Abstract

Unsupervised domain adaptation (UDA) based on generative adversarial networks (GANs) or diffusion models has recently demonstrated impressive performance in various medical imaging tasks, including nucleus or cell detection in cross-modality microscopy images. However, these UDA methods often do not consider low-resource settings in the target domain, i.e., only limited target training data is available. Therefore, they suffer from significant performance degradation in low-data regimes. In this paper, we propose a novel one-sided UDA method for low-resource nucleus detection in cross-modality microscopy images. Specifically, we design and incorporate a new masked contrastive learning module into a one-sided GAN for unpaired image translation, which explicitly encourages consistent feature encoding between corresponding source and translated patches. This module is further coupled with a nucleus detector to jointly augment the discriminator and foster semantic content preservation during image translation for domain adaptation. We evaluate the method on multiple public cross-modality microscopy image datasets. In low-data regimes, the method provides significantly superior nucleus detection performance to the reference baseline and recent state-of-the-art approaches, including those diffusion-based models.


Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1395_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XinFuy_LowResource_MICCAI2026,
        author = { Xing, Fuyong AND Ren, Siyang},
        title = { { Low-Resource Cross-Modality Nucleus Detection via One-Sided Domain Adaptation } },
        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 addresses a highly relevant and challenging problem in computational pathology: nuclei detection in cross-modality histopathology images under extreme low-resource conditions. The authors propose a novel framework that leverages a unidirectional GAN architecture enhanced with a Masked Contrastive Learning (MCL) module and joint task-aware supervision from a nuclei detector. Extensive experiments on multiple datasets demonstrate significant improvements over state-of-the-art unsupervised domain adaptation and diffusion-based approaches.

  • 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 Masked Contrastive Learning (MCL) module is a key contribution. By using ground-truth nuclei masks from the source domain to guide patch sampling, MCL enforces semantic consistency specifically in foreground regions during style transfer. 2.The experiments are thorough, covering multiple cross-stain (e. g. , IHC ↔ H&E) and cross-tissue scenarios.

  • 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 MCL module critically depends on accurate binary masks from the source domain. In real-world pathology annotations, labels may contain noise, missing objects, or boundary inaccuracies. It would strengthen the paper to either (a) discuss this limitation explicitly in the discussion section, or (b) include a brief robustness analysis (e. g. , using eroded/dilated masks or simulated label noise). 2.Please specify the exact differentiable augmentation policies used and Clarify the architecture of the detector R. 3.The paper does not mention whether the source code or trained models will be made publicly available.

  • 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.

    (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 weak acceptance because the paper addresses a clinically important low-resource cross-modality adaptation problem with a novel and well-motivated one-sided framework, featuring an effective Masked Contrastive Learning module that significantly outperforms existing methods under extreme data scarcity, as rigorously validated by comprehensive experiments.

  • 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.

    Good paper, accept.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a one-sided GAN for cross-modality nucleus detection in low-resource UDA (target N_T = 1). It combines (i) a Masked Contrastive Learning (MCL) module on top of the generator encoder, where a binary mask derived from source labels constrains query/positive patches to nucleus regions and negatives to background; (ii) joint discriminator augmentation, where a nucleus detector applied to translated images backpropagates a detection loss to the generator; and (iii) differentiable augmentation on target/translated images to mitigate discriminator overfitting. Experiments on one source and three target microscopy datasets show consistent gains over GAN- and diffusion-based UDA baselines.

  • 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. Low-resource UDA for rare-disease microscopy is a meaningful clinical scenario, and the choice of a one-sided GAN over CycleGAN/diffusion is well-justified for this regime.
    2. The foreground/background mask for patch sampling is a sensible adaptation of CUT to microscopy, where nuclei occupy a small fraction of the image.
    3. Across three target datasets and 5 seeds, the method achieves significant F1 gains over 8 baselines (Fluo: 80.5, PC: 91.6, IHC: 85.3), with roughly half the parameters of LRDGAN.
  • 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. MCL is essentially CUT [25] plus a foreground mask; joint task supervision during translation is conceptually close to PLDA [33] and LRDGAN [34]; differentiable augmentation is from [40]. The genuinely new ingredient is the mask, which is not isolated in the ablation.
    2. No scaling study across N_T ∈ {1, 5, 10, …}, so it is unclear when the advantage diminishes.
    3. The proposed method beats the supervised Target model by large margins (e.g., PRAUC 84.1 vs. 74.7 on Fluo); the paper does not clarify whether Target uses one image or the full target set.
  • 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.

    (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?

    Strong, well-organized empirical work on a clinically relevant problem with comprehensive baselines and proper variance reporting. The components are individually borrowed from CUT, prior task-supervised UDA, and DiffAugment, but their integration with the foreground mask works clearly better than the closest competitor LRDGAN at lower computational cost. Concerns about novelty, the N_T = 1-only evaluation, and the unexplained gap over the supervised Target baseline keep this from being a stronger accept, but the paper is above the threshold.

  • 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

    Paper described a GAN application for cross-modality nucleus detection in low-resource settings, using masked contrastive learning (MCL) module .

    As described in the paper, the MCL module enabled efficient single-directional image translation, while the joint discriminator augmentation explicitly enforces semantic content preservation with respect to prediction of nucleus labels for object detection.

    Results of the proposed method claim to significantly outperforms multiple recent state-of-the-art approaches including diffusion-based models.

  • 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.Clear problem statement. 2.Clear methodology. 3.Decent results.

  • 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.

    N/A

  • Please rate the clarity and organization of this paper

    Good

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

    The submission does not provide sufficient information for reproducibility.

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

    N/A

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

    N/A

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

    (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?

    1) While the paper is a classic classification problem under limited data, the approach is sound and the results are on par with the state-of-the-art.

    2) The evaluation on existing databases is solid and conclusion is assuring.

  • 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 original manuscript was in decent shape. Authors understood reviewers’ comments and concerns. The modified manuscript will be even stronger with incorporating the reviewers’ comments.



Review #4

  • Please describe the contribution of the paper

    This paper proposes a one-sided unsupervised domain adaptation (UDA) framework for low-resource cross-modality nucleus detection .

    The key idea is to improve GAN-based domain adaptation under limited target data by introducing a masked contrastive learning (MCL) module to enforce feature consistency between source and translated patches, coupling the MCL module with a task-specific nucleus detector to enhance semantic preservation, and adopting joint discriminator augmentation to stabilize training in low-data regimes. The method is evaluated on multiple cross-modality microscopy datasets and shows improved performance over several GAN- and diffusion-based baselines.

  • 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 focus on low-resource domain adaptation is important and under-explored, especially in medical imaging where target data is often scarce. 2.The experimental section is reasonable with evaluation on multiple cross-modality datasets and comparison with both GAN-based and diffusion-based 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.Limited methodological novelty (main concern) The proposed method is largely built upon existing components, including GAN-based UDA, contrastive learning (patch-wise / InfoNCE) and task-aware discriminator augmentation While the combination is reasonable, the overall framework appears incremental, and it is unclear whether it provides fundamentally new insights beyond existing UDA paradigms. 2.Although several baselines are included, the comparison is still somewhat limited. Most methods are GAN or diffusion-based image translation approaches and missing comparison with more recent feature-level / representation-level adaptation methods such as source-free methods, test-time adaptation methods, self-supervision/training method. 3.While the MCL module is effective, the paper lacks deeper analysis on how much improvement comes specifically from masked sampling vs standard contrastive learning and whether the gains are due to low-resource robustness or general modeling improvements. 4.The method is specifically designed for low-resource scenarios, but it is unclear how it performs in standard (non-low-resource) settings and the broader applicability of the approach is not fully explored.

  • 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

    See weakness

  • 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 addresses a relevant problem and presents a technically sound approach with solid experimental validation. However, the methodological novelty is limited, and the work is largely an incremental extension of existing UDA frameworks. Overall, I find this paper to be on the borderline for acceptance. I am currently tend to a borderline (weak reject) decision. However, I am open to increasing my score if the authors can adequately address the concerns above.

  • 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.

    Thanks for the rebuttal. Authors’ response solve my major concerns and I recommend to accept this paper.



Author Feedback

We appreciate the valuable comments from reviewers (R2, R3, R4 and R5).

Q1: Robustness to noisy masks (R2) A1: We will add a discussion of the limitation of masked contrastive learning (MCL): it relies on source foreground masks, which may be a bottleneck under severe annotation noise. However, the masks are created from labeled nucleus centers with a fixed radius rather than perfect boundary annotations, making MCL likely tolerant to mild or moderate annotation errors. Severe inaccurate or missing nucleus labels may degrade MCL’s performance.

Q2: Detector architecture and differentiable augmentation (R2) A2: The detector uses a U-Net-like encoder–decoder architecture [32]. The encoder has 4 residual learning blocks, each containing two 3×3 convolutional layers, instance normalization and exponential linear units. Strided convolutions connect adjacent blocks. The decoder mirrors the encoder structure and uses transposed convolutions for upsampling.

For differentiable data augmentation, we implement image transformations using native PyTorch tensor operations to enable gradient backpropagation through data augmentation to the generator. We tailor the augmentations [20] specifically for pathology images, including pixel blitting (flipping, right-angle rotation and translation), geometric transformations (scaling and random-degree rotation), and color/intensity adjustments (brightness, contrast, saturation, and hue).

Q3: Source codes (R2 and R4) A3: Codes and trained models will be released upon paper acceptance.

Q4: Novelty and masked sampling (R3 and R5) A4: The primary novelty of this work is development of an integrated unsupervised domain adaptation (UDA) framework for cross-domain nucleus detection in low-resource settings, which is an underexplored and challenging problem. The consistent improvements across multiple datasets demonstrate its technical significance and practical value.

Compared with CUT [25], our method uses MCL to learn from clinically meaningful regions in low-data regimes. The MCL further cooperates with nucleus detection to explicitly enhance prediction consistency between corresponding image patches.

Unlike prior task-supervised UDA [7][18][26][33][34], our method uses only one generator–discriminator pair, thus reducing learnable parameters. Its task supervision is also combined with adapted differentiable augmentation to better address limited target data.

A4.1: We have isolated masked sampling (MCL in Fig. 2) from the framework and compared with non-masked, standard contrastive learning (CL), i.e., CUT [25] in Table 1.Experiments show that masked sampling improves PRAUC performance on Fluo, PC, and IHC datasets (81.5, 74.9 and 80.5 respectively) compared with standard CL (34.7, 50.6, and 48.5 respectively). We will clarify this in the manuscript.

Q5: More target images and non-low-resource settings (R3 and R5) A5: Our method does not degrade in non-low-resource settings (not penalizing use of more target training data). However, due to the design for low-resource settings, its relative advantage gradually decreases as more target images are used. The performance gain is most noticeable with limited target data (Table 1), and this gain drops when using more target images (e.g., over 25 images for Fluo and IHC datasets).

Q6: Target model (R3) A6: The target model is trained with one target image (low-resource setting). The target model overfits limited data, while our UDA uses source data to get better results.

Q7: Comparison with feature-level UDA (R5) A7: Our method outperforms a recent feature-level model, AIF-SFDA (AAAI-25, pp. 4716-4724), by 22.1/18.2/4.3 in F1 for Fluo/PC/IHC. We mainly compare with pixel-space UDA because our method is pixel-space based, and many feature-level models (e.g., source-free or test-time adaptation) assume unavailable source data, which does not match our setting.

Q8: Others (R5) See A4 (novelty), A4.1 (masked sampling), and A5 (non-low-resource)




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.

    This paper developed one sided GAN with masked contrastive learning for extreme low resource cross modality nucleus detection. In rebuttal, clarify detector architecture and differentiable augmentation details, and state code and model release plans. Justify novelty beyond CUT and prior task supervised UDA by isolating the mask in ablations. Explain Target baseline training data and why UDA surpasses it. Discuss robustness to noisy source masks and add scaling across more target images. Please do not feel limited by this meta review and provide a careful comprehensive response to all reviewer concerns.

  • 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 reviewers unanimously provided positive recommendations, and I agree with their decision.



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.

    The reviewers are satisfied with the rebuttal.



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 is a borderline paper with mixed reviews. Reviewers agreed that the Masked Contrastive Learning (MCL) module effectively leverages foreground nuclei masks to improve semantic consistency in low-resource unsupervised domain adaptation (UDA), and that the method achieves strong empirical performance across diverse cross-stain and cross-tissue microscopy datasets. However, they also raised serious concerns about the paper’s novelty and analytical depth: (1)​the core design is seen as an incremental combination of existing techniques (CUT, task-aware discriminators, differentiable augmentations) without isolating the mask’s individual contribution through targeted ablations; (2)​the lack of scaling studies and clarification regarding supervised-target comparisons leaves the limits and validity of the reported gains unclear; and (3)​the evaluation scope is narrow, omitting comparisons with recent feature-level adaptation paradigms such as source-free or test-time adaptation methods. The rebuttal has successfully addressed most concerns. Since the review scores are all positive after rebuttal, I recommend accepting this manuscript.



back to top