Abstract

Deep learning has demonstrated remarkable success in high-throughput histopathology image analysis. However, the performance of learning-based models critically depends on the quality and size of annotation by expert pathologists—a resource-intensive and time-consuming process. To address the limitations of data scarcity and annotation burden, several methods have been proposed to synthesize paired histopathology data. Nevertheless, these frameworks typically still require annotation data, albeit in reduced quantities, to impose structural constraints during training. In this work, we present CHIS, a plug-in framework that guides the sampling trajectory of a pretrained diffusion model through two key stages: structural initialization at the start and textural modulation during generation. The initial noise state is refined by fusing the phase information from a prior mask with the amplitude of Gaussian noise in the frequency domain, yielding a structurally-informed starting point. During the reverse diffusion process, we adaptively modulate both coarse- and fine-grained textures at different wavelet decomposition levels. This enables a diffusion model pretrained solely on unlabeled images to generate outputs that align with prior structural masks while preserving the reference tissue style. We conducted extensive experiments demonstrating the superiority of CHIS in generation fidelity and its substantial benefits for downstream segmentation tasks.Code is available at https://github.com/IBIL-Code/CHIS.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1821_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/IBIL-Code/CHIS

Link to the Dataset(s)

N/A

BibTex

@InProceedings{QiuYuh_Controllable_MICCAI2026,
        author = { Qiu, Yuheng AND Luo, Jingyi AND Ye, Chenfei AND Ma, Ting AND Cao, Jianfeng},
        title = { { Controllable Histopathology Image Synthesis with Training-Free Structural Initialization and Textural Modulation } },
        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 core contribution of the paper is CHIS, a training-free, plug-in framework​ that enables a pretrained diffusion model (trained only on unlabeled histopathology images) to generate high-quality, structure-controlled synthetic images. In this paper, the authors have proposed two novel technical modules: 1) Structural Initialization: It refines the initial noise state in the frequency domain by fusing the phase information from a prior structural mask with the amplitude of Gaussian noise, providing a structure-informed starting point for the diffusion sampling trajectory. 2) Adaptive Textural Modulation: During the reverse diffusion process, it progressively guides both coarse and fine-grained textures at different wavelet decomposition levels to align the output with a reference tissue style while maintaining structural fidelity. The framework eliminates the need for paired annotation data for training and demonstrates superior results in boosting segmentation performance and image synthesis.

  • 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.Addresses an interesting practical problem: It tackles the significant data scarcity and annotation burden in computational pathology, offering a practical solution for generating high-fidelity synthetic data without requiring expensive paired labels for model training. 2.Novel technical modules: The proposed two-stage guidance mechanism, spectral decomposition for structural initialization and wavelet-based fusion for textural modulation, is innovative and well-motivated by signal processing principles. 3.Demonstrates Superior Performance in Structural Alignment: CHIS achieves the highest FS1 scores and lowest HD95 distances in Table 1. 4.Substantially Improves Downstream Task Performance: When used for data augmentation, CHIS-synthesized images lead to the highest performance gains for the downstream nuclei instance segmentation task (Hover-Net) across all three datasets without using paired data, as shown in Table 2.

  • 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.Insufficient Ablation and Sensitivity Analysis: While hyperparameters (γ, χ, ζ, cutoff radius r) are introduced, the paper lacks a comprehensive ablation study showing the individual effectiveness of the structural initialization and the textural modulation, and the sensitivity of results to these key parameters.

    2.It is suggested to provide the standard deviation and mean value of each numerical result, via using multiple random seeds.

  • Please rate the clarity and organization of this paper

    Good

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    1.The proposed method has been well compared with the existing related approaches. 2.The proposed method contains two novel techinical modules.

  • 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.

    My suggestion is ‘weakly accept’.



Review #2

  • Please describe the contribution of the paper

    CHIS is a training-free framework that guides a pretrained diffusion model to generate structure-aligned histopathology images by injecting mask structure via phase fusion and modulating textures via wavelet decomposition. It consistently improves downstream segmentation performance on three public 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) CHIS adopts a pretrained diffusion model without retraining or paired annotations, ideal for annotation‑scarce pathology nuclei dataset. (2) Structural information from the prior nuclei mask is injected through the phase component of the initial noise, while texture is modulated separately at coarse and fine granularities via wavelet decomposition. This explicit decoupling offers more transparent and controllable generation compared to implicit conditional approaches. (3) CHIS consistently improves the performance of nuclei segmentation model on various datasets. It approaches or even surpasses fully supervised models like NuDiff on some metrics, demonstrating that the synthesized data is practically useful beyond image‑level quality scores.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    (1) The proposed mask aggregation relies only on three heuristic statistics (count, area, dist), ignoring cell type information (e.g., epithelial, immune, stromal) that is readily available in datasets like MoNuSAC and PanNuke. Moreover, realistic spatial arrangements of different cell types (e.g., lymphocyte clustering, tumor cell nesting) are not modeled. As a result, the generated masks may fail to capture the long‑tail distribution of real histopathology tissue architecture. (2) Although the datasets provide per‑nucleus type annotations, CHIS only performs binary nucleus segmentation, limiting its practical utility for multi‑class tasks. Besides, no ablation experiments are conducted to isolate the contribution of the structural initialization module versus the texture modulation module. (3) The integration of PixCell as the pretrained backbone is not clearly described. PixCell is itself a conditional generation model that uses features from a pathology foundation model as control signals. The paper does not explain whether PixCell’s original conditioning mechanism is discarded. (4) The compared supervised model NuDiff (MICCAI 2023) is relatively outdated. A more relevant and recent competitor, ConNucDA (MICCAI 2024), which employs a ControlNet structure to jointly incorporate mask and type information with a pretrained diffusion model, is neither compared nor discussed. While CHIS preserves nucleus locations and reference texture styles, its generated images often exhibit blurred or unclear textures, whereas ConNucDA produces significantly sharper and more realistic texture details through model fine-tuning.

  • 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

    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?

    This paper presents CHIS, a training-free framework that guides a pretrained diffusion model to synthesize structure-aligned histopathology images by injecting mask structure via phase fusion and modulating textures via wavelet decomposition. The method is technically sound, avoids paired annotations, and demonstrates improved downstream segmentation performance on three public datasets. However, several critical limitations prevent a higher score. The mask generation relies on heuristic statistics while ignoring readily available cell-type information and realistic spatial arrangements. No ablation studies isolate the contribution of the two core modules. The integration of the PixCell backbone is not clearly described. Crucially, the comparison lacks recent strong baselines like ConNucDA (MICCAI 2024).

  • 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 addressed my main concerns. The authors clarified that CHIS is training-free and compared against appropriate baselines. The ablation study shows clear contributions from each component. The method also extends to multi-class synthesis with cell-type priors, which adds practical value. Overall, the work is complete and the results are solid.



Review #3

  • Please describe the contribution of the paper

    CHIS, a training‑free, plug‑and‑play framework that guides a pretrained diffusion model to generate structure‑aligned histopathology images. It achieves this through spectral phase fusion for structural initialization and adaptive wavelet‑based textural modulation, enabling a diffusion model trained solely on unlabeled images to produce outputs that respect prior masks while preserving reference tissue style—without requiring paired data or additional training, and with demonstrated improvements in both generation fidelity and downstream segmentation performance.

  • 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.Training‑free and plug‑and‑play: CHIS works with any pretrained diffusion model without requiring additional training, fine‑tuning, or paired annotation data, making it highly efficient and widely applicable. 2.Effective structural and textural control: By combining frequency‑domain phase fusion for initialization and wavelet‑based adaptive modulation during sampling, CHIS achieves strong alignment with prior masks while preserving realistic tissue style. 3.Demonstrated downstream utility: Extensive experiments show that CHIS not only generates high‑fidelity images (better FID/IS scores) but also significantly boosts instance segmentation performance when used for data augmentation, rivaling fully supervised methods.

  • 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.Dependence on a pretrained diffusion model: CHIS requires a high‑quality, histopathology‑specific pretrained diffusion model (e.g., PixelCell). If the backbone is suboptimal or trained on a different domain, the generation quality may degrade significantly. 2.Manual hyperparameter tuning: The framework introduces several hyperparameters that need to be set empirically. The paper does not provide an automatic selection strategy, which may limit generalizability across different datasets or tissue types.

  • 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?

    The paper presents a novel training‑free framework (CHIS) for controllable histopathology image synthesis, with technically sound contributions in frequency‑domain initialization and wavelet‑based texture modulation, and demonstrates clear improvements in both generation fidelity and downstream segmentation. However, the method relies on a well‑pretrained diffusion backbone, requires manual hyperparameter tuning, and is evaluated only on nuclei‑level segmentation tasks. Given the solid novelty and practical utility, but also these non‑trivial limitations, my recommendation is weak accept.

  • 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 authos claim that ConNucDA is supervised while CHIS is training-free, which is reasonable. The authors add some ablation results on methods and parameters, which is also reasonable.



Author Feedback

We thank all reviewers for the constructive feedback, and appreciate the recognition of our novel technical modules (R1, R3, AC), training-free plug-and-play design (R3, AC), explicit structure–texture decoupling (R2), and consistent gains across three datasets (all). We address the main concerns below.

Q1 [R2, AC]: Comparison with ConNucDA. CHIS and ConNucDA address paired data generation under fundamentally different settings. ConNucDA trains a ControlNet with paired mask–image–type annotations, whereas CHIS is strictly training-free and uses no paired data. Our evaluation thus focuses on methods sharing a similar setting (training-free and GAN-based unpaired baselines, Tables 1–3). We also included fully-supervised NuDiff [23] and SDM [21] as reference anchors to inspect the gap to the supervised upper bound, and we observe CHIS not only closes this gap but even surpasses NuDiff on MoNuSAC FS1 (Table 1) and achieves the best FID/IS across all three datasets (Table 3). We further benchmarked against ConNucDA on PanNuke: fully-supervised ConNucDA outperforms our training-free CHIS (AJI: 0.6460 vs 0.6318), however CHIS surpasses ConNucDA under 20% paired supervision (AJI: 0.6318 vs 0.6196). Notably, ConNucDA is inherently limited to in-domain synthesis, while CHIS naturally supports cross-domain generation.

Q2 [R1, R2, R3, AC]: Ablation and hyperparameter sensitivity. 1) Component effectiveness. Structural Initialization (SI) and Textural Modulation (TM) work seamlessly for mask-aligned generation. Over the PixCell baseline (AJI: 0.5795), SI alone yields +2.87%, TM alone +4.32%, and SI+TM jointly +9.02%, confirming their complementarity. 2) Hyperparameter sensitivity. CHIS uses one fixed set (γ=0.2, χ=0.95, ζ=0.6) across three heterogeneous datasets, with ±25% perturbations causing <3.28% FS1/FID variation, demonstrating strong robustness. Due to space limits, we prioritized downstream usability evaluation (Tables 1–2).

Q3 [R2, R3]: PixCell integration. PixCell is used as a frozen backbone; all parameters and its original foundation-feature conditioning (for style guidance) are fully preserved. CHIS operates externally to the backbone: (i) it replaces the initial Gaussian noise with the phase-fused latent at t=T; (ii) injects wavelet-modulated guidance during the first 40 sampling steps; and (iii) allows PixCell to operate freely in the final 10 steps.

Q4 [R2, AC]: Heuristic masks, cell types, and multi-class tasks. Adopting a simple aggregation strategy is a dedicated choice to stay consistent with the training-free principle, and a learned mask generator would contradict our core contribution. Despite this simplicity, CHIS delivers competitive results on the challenging instance segmentation task (Table 2, Dice/AJI on PanNuke, Kumar, MoNuSAC), where precise per-nucleus boundaries are required. Moreover, as stated in Sec. 2.2, the three metrics {count, area, dist} are category-specific and can be “specified directly by pathologists”, enabling expert-derived cell-type priors (e.g., lymphocyte clustering, tumor cell nesting) to be injected without retraining. Extending CHIS to multi-class synthesis is also straightforward: (a) aggregate per-class cell masks into a multi-channel prior mask, and (b) apply SI/TM channel-wise, since both modules are architecturally agnostic to mask semantics. We will flag this as promising future work.

Q5 [R1, R3]: Backbone dependence. We adopted PixCell as an advanced backbone for high-quality image generation, and training-free methods naturally inherit backbone quality. However, the backbone itself cannot generate matching masks—it only synthesizes images conditioned on style reference. Our contribution precisely lies in equipping the backbone with this critical mask-controllable function in a training-free manner. Advances in the backbone will further boost CHIS, but optimizing the backbone itself lies beyond the scope of this work.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

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

    The reviewers pointed out that the paper addresses an important practical problem in computational pathology by reducing the need for costly paired annotations and enabling high-fidelity synthetic data generation for annotation-scarce nuclei datasets. The reviewers further noted that the proposed structural-textural guidance strategy is well motivated, combining phase-based structural initialization with wavelet-based texture modulation to provide more controllable image synthesis. Most importantly, they highlighted that CHIS achieves strong structural alignment and consistently improves downstream nuclei instance segmentation performance across multiple datasets, demonstrating its practical value beyond visual realism.

    The reviewers pointed out several weaknesses as well. The paper lacks sufficient ablation, sensitivity analysis, and reporting over multiple random seeds, making the contribution of each component unclear. They also noted that the mask generation is heuristic, ignores cell-type information and realistic spatial patterns, and the method is limited to binary segmentation despite available multi-class annotations. Additional concerns include unclear PixCell integration, dependence on a strong pretrained pathology diffusion model, manual hyperparameter tuning, limited comparison with recent baselines such as ConNucDA(MICCAI 2024).

    I find the work interesting and useful, particularly because it can be applied without fine-tuning. However, there are several concerns regarding validation, especially those raised by R2, who gave Weak Reject. I recommend that the authors clarify during the rebuttal how the proposed method compares with ConNucDA and address key concerns, including heuristic mask generation and missing ablation studies.

  • 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 authors have satisfactorily addressed most of the reviewers’ questions. Two reviewers are satisfied with the revision and have changed their recommendations to accept. In particular, R2, who initially gave a weak reject, now supports acceptance. Considering the overall discussion and the reviewers’ consensus in favor of acceptance, I also recommend acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After rebuttal, this paper receives three accept recommendations, including one reviewer who changed from weak reject to accept. The AC acknowledges that the paper addresses an important practical problem in computational pathology by enabling controllable histopathology image synthesis for annotation-scarce nuclei datasets without paired training data. The author has addressed most of the concerns in the rebuttal. Some limitations remain, including dependence on a strong pretrained pathology diffusion backbone, manually selected hyperparameters, and heuristic mask construction. However, these concerns do not outweigh the paper’s novelty, efficiency, and demonstrated downstream utility. Therefore, the AC recommends acceptance.



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.

    This paper was reviewed by three experts in the field. Based on the reviewers’ post-rebuttal feedback, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready version.



back to top