing, Xudong},
title = { { Addressing Tissue and Appearance Heterogeneity in Text-Guided Few-Shot WSI Classification } },
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 a method for constructing low-magnification (5×) features from high-magnification (20×) local features in order to reduce the impact of inter-slide domain shift when building a text-guided few-shot WSI classifier. The method is motivated by the observation that inter-slide domain shift, caused by factors such as staining variation and scanner-induced artifacts, affects low-magnification features more strongly than high-magnification features. The experiments suggest the effectiveness of the proposed approach.
- 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) The paper explicitly proposes a method for addressing inter-slide domain shift.
I believe many in the audience would agree that appearance heterogeneity in WSIs is one of the major factors that degrades classification performance. A particular strength of this paper is that it focuses on inter-slide domain shift as a cause of performance degradation in WSI classification and clearly demonstrates, through the experimental results in Fig. 1, that the extent of its impact differs between high- and low-magnification representations.
(2) Based on the above point, the paper addresses domain shift among slides by constructing low-magnification (5×) representations from high-magnification (20×) features.
The experiments show that suppressing the effect of inter-slide heterogeneity by deriving global low-magnification features from high-magnification features, which are less sensitive to domain shift, actually leads to improved performance. The results in Table 1 confirm that this effect is observed regardless of the number of shots. The results shown in Fig. 3 are also impressive.
(3) The method appears easy to adapt from existing MIL frameworks.
As shown in Eq. (9), the proposed method computes attention weights by adding text-based prior information to patch features. This is an appealing proposal because it does not require substantial changes to the MIL architecture.
(4) The paper carefully addresses the imbalance in the number of samples across patch types during training.
Instead of using all patches together as training data, the method first clusters the patches and then performs training after balancing the samples across clusters. Since lesion regions often occupy only a small portion of a WSI, I believe this kind of careful preprocessing for training should be appreciated.
- 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 comparative experiments are insufficient to demonstrate the benefit of constructing low-magnification features from high-magnification features.
In the proposed method, the function \phi in Eq. (6) aggregates sixteen high-magnification feature vectors arranged in a 4×4 grid to construct a low-magnification feature representation. To provide a fairer comparison and better demonstrate the effectiveness of the proposal, the authors should also compare their method with an alternative approach that maps 5× features into vectors of the same dimension using, for example, an MLP. By learning such a mapping function for the 5× features, it may be possible to transform them into representations that are more robust to domain shift.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Please see the comments in the major strength.
- 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?
Please see the comments in the major strength.
- 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.
I have read the comments from the other reviewers as well as the authors’ responses. Based on these, I maintain my original evaluation and recommend acceptance. Although Reviewer 2 provided a negative assessment, I support the authors’ response to those criticisms.
Review #2
- Please describe the contribution of the paper
This paper studies few-shot WSI classification with text-guided MIL and shows that performance drops mainly due to differences across slides, such as tissue composition and appearance. These differences make it hard to select important regions and combine features, especially at low magnification like 5x. To solve this, the authors propose a method that balances tissue types during patch selection and builds more stable features using 20x information instead of raw 5x images. Their approach improves robustness and makes predictions more reliable under limited annotations.
- 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.The paper successfully identified the key issues in text-guided few-shot learning, which are tissue-composition difference and appearance heterogeneity, and proposed two designs to mitigate those issues. The idea is well-motivated. 2.The experimental results look convincing. The proposed method consistently outperformed baselines under different settings and tasks.
- 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 proposed method is tested on three datasets. However, two of them are binary classification tasks, and one is a three-way classification. The robustness of the method in a multiple-class problem is not demonstrated. 2.The test settings are 4-shot, 8-shot, and 16-shot. The authors did not justify why these numbers were selected. 3.The caption of Figure 2 is too short. It is recommended that the caption should be self-contained, i.e., explain each part of the model. 4.The novelty of the core components is limited. The proposed TCBM and ARAM modules are essentially a relatively simple clustering operation.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Please see the weaknesses mentioned above.
- 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.
The rebuttal from the authors did not address my concerns, particularly regarding parameter choices and evaluation tasks.
Review #3
- Please describe the contribution of the paper
This paper proposes a Heterogeneity-Aware Text-guided MIL framework that consists of two modules:1.The Tissue-Composition Balancing Module (TCBM) reduces composition bias by selecting patches in a balanced way from different tissue-related clusters. 2.The Appearance-Robust Aggregation Module (ARAM) builds low-magnification contextual representations using stable 20× local structural patterns, which improves robustness against appearance heterogeneity.
- 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 paper identifies tissue composition and appearance heterogeneity as two major sources of slide-to-slide variation in text-guided few-shot WSI classification and demonstrates their effects through controlled experiments. 2.This paper proposes two dedicated modules: TCBM and ARAM. TCBM addresses composition bias by performing tissue-balanced patch selection across tissue-related clusters, whereas ARAM constructs low-magnification contextual representations from stable high-magnification structural information. Extensive experiments show that, compared with existing methods, both classification performance and prediction stability are consistently improved.
- 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.”ARAM improves robustness to appearance heterogeneity by constructing a low-magnification context proxy from stable 20× local structures, rather than directly using unstable 5× appearance features. “ .Why choose 20×, rather than 10× or 40×？The authors do not explain the reason. 2.Some hyperparameters in the paper are not clearly specified, such as the number of clusters Lt and the number of selected instances Nt in the TCBM module. 3.In the visualization results section, why were only three comparison methods selected for visualization instead of including all methods? 4.The authors should clarify why ABMIL was selected as the classification backbone, rather than adopting newer multi-instance classification models.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
1.The paper identifies tissue-composition and appearance heterogeneity as two major sources of slide-to-slide variation in text guided few-shot WSI classification. Accordingly, it designs two modules—TCBM and ARAM—to specifically address these two issues. The proposed method is well-aligned with the problem formulation. 2.The method is evaluated on three datasets and compared against nine existing approaches. The experimental results demonstrate consistent and significant improvements over the state-of-the-art.
- 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 authors have addressed my primary concerns, and the methodology is presented with clear logic. While the experimental section could still be further enhanced, the current revision is acceptable.
Author Feedback
We thank the reviewers for their constructive feedback and valuable suggestions. To R1 The 5× magnification provides broader tissue-level context, which helps capture global tissue distribution and structural relationships. The results in Fig. 1 also show that introducing 5× context improves performance compared with using 20× alone, indicating the benefit of low-magnification information. However, directly using raw 5× features may also introduce staining variation, artifacts and coarse appearance heterogeneity. Therefore, ARAM does not simply fuse 5× and 20× features, but constructs a more stable low-magnification context proxy from 4×4 20× local structural units. The fact that direct 5×+20× methods in Fig. 1 still perform worse than ARAM further supports this design motivation. We will clarify this point in the revised manuscript.
To R2 1.The main contribution of this work is not to propose a universal MIL framework for both binary and multi-class classification, but to identify that inter-slide heterogeneity is an important factor causing performance degradation in text-guided few-shot MIL. We further decouple such heterogeneity into substantial variations in tissue composition and appearance across different slides, and design TCBM and ARAM to mitigate them respectively. The experimental results validate the effectiveness of the proposed method. Regarding multi-class classification, we agree that it is a valuable extension and will explore it in future work. 2.The 4/8/16-shot settings are commonly used in few-shot learning, corresponding to extremely limited, moderate, and relatively sufficient few-shot supervision, respectively. 3.For the figure issue, we will improve Figure 2 and its caption in the revised manuscript to make the overall pipeline and module functions clearer.
To R3 1.We choose 5× and 20× because they are commonly used magnifications in WSI classification and provide complementary information: 20× preserves cellular morphology and local tissue patterns, while 5× provides broader tissue-level context and global tissue distribution. In contrast, although 40× contains finer cellular details, it introduces significantly higher computational cost and is less commonly adopted than 20× in existing WSI MIL methods. Compared with 5×, 10× has a smaller field of view and is less effective in capturing global tissue organization and contextual information. Therefore, we adopt the 5×/20× combination as the default setting in this work. 2.Regarding experimental parameters, the main hyperparameters, such as (L_t) and (N_t), have already been reported in the experimental section, and we will provide more detailed implementation settings and hyperparameters in the revised manuscript. 3.For visualization, we select TOP, MSCPT, and MOC because they represent text-guided few-shot WSI classification, multi-magnification few-shot modeling, and recent few-shot WSI SOTA methods, respectively, covering the main types of relevant baselines. 4.ABMIL is used as the backbone because it is a classic and stable MIL framework, which allows us to better isolate and evaluate the contributions of TCBM and ARAM. Meanwhile, HAT itself is not dependent on ABMIL and can be combined with other MIL aggregators.
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.
The reviewers generally agree that the paper addresses an important problem in few-shot WSI classification, particularly focusing on inter-slide heterogeneity. The proposed approach is well-motivated, and the experimental results consistently demonstrate improvements over existing methods. However, several concerns remain. The main issue is the limited experimental validation of key design choices, including the lack of comparisons with alternative formulations (e.g., directly improving low-magnification features) and insufficient justification of hyperparameters and architectural decisions (e.g., magnification level, clustering design). In addition, there are questions regarding the novelty and necessity of individual components, as some modules may appear relatively incremental. The final decision will depend on how convincingly these concerns are addressed.
- 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 an important problem in text-guided few-shot WSI classification by explicitly focusing on inter-slide domain shift caused by tissue composition and appearance heterogeneity. Reviewers appreciated the clear motivation, the careful analysis of slide heterogeneity, and the proposed TCBM and ARAM modules, which effectively improve robustness and prediction stability across multiple datasets and shot settings. Several concerns were initially raised regarding the limited evaluation settings, parameter choices, and the novelty of certain components. However, the rebuttal clarified many of these issues and provided reasonable justification for the design choices and experimental protocol. Multiple reviewers acknowledged that the proposed framework is logically well motivated and practically effective. Overall, while there remain some limitations that could be further explored in future work, the paper presents a meaningful and practically valuable contribution to robust few-shot pathology learning. I therefore support acceptance.
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.
The paper got two accept and one reject after rebuttal. I recommend acceptance for the paper. I think the concerns of the reviewer, who gave reject, are somehow addressed, including the 4,8,16-shot setting in few-shot learning and the figure caption. While generalization to multi-class tasks can not be responded without additional experiments.
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 paper identifies tissue composition and appearance heterogeneity as two major sources of slide-to-slide variation in text-guided few-shot WSI classification. To address these challenges, it proposes two dedicated modules, TCBM and ARAM. Most of the reviewers’ concerns have been addressed, and although the experimental evaluation could be further strengthened, particularly through a more thorough assessment of hyperparameter sensitivity, the current contribution is significant and well-motivated, and is suitable for presentation at MICCAI.