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

Abstract

Nuclei segmentation and classification are fundamental to computational pathology, yet the performance of supervised deep learning approaches is often hampered by the scarcity of large-scale, accurately annotated datasets. Although generative data augmentation offers a promising solution, existing methods remain constrained by inefficient training and limited efficacy on small-scale annotated datasets. To address these issues, we propose a novel multimodal conditioned framework based on optimal transport flow matching (OTFM) to synthesize high-fidelity nuclei mask and pathology image pairs. Our framework employs synchronous multimodal conditioning to efficiently utilize small-scale datasets for generating aligned nuclei mask-image pairs, and adopts few-step sampling to accelerate inference. To ensure biologically plausible mask generation, we integrate cell center points and text descriptions as conditions to preserve spatial layout consistency. Furthermore, self-supervised learning (SSL) feature representations extracted from pathology foundation models are incorporated as additional conditions, providing morphological guidance for synthesizing high-fidelity, accurately aligned pathology images. Extensive experiments on pathology nuclei datasets show that our framework consistently improves performance in nuclei segmentation and classification tasks. Code is available at https://github.com/zhangyn1415/MMCFM-PNDA.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2543_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/zhangyn1415/MMCFM-PNDA

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhaYan_MultimodalConditioned_MICCAI2026,
        author = { Zhang, Yanan AND Wang, Yinuo AND Liu, Mengyang AND Bai, Xiangzhi},
        title = { { Multimodal-Conditioned Flow Matching for Pathology Nuclei Data Augmentation } },
        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

    The paper proposes a multimodal-conditioned flow matching method for synthesizing paired histopathology images and nuclei masks. It conditions generation on cell centers, text prompts, and self-supervised features from pathology foundation models, enabling biologically plausible and high-fidelity data augmentation. Experiments show consistent gains on downstream segmentation and classification tasks across three datasets.

  • 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.Novel conditioning strategy: Effectively combines geometric (centers), semantic (text), and visual (CPathFM features) cues. 2.Clear presentation: Well-written with informative figures.

    1. Outperforms latest GAN/diffusion 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.

    1.The performance may depend on accurate textual and geometric inputs. In practice, cell centers may be noisy (e. g. , from weak detectors), and text metadata might be missing, ambiguous, or mislabeled. The robustness of the model under such imperfect conditions is not evaluated. 2.Although image quality metrics are reported, the paper does not explicitly analyze the diversity of the generated nuclei populations—e. g. , variation in shape, size, texture, or spatial arrangement. High-fidelity but low-diversity synthesis could lead to overfitting in downstream models. Metrics like nucleus-level variance or nearest-neighbor diversity scores would provide valuable insight.

  • 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.

    (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 is a strong, well-executed contribution that innovatively leverages foundation models and flow matching for realistic nuclei data synthesis. Results are compelling and clinically meaningful. I strongly recommend acceptance.

  • 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

    This paper proposes a multimodal-conditioned optimal transport flow matching framework for pathology nuclei data augmentation. The method synthesizes paired nuclei mask and pathology image samples by integrating multiple complementary conditions, including nuclei type maps, cell center points, text embeddings, and SSL feature representations extracted from pathology foundation models. The framework is trained end-to-end and supports few-step sampling, improving both training and inference efficiency relative to diffusion-based alternatives. Experiments on PanNuke, EndoNuke, and Lizard show that the synthesized data improves downstream nuclei segmentation and classification performance across multiple datasets.

  • 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 addresses a practically important problem in computational pathology: data augmentation for nuclei segmentation and classification under limited annotations. This is a relevant and timely problem with clear downstream utility. 2.The proposed framework is technically well structured. The use of multimodal conditioning, including text, nuclei spatial cues, and SSL features from pathology foundation models, is well motivated and aligns with the need to preserve both semantic and morphological fidelity in synthesized pathology images. 3.The use of optimal transport flow matching is appealing from an efficiency perspective. The few-step sampling results and comparison to DDIM-style sampling support the claim that the method is more efficient than diffusion-based alternatives. 4.The experimental evaluation is reasonably comprehensive. The paper includes generation-quality comparison, ablation studies, efficiency analysis, feature visualization, and downstream evaluation on three datasets. This makes the empirical section stronger than many purely generative pathology papers. 5.Most importantly, the synthetic data leads to consistent improvements on downstream tasks, including Dice, IoU, mPQ, and F1, which strengthens the practical significance of the work.

  • 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 methodological novelty is moderate rather than strong. The paper combines existing ingredients, including flow matching, multimodal conditioning, text guidance, and SSL-feature guidance, into a unified framework, but the individual technical components are not fundamentally new. 2.The gains, while consistent, are relatively incremental in several downstream metrics. The paper demonstrates utility, but the magnitude of improvement is not always large enough to clearly establish a major practical advance. 3.The evaluation of generation quality relies heavily on indirect proxy metrics and downstream model performance. Additional qualitative assessment by pathology experts, or stronger analysis of biological plausibility, would improve confidence in the realism of the generated samples. 4.The contribution of each external pretrained model is not fully characterized. Since the framework depends on pathology foundation model SSL features and pathology VLM text embeddings, it would be helpful to better analyze sensitivity to encoder choice and the added complexity/cost of these components. 5.Although the paper claims improved efficiency, the computational cost of training with multiple conditioning modules and pretrained encoders is not fully discussed. A more explicit efficiency comparison, beyond sampling-step count, would strengthen the claims.

  • 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

    This is a solid and well-executed paper on pathology nuclei data augmentation. I appreciate that the paper goes beyond image synthesis quality and demonstrates consistent downstream improvements on segmentation and classification tasks across multiple datasets.

    The strongest aspect of the work is the integration of complementary conditions, especially SSL features from pathology foundation models, into an efficient flow matching framework. The empirical study is also reasonably comprehensive.

    To further strengthen the paper, I encourage the authors to clarify the degree of methodological novelty relative to existing conditional generative approaches, provide deeper analysis of sensitivity to the choice of pretrained encoders, and include additional discussion of biological plausibility and computational cost. Stronger expert-based or pathology-aware qualitative validation would also improve the work.

    Overall, I find the paper promising and relevant to the MICCAI audience.

  • 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 recommend a weak accept because the paper addresses a relevant problem in computational pathology and presents a technically coherent multimodal flow matching framework with solid empirical validation. The integration of SSL features, text guidance, and spatial conditioning is well motivated, and the method consistently improves downstream nuclei segmentation and classification across multiple datasets.

    However, the work is more of a strong integration of existing ideas than a major methodological breakthrough. In addition, some of the gains are incremental, and the evaluation could be strengthened with more pathology-aware realism analysis and a clearer discussion of computational trade-offs. Overall, the paper is solid and likely of interest to the MICCAI community, but it is not at the level of a standout 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 #3

  • Please describe the contribution of the paper

    The paper proposes a multimodal-conditioned flow matching framework (OTFM-based) for generating paired nuclei masks and pathology images. It combines multiple conditioning signals (mask, text, SSL features, cell centroid) to improve controllability, and claims improved downstream performance via generative data augmentation.

  • 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 idea of combining multiple conditioning modalities is reasonable and aligns with current trends in controllable generation.
    • Flow matching seems to be interesting over diffusion method from an efficiency aspect.
    • Multiple datasets are covered in validation.
  • 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 framework introduces multiple conditioning signals, but it is not clear whether all of them are necessary. The analysis is limited, and there is little insight into why this particular combination works or which components contribute most.

    • There are concerns around overclaiming and insufficiently supported evidence, especially in Fig. 4 and Fig. 5.Fig. 4 is not convincing as it only shows selected qualitative samples without comparison to baselines or failure cases. There is no quantitative or structural consistency evaluation to support claims of “high-fidelity” or “biologically plausible” synthesis. Fig. 5 relies on t-SNE visualization to argue distribution alignment and generalization. However, t-SNE is highly sensitive to hyperparameters and distorts global structure. The paper does not provide sufficient details on the setup, and the claim that synthetic data matches real distributions or improves generalization is not justified by this visualization alone.

    • The improvements reported in Table 3 are relatively modest. While consistent, they do not strongly support the claimed impact of the method.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The 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?

    The paper presents a well-engineered approach, but the contribution is incremental and the experimental evidence is not strong enough to support the main claims.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank all three Reviewers for their thoughtful feedback and constructive criticism. R1: In our setting, cell center points are derived from ground-truth instance masks, ensuring accurate conditioning. Text descriptions are extracted from available metadata, providing reliable semantic guidance. Under these realistic conditions, the strong alignment between synthetic and real distributions in Fig. 5(a) demonstrates the stability of our framework. Beyond the t-SNE visualization in Fig. 5(a), we have additionally quantified the morphological distributions of synthetic nuclei masks (area, perimeter, eccentricity, solidity). The results closely match those of real masks, confirming that our method generates diverse nuclei populations. These analyses will be included in the camera-ready version. R2: Our contribution lies in the synergistic integration of FM with multimodal conditioning tailored for pathology nuclei augmentation, supported by consistent improvements across three public datasets. The diversity of synthesized samples is further enhanced by SSL features from pathology foundation models, which introduce beneficial variability beyond the available training data. We acknowledge that expert evaluation would be valuable. However, manual annotation of the nuclei type and boundary is labor-intensive and subject to variability, even among pathologists. Thus, standardized image quality metrics and downstream performance serve as practical, objective proxies. The consistent improvements in perceptual metrics and model-based evaluations (Table 1, 3) jointly support the realism and utility of our synthesized data. Our framework is designed to benefit from pre-trained pathology foundation models. Fig. 5(d) shows consistent performance with different SSL encoders (UNI, Virchow2), demonstrating robustness. A systematic sensitivity analysis and computational cost breakdown will be addressed in future work. R3: Our framework supports end-to-end training with only a few additional parameters introduced by the SSL feature and text embedding modules (Fig. 2). This improves training efficiency compared to multi-stage approaches (ConNucDA). Quantitative comparisons of training time and model size are omitted due to space constraints. The ablation study (Table 2) confirms that each conditioning component contributes meaningfully: text embeddings provide semantic tissue-level guidance, while SSL features encode fine-grained morphological and textural details. In our experiments, t-SNE with multiple hyperparameter settings produced consistent alignment. More importantly, our “high-fidelity” and “biologically plausible” claims are primarily supported by quantitative metrics, not visualizations alone. CLIP-FID uses pathology VLM encoders widely validated to capture biologically meaningful features, providing a robust assessment of synthesis quality. For Fig. 4, representative samples are shown for illustration, while systematic quantitative comparison with baselines is presented in Table 1 and Fig. 3.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

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

    The reviewers pointed out that the paper addresses a timely and practically important problem in computational pathology: improving nuclei segmentation and classification under limited annotations through synthetic data augmentation. Most of all, they appreciate the novel multimodal conditioning strategy, which combines geometric, semantic, and visual cues using nuclei centers, text prompts, and pathology foundation model features. They also view the use of optimal transport flow matching positively because it offers an efficient alternative to diffusion-based methods. The reviewers also mentioned that the paper is clearly written and well presented, with informative figures and a reasonably comprehensive evaluation. Most importantly, the reviewers note that the generated synthetic data consistently improves downstream performance, which supports the practical value of the proposed method.

    The reviewers identify several weaknesses. First, the methodological novelty is considered moderate because the framework mainly combines existing components rather than introducing fundamentally new ones. Second, the necessity and contribution of each conditioning signal are not fully analyzed, including sensitivity to different pretrained encoders and the added complexity they introduce. Third, the robustness of the proposed method remains unclear in realistic settings, such as noisy inputs. Reviewers also note that the reported downstream gains are relatively modest. Finally, the evaluation of generation quality is considered insufficient, with limited analysis of diversity, biological plausibility, structural consistency, failure cases, and expert/pathologist assessment.

    Although one reviewer provided a negative assessment slightly lower than borderline (rating 3), the other two recommended acceptance, with ratings of 4 and 5.I also think the integration of optimal transport flow matching with pathology vision-language and foundation models to be a strong contribution toward high-fidelity pathology image synthesis. Overall, the paper’s strengths outweigh its weaknesses, and I therefore recommend acceptance.



back to top