List of Papers Browse by Subject Areas Author List
Abstract
Fusing whole-slide images (WSIs) and magnetic resonance imaging (MRI) provides a comprehensive perspective from both microscopic and macroscopic levels, and could potentially improve the accuracy of human epidermal growth factor receptor 2 (HER2) status prediction. However, directly applying existing fusion methods to such disparate modalities faces two major limitations. First, they often directly interact across these highly heterogeneous modalities, failing to adequately consider the huge domain gap between WSI and MRI. WSI and MRI differ fundamentally in imaging perspectives, information density, and dimensionality, making their correspondence inherently ambiguous. Second, they fail to effectively resolve modality imbalance caused by differences in information density through mere feature-level strategies. WSI contains substantially richer information and usually dominates the fusion process, causing models to inadequately learn from the non-dominant MRI modality. Consequently, both issues limit the model’s ability to leverage multimodal information, sometimes even yielding performance inferior to unimodal baselines. To address these issues, we propose a novel Imbalance-Aware Distributional Alignment (IADA) framework for fusing WSI and MRI via distributional representations and dynamic gradient modulation. Specifically, a heterogeneous modality alignment strategy with distributional representations is introduced to mitigate the domain gap by indirectly characterizing and promoting semantic consistency across modalities. Furthermore, a dynamic gradient modulation module is incorporated to mitigate modality imbalance by adaptively suppressing the gradient contribution of the dominant modality, enabling both modalities to be fully leveraged in the fusion process. Extensive experiments on data from three medical centers demonstrate the superior generalization and effectiveness of the proposed method.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4758_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/embers819/IADA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HuaXim_ImbalanceAware_MICCAI2026,
author = { Huang, Ximeng AND Tang, Wenjie AND Wu, Yeli AND Cai, Jinhui AND Liu, Qingyu AND Wu, Weiwen AND Guo, Yuan AND Zhang, Jianjia},
title = { { Imbalance-Aware Distributional Alignment of Heterogeneous Modalities for HER2 Status Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper addresses an important medical problem, and collecting large-scale multimodal MRI–WSI data is highly challenging, making this dataset a valuable contribution.
- 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 addresses an important and clinically relevant problem by exploring multimodal fusion of WSI and MRI for HER2 status prediction. The proposed framework is well-motivated, particularly in tackling domain gap and modality imbalance, and the combination of distributional alignment and dynamic gradient modulation is intuitive. The method is evaluated on a multi-center dataset with consistent performance improvements over several baselines.
- 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 experimental validation is not sufficiently comprehensive. In particular, the dynamic gradient modulation module lacks stability analysis and sensitivity evaluation, making its robustness unclear. In addition, although the paper introduces “distributional alignment,” the approach essentially reduces to KL-based alignment, which is a commonly used technique, and the necessity and advantage of modeling features as distributions are not well justified. Overall, the methodological novelty is limited and the design choices are not thoroughly validated.
- 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.
(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?
I give a weak reject recommendation because, despite the reasonable motivation and empirical improvements, the methodological novelty is limited and the experimental validation is not sufficient to convincingly justify the proposed design choices.
- Reviewer confidence
Very confident (4)
- [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 author’s response has not alleviated my concerns about the paper.
Review #2
- Please describe the contribution of the paper
This paper proposes IADA, which achieves modal alignment by modeling WSI and MRI features in a distributed space, while dynamic gradient modulation adjusts the relative contributions of WSI and MRI. The paper collected data from three independent medical centers to evaluate the model’s performance in predicting HER2 status, presenting results from both the validation set and an independent test set.
- 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 methodology is well-defined and proposes clear, effective solutions to the bottlenecks in cross-modal fusion. 2) The performance evaluation takes into account both single-modality and multimodal fusion approaches.
- 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 paper’s Dynamic Gradient Modulation is applied to individual training batches, and since the authors set the batch size to 2, this ratio may be sensitive to noise and may not reliably reflect true modal importance. 2) Although the authors conducted extensive performance comparisons, their choice of baselines did not include comparisons with other HER2 prediction methods, nor did they include other multimodal medical imaging methods (only HE+Gene was observed). Furthermore, among single-modality WSI-only methods, the paper only includes RRT-MIL (2024) as the most recent study; newer research has not been incorporated into the discussion, and there is insufficient evidence to demonstrate that Fusion outperforms WSI-only approaches. 3) Are the figures in Table 1 averages or one-off results? Including standard deviations might provide a more comprehensive understanding of the results.
- 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.
(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 proposes a promising multimodal fusion approach that could lead to a breakthrough in performance; however, current experiments are unable to demonstrate the method’s superiority. Although the study collected data from 866 patients across three centers, it was not validated on a public dataset, which weakens the persuasiveness of the results.
- 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 provided a targeted and largely satisfactory response to my three core concerns:
Experimental completeness and baselines (Q1–Q3 combined): They clarify that 21 methods were compared in total (including HER2 prediction, recent WSI‑only methods like MambaMIL and AttriMIL, and multimodal fusion methods like M2Fusion and HARM3Fusion), and only the most competitive were reported due to page limits. They also confirm that Table 1 reports 5‑fold averages, with standard deviations already computed and to be added in the final version. This resolves my concerns about insufficient baselines and missing statistics.
DGM robustness and batch size (Q4): They explain that batch size = 2 is the hardware limit on a 48GB GPU, which is larger than common WSI settings (e.g., WiKG and CHIEF use batch size 1). They further argue that DGM modulates gradients only slightly per batch, with a smooth modulation curve (Fig. 2) that demonstrates robustness. This is an acceptable justification.
Lack of public dataset validation (Q5): They acknowledge that high‑quality paired WSI‑MRI public datasets are scarce, and note that their 3‑center, 866‑patient dataset is already substantial. They also commit to releasing both code and data publicly. While this does not fully replace public dataset validation, the open‑source commitment significantly improves reproducibility and persuasiveness. I accept this as a reasonable trade‑off.
The rebuttal has successfully resolved my major concerns, particularly the issues of baseline completeness and result transparency. The method is well motivated, and the multimodal fusion problem is clinically important. I therefore raise my score from Weak Reject to Weak Accept.
Review #3
- Please describe the contribution of the paper
The authors propose a novel Imbalance-Aware Distributional Alignment (IADA) framework that fuses WSI and MRI through distributional representations and dynamic gradient modulation.
- 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.
To address the huge domain gap between WSI and MRI, the authors propose a novel distributional cross-modality alignment strategy that models both modality features as distributional representations and enforces their semantic consistency.
To effectively mitigate modality imbalance, the authors incorporate a dynamic gradient modulation module that adaptively adjusts the contributions of WSI and MRI during optimization, enabling more effective use of fused information.
- 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.
Although the results seems good to me, the complexity and statistical significance should be discussed.
Could the authors provide further evidence of the advantage of combining WSI and MRI?
- 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?
See above parts.
- 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 rebuttal has addressed my main concerns about complexity and statistical significance.
Author Feedback
R1 Q1 Robustness of Dynamic Gradient Modulation (DGM) DGM’s robustness can be demonstrated by the theory and experiment: (1) Theoretically, DGM modulates gradients very slightly per batch and gradually becomes obvious over many iterations and epochs due to the small learning rate(1e-4) and the gradient modulation coefficient∈(0,1] in Eq.(7). (2) Experimentally, DGM’s modulation effect curve in Fig. 2 generally shows a smooth trend with as small fluctuations as the model without it, indicating its robustness.
Q2 Necessity of Modeling Features as Distributions This design is motivated by the huge WSI-MRI domain gap: they differ greatly in imaging perspective, information density, and dimensionality, making their correspondences ambiguous and hard to model. Modeling features as distributions can better capture their implicit alignment rather than directly modeling their explicit correspondence. Its necessity has also been validated in Table 2.It improves AUC both in internal(by 1.1%) and external(by 3.4%) tests over the SOTA explicitly modeling methods.
Q3 Methodological Novelty and Using KL Alignment Our novelty is bridging the WSI-MRI domain gap by introducing distributional alignment. This strategy is specifically proposed by considering the challenging nature of the problem. As far as we know, this is the first work to use this strategy for addressing the problem. The commonly used KL alignment fits our task well, and KL itself has not been claimed as our novelty. Will clarify this.
R2 Q1 Setting Batch Size to 2 We set the batch size to 2 due to the huge size of WSI samples(usually 1–5 GB), which was already the maximum on advanced NVIDIA RTX A6000 GPU with 48 GB memory. In fact, this is larger than common WSI settings, e.g., WiKG(CVPR24) and CHIEF(Nature24) both use batch size 1.As mentioned in R1Q1, DGM clearly shows robustness with this setting. Will clarify this to avoid confusion.
Q2 Competing Methods As many as 21 competing methods are reported in our study. Although other recent methods were also compared, we only reported the most competitive ones due to page limit. Our model achieves higher AUC than HER2 prediction methods TransformerMIL(Commun Med24), recent WSI-only methods MambaMIL(MICCAI24) and AttriMIL(MedIA25) with large improvement margins (>10%), and relevant multimodal medical fusion methods M2Fusion(MICCAI24) and HARM3Fusion(MICCAI25) with considerable improvement margins (>5%).
Q3 Are the Figures Averages or One-off Results? The results in Table 1 are averaged over 5 folds. The stds of all the methods involved in comparison were already calculated, but omitted due to the table space limit. Will add them in the final version.
Q4 Public Dataset High-quality, large-scale paired WSI-MRI public datasets are very scarce, if any. Our dataset contains as many as 866 patients collected across 3 centers, and our comprehensive validation with both internal and external tests can well validate the effectiveness of the proposed method. Both the dataset and code will be publicly released to ensure persuasiveness.
R3 Q1 Complexity and Statistical Significance Our fusion module is lightweight, with only 1.054M parameters(6.85% of the backbone). The AUC improvements are statistically significant over both the best WSI baseline RRT-MIL(P<0.001 in internal test and P=0.031 in external test) and the best MRI baseline DenseNet(P<0.001&P<0.001). Due to page limit, these details were omitted.
Q2 Further Evidence for the Advantage of Combining WSI and MRI Besides the statistically significant performance improvement, combining WSI and MRI demonstrated: (1) Stronger generalization. In the external test with 3-center dataset, our model outperforms the WSI SOTA RRT-MIL by 5.7% and MRI SOTA DenseNet by 18.4%. (2) Higher stability. Our model achieved an AUC std of only 3.1%, lower than RRT-MIL by 1.2% and DenseNet by 1.7% in internal test.
Q3 Code Availability Both dataset and code will be publicly released.
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 paper addresses a clinically relevant HER2 prediction problem and presents a valuable multi-center MRI–WSI dataset. However, the reviews lean negative, with two Weak Rejects and one Weak Accept, so a rebuttal is needed before a final decision. The main concerns are that the proposed distributional alignment appears to rely mainly on KL-based alignment, and the necessity of modeling features as distributions is not sufficiently justified. The dynamic gradient modulation module may also be unstable with a batch size of 2, and the paper lacks sensitivity or stability analysis for this component. The baseline comparison is incomplete, with limited comparison to HER2 prediction methods, recent WSI-only methods, and relevant multimodal medical fusion models. The authors should also clarify whether the reported results are averaged over multiple runs, provide standard deviations or confidence intervals, and give stronger evidence that WSI+MRI fusion offers a clear advantage over single-modality approaches.
- 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 HER2 prediction problem and provides a valuable multi-center paired WSI–MRI dataset. The rebuttal reasonably addresses concerns about DGM robustness, baseline completeness, statistical significance, result averaging, and the benefit of WSI+MRI fusion. Although the methodological novelty is moderate and the distributional-alignment claim should be presented carefully, two reviewers support acceptance after rebuttal, so I consider the paper slightly above the threshold.
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.
This paper proposes an imbalance-aware distributional alignment framework for HER2 status prediction using WSI and MRI data, supported by a multi-center multimodal dataset. Reviewers appreciate the clinical relevance of the problem and the value of the collected dataset. After rebuttal, most concerns such as result transparency, statistical significance, model complexity, and the advantage of WSI-MRI fusion are reasonably addressed. The revision should incorporate these clarifications and quantitative details into the manuscript, while clearly stating the remaining limitations regarding distributional modeling justification and the stability of dynamic gradient modulation.
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.
While fusing WSI and MRI is an interesting and novel direction, this submission has notable limitations. The experiments in the current manuscript are somewhat insufficient; although the authors provided additional results during the rebuttal, I am concerned about the space constraints to properly integrate them into the camera-ready version. More fundamentally, I question the direct alignment of MRI and WSI, as it remains unclear whether sufficient shared semantic information exists between the two modalities to justify this approach. Despite these methodological doubts, the novelty of the topic leads me to lean toward a weak accept. However, this is a very borderline case, and a decision to reject would also be entirely reasonable.