Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Accurate classification of cell types in H\&E whole-slide images requires integrating evidence across scales, from cell morphology to the local micro-environment and a broader tissue context. However, existing models rely on a fixed field-of-view that frequently misses decisive out-of-view evidence or dilutes local evidence with irrelevant surroundings. We therefore propose a mixture-of-experts (MoE) framework that retrieves multi-field-of-view evidence via query-guided attention pooling, allowing the model to select the right visual cues across views for cell type classification. To ensure that experts effectively collaborate, we deploy an uncertainty-aware router and train the model with a responsibility-driven objective, assigning per-sample credit based on ground truth support to drive complementary specialization across the experts. We validate on public and in-house spatial-transcriptomic-derived datasets of over 400k cell-centered patches, covering lung and skin tissues and diverse cell types, where our method outperforms strong baselines. Finally, the learned pooling weights provide intuitive evidence maps; a small pilot review by a practicing pathologist suggests that these highlighted regions show qualitative agreement with human pathological reasoning.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4161_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yrzzz/TRACE

Link to the Dataset(s)

Xenium skin dataset: https://www.10xgenomics.com/datasets/xenium-prime-ffpe-human-skin Xenium lung dataset: https://www.10xgenomics.com/datasets/xenium-human-lung-cancer-postxenium-technote

BibTex

@InProceedings{YuaRui_Learning_MICCAI2026,
        author = { Yuan, Ruizhi AND Zhao, Chongyue AND Liu, Tianhao AND Wang, Qian AND Yu, Zeqiu AND Tang, Lu AND Huang, Heng AND Chen, Wei},
        title = { { Learning Where to Look: Pathologist-Inspired Multi-field-of-View Evidence Retrieval for Cell Type Classification in H&E } },
        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 mixture-of-experts framework, in which individual experts perform attention-based, token-level retrieval of classification evidence across nested, cell-centered views. The proposed method aims to address the limitations of existing classification approaches that strictly rely on fixed fields of view, which often lead to the omission of decisive contextual evidence outside the predefined boundaries. By bridging this gap, the method aligns the automated cell classification paradigm more closely with the actual diagnostic workflow of human pathologists.

  • 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 framework elegantly assigns visual features across three distinct scales (cellular, local, and contextual) to three specialized experts. Recognizing the inherent differences in visual information across these scales, the model employs a tailored and highly reasonable processing strategy: the first expert explicitly computes the embedding only from tokens that fall inside the target cell, while the remaining experts aggregate their respective views via cosine-attention based retrieval. This scale-aware Mixture-of-Experts design is not only highly intuitive but also methodologically innovative for pathology image analysis. 2.The proposed “responsibility-driven” training mechanism represents an original and effective way to utilize the data. By calculating a per-sample responsibility weight based on each expert’s support for the ground-truth class, and updating the experts’ parameters proportionally, the approach maximizes the utility of the annotations. This operation significantly amplifies the supervisory signal of the cell classification labels, as the ground truth dynamically guides the parameter updates through two distinct channels: the final classification output and the experts’ supportive confidence in the predictions.

  • 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 performance metrics used in Table 1 are limited solely to the F1-score. Cell type classification is notoriously plagued by severe class imbalance. Therefore, relying only on F1 and Accuracy in Table 2 is insufficient. Standard evaluations in this domain mandate the inclusion of Precision, Recall, and potentially AUC, to transparently demonstrate whether the model suffers from over-predicting the majority class or failing to detect minority classes. 2.The model is evaluated on only Lung and Skin. Given that the paper proposes a fundamental, pathologist-inspired paradigm shift for H&E cell classification, evaluating on just two tissue types is insufficient to claim generalizability. The authors should evaluate their model on established multi-organ datasets such as PanNuke or CoNSeP to prove the robustness of the multi-FoV retrieval mechanism across different histological patterns. 3.The paper completely lacks an analysis of FLOPs, parameter counts, and inference speed. Without this, the clinical feasibility of deploying this model for whole-slide scale analysis remains highly questionable.

  • 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.

    (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?

    I am leaning towards a Weak Accept (4) for this submission. The paper introduces a highly intuitive, pathologist-inspired Mixture-of-Experts architecture that effectively addresses the fixed-FoV bottleneck in cell classification. The scale-aware design for nested views, coupled with the original responsibility-driven training mechanism, demonstrates strong methodological innovation. The approach is theoretically sound and aligns well with the clinical diagnostic workflow. However, my enthusiasm is dampened by two critical limitations that prevent a higher score: 1.Claiming a fundamental paradigm shift requires validation across diverse histological patterns. Evaluating only on Lung and Skin is a significant limitation. 2.Relying solely on the F1-score is inadequate for cell type classification, which is notoriously imbalanced. The absence of Precision, Recall, and AUC masks potential predictive biases. 3.The multi-FoV token-level retrieval mechanism inherently risks introducing high computational overhead. The complete lack of analysis regarding FLOPs, parameter counts, and WSI-level inference speed leaves the clinical feasibility of this model highly questionable.

  • 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

    Authors propose TRACE, a multi-field-of-view (FOV) framework designed to address limitations of existing methods that rely on a fixed FOV for cell type classification in H&E whole-slide images. The model leverages multiple visual contexts to select the most informative cues across views, improving cell-type prediction. It is evaluated on both public and in-house spatial transcriptomics-derived datasets. Additionally, a small pilot study with a practicing pathologist is conducted to assess whether the model’s highlighted regions align with human pathological reasoning.

  • 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.
    • Model design and novelty: TRACE is a novel contribution, particularly through its multi-expert architecture over nested multi-field-of-view inputs. The integration of uncertainty-aware routing and probability-space fusion provides a principled mechanism for combining expert predictions in a more adaptive and robust manner.
    • Clinical relevance: The qualitative agreement is a valuable addition, suggesting that the learned evidence maps may align with human pathological reasoning and supporting the potential clinical interpretability of the method.
  • 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.
    • Limited methodological transparency: While the model components are described, the method section remains relatively concise, which may hinder reproducibility and understanding. Some variables and acronyms are not explicitly defined, and a more structured presentation of the MoE routing, multi-FOV aggregation, and responsibility-driven objective would improve clarity.
    • Inconsistencies in reported numerical results: Inconsistencies are present in reported values across tables and text (e.g., differences in F1 scores for specific cell types such as Fibroblasts and Vasculature in lung, and slight inconsistencies in ablation improvements such as 0.67 → 0.68 on skin). A careful consistency check across all reported numbers is recommended.
    • Limited fairness of experimental comparison: The evaluation is not fully fair due to asymmetry in input information. TRACE uses paired multi-FOV (local + context) inputs, while it remains unclear whether baselines are provided with equivalent multi-scale or expanded-context views. In addition, baselines do not appear to be strengthened with comparable multi-scale augmentation, which may disadvantage them relative to TRACE. A fairer comparison can attribute primarily the model architecture rather than differences in available spatial context.
    • Unclear optimization and convergence justification: Convergence cannot be confirmed from the reported setup. While a constant learning rate with large batch size and fixed training duration can be effective in a frozen-backbone setting, the absence of learning rate scheduling or convergence analysis leaves the optimization procedure insufficiently justified. Furthermore, although expert heads and the gating network are stated to be optimized, the exact optimization strategy is not fully specified.
    • Inconsistencies in qualitative evaluation: The qualitative evaluation is not fully clear regarding the number and role of evaluators; the abstract refers to a practicing pathologist, while the results section suggests multiple readers, making the exact setup of the qualitative agreement ambiguous.
  • 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 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
    • Reference replacement: Reference 5 should be replaced with the full journal citation to ensure completeness and consistency. The correct version is: Hörst, Fabian, et al., “CellViT: Vision Transformers for Precise Cell Segmentation and Classification”, Medical Image Analysis, 94 (2024), 103143.- Limited clarity in Figure 2: In Fig. 2, the context view would benefit from also displaying the green contour to improve visual consistency between local and contextual representations. In addition, the method used to compute the visualized attention or evidence maps is not clearly described, and a more explicit explanation of the underlying technique would improve interpretability.
    • Minor typographical issues in the manuscript: typos should be corrected to improve readability. For example, in the Method section, phrases such as “an detected cells” require correction, along with a general proofreading pass to remove similar minor grammatical errors throughout the text.
  • 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 thank the Authors for their work on TRACE. The paper is overall interesting and addresses an important problem in histopathology representation learning. However, some issues limit the clarity and strength of the contribution. The method description is relatively concise but lacks sufficient detail on key components. There are also inconsistencies in reported numerical results across tables and text. The experimental comparison seems not fully fair. In addition, convergence and optimization are not sufficiently supported. Finally, the qualitative evaluation lacks clarity regarding the number and role of evaluators, with inconsistencies between abstract and results sections.

  • 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 believe that most of the concerns raised in the revision have been adequately addressed, and I appreciate the Authors’ efforts. Therefore, I recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    This manuscript proposes a mixture of experts framework that retrieves multi-field-of-view evidence, mimicking the diagnostic process of a pathologist. The framework consists of cell, local, and context-based views, where tokens are extracted from each viewpoint. Furthermore, the overall structure is well-organized in accordance with MICCAI guidelines.

  • 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 framework consists of cell, local, and context-based views, where tokens are extracted from each viewpoint. The integration of these multi-faceted perspectives into the model design is impressive and provides a highly reasonable approach to pathological analysis.

  • 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) Dataset Xenium and Visium HD were utilized to generate the labels for the dataset. While both methods are state-of-the-art, there remains a concern regarding whether they provide sufficiently robust ground truth for an AI model. The model trained on this data may inherit the intrinsic limitations and biases associated with these spatial technologies during prediction and analysis.

    2) Method -crop size The cropping pixel sizes are fixed at 244 * 244 and 1024 * 1024 to generate tokens across different viewpoints. However, this fixed-scale approach may introduce bias, potentially favoring specific image characteristics or structures.

    3) Experimental results Although the F1 score is used to demonstrate the efficacy of the proposed approach, relying solely on it can skew results and mask underlying weaknesses, such as class imbalance. Therefore, precision and recall should be reported alongside the F1 score. To establish greater confidence in the results, additional metrics—including precision-recall curves, class-wise performance, and confidence intervals—should be provided.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    As described in the previous comment, several concerns should be addressed to improve the completeness of the manuscript.

  • 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.

    I still think that treating Xenium or Visium HD data as absolute ground truth and employing it as label data for strong supervised learning carries substantial risks, given the inherent technical noise and capture inefficiencies of spatial platforms. The authors must explicitly address these risks and limitations within the paper.



Author Feedback

We thank the AC and reviewers for recognizing the importance, novelty, and clinical relevance of our method, as well as for identifying the main issues to address. We focus our response on dataset scope, evaluation metrics, comparison fairness, methodological clarity, and consistency.

1) Limited dataset scope (AC/R1). We agree that broader multi-organ validation would further strengthen generalization. We choose lung and skin as primary applications intentionally, because they span major cell types and include both morphology-driven and context-dependent cases. Our pathologist also has the best expertise on these two important tissues, making the evidence-map assessment clinically informed rather than generic. Our public Xenium and in-house Visium HD data include 10 WSIs and >400k cells across different data sources and staining appearances, providing initial support for generalization beyond a single cohort. Broader multi-organ validation remains important future work.

2) Metrics under class imbalance (AC/R1/R3). Reviewers asked for additional metrics to better assess class imbalance. We agree that precision/recall curves and confidence intervals would further support the results and we did have those results. However, due to space limits and similar results, we used F1-based metrics because F1 balances precision and recall, penalizing both over-prediction and missed detections. For class imbalance concern, Table 1 reports per-class F1, so performance on each class is shown explicitly rather than hidden by majority classes; Table 2 reports macro-F1, which averages class performance equally instead of weighting classes by their frequency.

3) Fairness of the comparison (AC/R2). R2 raised whether TRACE benefits from extra input views compared with fixed-FOV baselines. TRACE tests whether cell typing benefits from evidence beyond a fixed crop; local/context views are therefore the core design under evaluation, not unfair extra information. For fixed-FOV baselines, we used their original architectures and input settings instead of forcing unsupported multi-view inputs; naively enlarging crops would violate their training setup and may reduce performance by diluting local evidence. We also include NuClass as a comparable multi-view baseline using the same views/input scale as our method.

4) Method/optimization clarity concern (AC/R2). We agree the presentation can be clearer, but the manuscript already provides the core TRACE pipeline and math details. The issue is clarity of presentation, not a missing method component. We will explicitly define variables/acronyms and present model workflow step by step. For optimization, all modules except the feature extractors are trained jointly with AdamW using a fixed learning rate, as only lightweight task-specific networks are updated. The epoch number is the maximum budget, with best-checkpoint selection based on validation performance and early stopping.

5) Consistency and clarity fixes (AC/R2). Tables report the final values; text inconsistencies are synchronization errors and will be corrected. The pilot review involved one pathologist and one medical student; the abstract emphasized the pathologist due to space and expertise. We will also clarify evidence-map visualization, update the CellViT citation, and fix typos.

6) R3 factual corrections. We respectfully disagree with two R3 criticisms. First, questioning Xenium/Visium HD labels as insufficient ground truth is unsupported. Spatial transcriptomics technologies are among the strongest available methods for obtaining cell-type labels with spatial localization and direct H&E alignment. Our pathologist further supports label reliability. Second, the crop-size criticism is not correct: the manuscript uses 224×224, not 244×244, for the local crop. The “fixed-scale bias” claim also misreads our method, which combines cell/local/context views with token-level retrieval. Thus, these points should not be treated as method flaws.



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 proposes a multi-field-of-view mixture-of-experts framework for cell type classification in H&E images, inspired by pathologist workflows. The reviewers’ scores are 4, 3, and 2, indicating mixed opinions. The reviewers agree that the problem is important and that the proposed framework is conceptually interesting, particularly in leveraging multi-scale contextual evidence. However, concerns remain regarding the completeness and fairness of the experimental evaluation, including limited dataset diversity, reliance on a restricted set of evaluation metrics, and potential imbalance in comparison settings. In addition, clarity and consistency in the methodological description and reported results could be improved. The AC therefore recommends inviting the paper for rebuttal.

    The authors are encouraged to carefully address the reviewers’ comments in the rebuttal, particularly regarding strengthening the experimental validation (e.g., broader datasets and evaluation metrics), clarifying the methodological details, and ensuring fair and consistent comparisons with baselines.

  • 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 proposes TRACE, a pathologist-inspired multi-field-of-view evidence retrieval framework for cell type classification in H&E whole-slide images. The method uses nested cell-, local-, and context-level views, with a mixture-of-experts design that retrieves token-level evidence across different fields of view. The goal is to overcome the limitations of fixed-FoV cell classification by allowing the model to selectively use local cellular morphology and broader tissue context.

    The initial reviews were mixed, with scores of 4 / 3 / 2.Reviewer 1 was positive and appreciated the intuitive scale-aware mixture-of-experts design and the responsibility-driven training mechanism, while noting concerns about limited evaluation metrics, limited tissue scope, and missing computational analysis. Reviewer 2 found the problem and model design interesting but raised concerns about limited methodological transparency, inconsistencies in reported numbers, fairness of comparisons, optimization details, and clarity of the qualitative pathologist-alignment evaluation. Reviewer 3 was more critical, raising concerns about the use of Xenium/Visium HD-derived labels as ground truth, fixed crop sizes, and insufficient reporting of precision/recall and uncertainty under class imbalance.

    The rebuttal addressed several of these points. The authors clarified that the dataset includes public Xenium and in-house Visium HD data across 10 WSIs and more than 400k cells, while acknowledging that broader multi-organ validation would further strengthen generalization. They explained that per-class F1 and macro-F1 were already reported to mitigate concerns about class imbalance, and that precision/recall curves and confidence intervals could improve transparency. The rebuttal also clarified the comparison protocol, arguing that the local/context views are the core design of TRACE rather than unfair additional information, and that NuClass was included as a comparable multi-view baseline. The authors also addressed method-presentation and optimization concerns, clarified inconsistencies in reported numbers, explained the pilot reviewer setup, and corrected the crop-size issue raised by Reviewer 3. Some limitations remain. The validation is still limited to lung and skin, and the generalizability of the multi-FoV retrieval mechanism across broader histological patterns is not fully established. The use of spatial transcriptomics-derived labels is reasonable for scalable H&E cell typing, but the final paper should explicitly discuss the potential noise and limitations of Xenium/Visium HD labels rather than treating them as absolute ground truth. The computational cost and WSI-scale feasibility of the multi-FoV retrieval design should also be better documented. In addition, precision/recall curves, confidence intervals, and clearer evidence-map visualization would improve the final version.

    After rebuttal, the effective reviewer summary is A / A / R, treating the no-update weak accept as unchanged. Reviewer 2 moved from weak reject to accept, stating that most concerns had been adequately addressed. Reviewer 3 maintained reject, mainly because of concerns about the reliability of Xenium/Visium HD labels as ground truth. I place some weight on this remaining concern, but I do not view it as a fatal flaw because all methods are evaluated under the same label source and because spatial transcriptomics-derived labels are among the strongest scalable supervision sources currently available for this task.

    I recommend Accept. The paper is borderline but above the acceptance threshold. The core idea of multi-field-of-view evidence retrieval for H&E cell typing is well motivated, clinically relevant, and reasonably novel. The rebuttal clarified several important methodological and evaluation concerns, and the post-rebuttal reviewer balance is positive. The final version should clearly acknowledge the limited tissue scope, label-source limitations, computational cost, and remaining evaluation/reporting limitations.



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.

    After the rebuttal, the ratings contain one accept, one weak accept and one reject. Most of the comments were about experimental evaluations, which were addressed in the rebuttal. The negative reviewer is still concerned about the label quality of the training dataset. This could be a separate problem to the objective in this paper.



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 proposes a novel and clinically motivated multi-field-of-view MoE framework for H&E cell type classification, and reviewers generally recognized the importance of the problem, the pathologist-inspired design, and the promising results. While concerns remain about dataset breadth, metric reporting, and label uncertainty from spatial transcriptomics, the rebuttal adequately clarified several key issues, Reviewer 2 changed to Accept, and the remaining weaknesses appear to be limitations rather than fatal flaws; therefore, I recommend acceptance.



back to top