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

Abstract

Pathology Foundation Models (FMs) have shown strong performance across a wide range of pathology image representation and diagnostic tasks. However, FMs do not exhibit the expected performance advantage over traditional specialized models in Nuclei Detection and Classification (NDC). In this work, we reveal that jointly optimizing nuclei detection and classification leads to severe representation degradation in FMs. Moreover, we identify that the substantial intrinsic disparity in task difficulty between nuclei detection and nuclei classification renders joint NDC optimization unnecessarily computationally burdensome for the detection stage. To address these challenges, we propose DeNuC, a simple yet effective method designed to break through existing bottlenecks by Decoupling Nuclei detection and Classification. DeNuC employs a lightweight model to detect nuclei, subsequently leveraging a pathology FM to encode input images and query nucleus-specific features based on the detected coordinates for classification. Extensive experiments on three widely used benchmarks demonstrate that DeNuC effectively unlocks the representational potential of FMs for NDC and significantly outperforms state-of-the-art methods. Notably, DeNuC improves F1 scores by 4.2% and 3.6% (or higher) on the BRCAM2C and PUMA datasets, respectively, while using only 16% (or fewer) trainable parameters compared to other methods. Code is available at https://github.com/ZijiangY1116/DeNuC.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2293_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ZijiangY1116/DeNuC

Link to the Dataset(s)

PUMA dataset: https://zenodo.org/records/15050523 BRCAM2C dataset: https://github.com/TopoXLab/Dataset-BRCA-M2C OCELOT: https://zenodo.org/records/8417503

BibTex

@InProceedings{YanZij_DeNuC_MICCAI2026,
        author = { Yang, Zijiang AND Kuang, Chen AND Fu, Dongmei},
        title = { { DeNuC: Decoupling Nuclei Detection and Classification in Histopathology } },
        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 introduces DeNuC, a framework that decouples nuclei detection and classification to prevent representation degradation in pathology Foundation Models Based on the insight that detection is an “easier” task that converges quickly while classification is harder, DeNuC uses a lightweight model for localization and reserves the FM to extract features at the detected coordinates for classification.

  • 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 observation that joint detection-classification optimization degrades FM representations is a highly valuable diagnostic insight for the computational pathology community. 2.The solution directly addresses the problem without over-engineering. Decoupling the tasks is easy to implement, reproduce, and integrate into existing pipelines. 3.Achieving a 3.6%–4.2% F1 boost while using less than 16% of the trainable parameters of competing methods is a good, practically relevant result. 4.Decoupling allows for pooling detection data across multiple datasets with different class taxonomies, which is a smart and practical side benefit.

  • 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 thoroughly ablates the detection backbones but neglects the classification side. It needs to show what happens if the classification stage uses a non-FM backbone (e. g. , a standard ResNet trained from scratch) or different pathology FMs (e. g. , UNI, CONCH, Virchow). This is crucial to prove that the performance gains come from unlocking the FM, rather than just the decoupling strategy itself. 2.If the FM is only used for classification, why is coordinate-based feature querying better than simply cropping patches around detected nuclei and running a standard classifier? The paper needs to explicitly compare against this straightforward baseline. 3.The assumption that detection is easier than classification holds for the tested datasets, but might break down in tissues with dense lymphocytic infiltrates, high nuclear overlap, or different stains (like IHC). A brief discussion on these edge cases would help.

  • 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.

    (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?

    This paper provides an insightful diagnosis of why pathology FMs struggle with joint detection and classification tasks and offers a simple, highly parameter-efficient solution. The core idea is sound and the results are strong. While there are some gaps in the ablations—specifically regarding the classification branch and a missing crop-and-classify baseline—the fundamental contribution is sound. Thus, I recommend acceptance.

  • 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

    1.This paper proposes DeNuC, a decoupled framework for nuclei detection and classification (NDC) in histopathology images. The key contribution is the identification of a previously underexplored issue: joint optimization of detection and classification leads to representation degradation in pathology foundation models (FMs). The authors demonstrate that FMs, while strong in semantic representation are not well-suited for coordinate regression and joint optimization forces the model to compromise its feature space.

    2.To address this, the authors propose to decouple detection and classification into two separate stages: (1) a lightweight detector for nuclei localization, and (2) a foundation model-based feature extractor with coordinate-guided feature querying for classification. This design allows the FM to preserve its pre-trained representation capability while improving overall task performance.

    3.Extensive experiments on three benchmark datasets (BRCAM2C, OCELOT, and PUMA) show that the proposed method achieves state-of-the-art performance with significantly fewer trainable parameters (≈4.3M, ~16% of competing methods). The paper also includes thorough ablation studies validating the effectiveness of decoupling and training strategies.

  • 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 demonstrates representation degradation in foundation models during joint NDC optimization with convincing empirical evidence. 2.The decoupled design is well-motivated and enables better task specialization while avoiding gradient interference. It is a simple and effective methodology. 3.Consistently achieves notable improvements across multiple benchmark datasets. 4.Delivers state-of-the-art results with a very small model (~4.3M parameters), making it computationally efficient. 5.Comprehensive ablation studies: Provides thorough experimental analysis validating the design choices and strengthening the claims.

  • 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.While the empirical insight is valuable, the core idea of decoupling detection and classification is not entirely new and has been explored in various forms in object detection and medical imaging pipelines. The novelty primarily lies in its application and analysis in the context of foundation models. 2.The claim of representation degradation is supported empirically but lacks deeper theoretical justification. A more rigorous analysis (e.g., gradient interference, feature space shifts) would strengthen the contribution. 3.The framework relies heavily on the accuracy of the detection stage. Although detection performance is high, the paper does not analyze the impact of detection errors on classification performance or overall robustness. 4.Experiments are conducted on three datasets, which are standard but relatively similar in domain. Additional validation on more diverse datasets or cross-domain settings would further support generalization claims. 5.The paper does not explore robustness to noisy inputs, domain shifts, or imperfect detections, which are important in clinical scenarios.

  • 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?

    The paper presents a well-motivated and empirically strong contribution with clear performance improvements and excellent parameter efficiency. The identification of representation degradation in foundation models under joint optimization is an important insight, and the proposed decoupled framework effectively addresses this issue.

    However, the methodological novelty is somewhat limited, and the lack of deeper theoretical analysis and robustness evaluation prevents it from being a clear accept. Despite these limitations, the paper provides meaningful contributions and practical value.

  • 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 #3

  • Please describe the contribution of the paper

    The primary contribution of this paper is identifying that jointly optimizing nuclei detection and classification causes severe “representation degradation” in pathology Foundation Models (FMs) due to inherent task difficulty disparities. To address this, the authors propose DeNuC, a novel framework that decouples the two tasks by employing an ultra-lightweight network for precise point-based localization and leveraging a frozen pre-trained FM for coordinate-guided classification. Extensive experiments across three benchmarks demonstrate that DeNuC significantly outperforms existing State-of-the-Art methods. Notably, this decoupled approach fully unlocks the representational potential of FMs while utilizing only 16% or fewer trainable parameters compared to current models, offering a highly resource-efficient paradigm for histopathology 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.
    • The paper provides a compelling and insightful analysis of the limitations of applying pathology Foundation Models (FMs) to multi-task learning. Specifically, it reveals that jointly optimizing nuclei localization (coordinate regression) and classification causes severe “representation degradation” in FMs due to a substantial disparity in task difficulty and convergence speeds. This finding logically challenges the conventional paradigm of joint multi-task optimization in Nuclei Detection and Classification (NDC) and provides a solid theoretical motivation for the proposed solution.
    • The proposed DeNuC framework offers an elegant, highly effective, and resource-efficient methodology by completely decoupling detection and classification. ** By delegating the relatively simple localization task to an ultra-lightweight point-detection network, the framework avoids redundant computations and allows for highly efficient spatial identification. ** Utilizing a coordinate-guided feature querying mechanism (via bilinear interpolation) allows the frozen pre-trained FM to focus purely on extracting fine-grained semantic representations for classification. This prevents the FM’s feature space from being perturbed by localization gradients, thereby fully unlocking its pre-trained potential.
    • The empirical results demonstrate a remarkable trade-off between performance and computational cost. DeNuC consistently and significantly outperforms existing state-of-the-art methods (including complex architectures like CellViT, SMILE, and MUSE) across three diverse and widely used benchmarks (BRCAM2C, OCELOT, and PUMA). Remarkably, it achieves this superior performance while requiring only 16% (or fewer) trainable parameters compared to competing methods, demonstrating high practical value for real-world clinical deployment where computational resources are often constrained.
    • The paper presents a rigorous and well-structured set of ablation studies that strongly validate the authors’ core claims. ** The experiments on cross-dataset joint training for the detection module effectively demonstrate how a decoupled, lightweight detector can leverage diverse datasets to boost localization robustness without increasing model complexity. ** The comparison of different optimization strategies for the classification network (Linear probing, LoRA, Full fine-tuning vs. End-to-end) provides concrete empirical evidence that isolating the classification task and freezing the backbone is the optimal strategy to prevent the aforementioned representation degradation.
  • 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 paper heavily emphasizes the drastic reduction in trainable parameters (e.g., 4.3M), it appears to conflate training parameter efficiency with overall computational efficiency during inference. Because the proposed framework still relies on a massive, frozen Foundation Model (e.g., UNI2-H) to extract high-dimensional feature maps for every input image, the actual inference latency, memory footprint (VRAM), and FLOPs might still be substantial. To present a truly balanced view of its “resource efficiency,” the authors should provide a comprehensive comparison of inference speed (FPS) and total FLOPs against baseline methods.
    • Incomplete analysis of error propagation from detection to classification. The decoupled design introduces an inherent dependency: classification performance is gated by detection quality, yet the paper provides only aggregate performance metrics. A false negative in detection irrevocably precludes correct classification, while false positives waste computational resources and may produce spurious classifications.
    • The reliance on a purely point-based detection architecture fundamentally restricts the framework from providing nuclei segmentation masks. While point detection is sufficient for counting and basic classification, many critical downstream clinical tasks (e.g., calculating the nuclear-cytoplasmic ratio, assessing nuclear pleomorphism, or spatial transcriptomics alignment) require precise morphological features (shape, area, boundaries). The inability to extract morphological context is a notable trade-off that should be explicitly acknowledged and discussed as a limitation.
    • The experimental setup for the motivational analysis in Figure 1(a) lacks sufficient methodological detail in the main text. It is difficult for the reader to fully grasp how the “Linear Probing” curve was generated dynamically during the joint optimization process (e.g., was a new linear classifier trained from scratch at every epoch solely to probe the evolving feature space?). Clarifying the exact protocol for this specific analysis is necessary to ensure reproducibility and to fully convince the reader of the degradation claim.
    • The coordinate-guided feature querying mechanism (Equation 5) relies on bilinear interpolation from the FM’s feature map. Given that modern FMs (like UNI2-H) are predominantly Vision Transformer (ViT) based, their feature maps are inherently patch-based (e.g., 14x14 or 16x16 pixel patches). For very small nuclei, or those located precisely at the boundaries of multiple patches, bilinear interpolation might inadvertently capture entangled features from neighboring structures or background stroma. The paper would benefit from an analysis of how the patch size or the chosen feature layer of the FM affects the classification performance of nuclei of varying sizes.
  • 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?
    • Valuable Insight into FM Behavior: The authors identify a crucial issue—”representation degradation”—that occurs when FMs are forced to jointly optimize for coordinate regression (localization) and classification. This observation challenges the prevailing multi-task learning paradigm in NDC and provides a strong, logical foundation for the proposed decoupling strategy.
    • The DeNuC framework is conceptually clean. By delegating the simpler localization task to a highly lightweight point-detector and using the frozen FM strictly as a feature extractor for coordinate-guided classification, the method perfectly aligns the architecture with the inherent difficulty of each task.
    • The experimental validation is robust. Achieving state-of-the-art results across three diverse benchmarks (BRCAM2C, OCELOT, PUMA) while utilizing less than 16% of the trainable parameters of competing methods (like CellViT or SMILE) is a substantial achievement. This demonstrates that high performance in NDC does not strictly require massive, complex, end-to-end architectures.
  • 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



Author Feedback

N/A



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    All reviewers gave rather positive evaluations of the paper. The observed degradation and the proposed framework provide valuable insights into the field. The authors are encouraged to further improve the paper’s quality based on the reviewers’ comments.



back to top