Abstract

Accurate 3D pelvic landmark localization is essential for hip morphology assessment and surgical planning. However, most models are trained on a single modality and therefore suffer substantial performance degradation when applied to a different modality due to severe domain shift. To alleviate this performance drop, unsupervised domain adaptation (UDA) methods based on image synthesis have been explored. Nevertheless, sparse landmark annotations make it difficult for existing synthesis methods to enforce semantic consistency between translated and real images, which limits the effectiveness of UDA. To address this limitation, we propose a UDA framework for 3D pelvic landmark localization using semantically consistent landmark-conditioned synthesis. Specifically, building on a CycleGAN backbone, we introduce landmark-conditioned adversarial learning (LCAL), which incorporates class- and laterality-conditioned local discriminators to promote semantically consistent image synthesis in regions surrounding each landmark. For localization, we employ modality-specific detectors to encourage consistent localization predictions across synthesized images of the same subject with different styles. Additionally, we propose a pseudo-label ensembling mechanism (PLEM) that produces reliable pseudo-labels by jointly considering localization confidence and cross-prediction spatial consistency, providing effective supervision in the target domain. Experiments on an unpaired hip CT-MR dataset demonstrate that our method outperforms other state-of-the-art UDA approaches on both CT-to-MR and MR-to-CT adaptation tasks. The code is available at https://github.com/KangqingYe/LCS-UDA.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2218_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/KangqingYe/LCS-UDA

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YeKan_Unsupervised_MICCAI2026,
        author = { Ye, Kangqing AND Luo, Simeng AND Gao, Xiaoru AND Zheng, Guoyan},
        title = { { Unsupervised Domain Adaptation for Pelvic Landmark Localization in CT and MR Using Landmark-Conditioned Synthesis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a UDA framework for 3D pelvic landmark localization that combines CycleGAN-based image synthesis with landmark-conditioned local adversarial learning (LCAL) and a pseudo-label ensembling mechanism (PLEM) to improve cross-modality robustness between CT and MRI without requiring target annotations.

  • 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 clinically relevant problem: cross-modality 3D anatomical landmark localization under limited annotations. 2.The proposed LCAL introduces landmark-aware local discrimination, which incorporates anatomical priors into GAN-based synthesis and improves local structural realism beyond standard CycleGAN. 3.The PLEM strategy effectively combines prediction confidence and spatial consistency to generate more reliable pseudo-labels for unsupervised target-domain supervision. 4.Experimental results show consistent improvements over baseline UDA methods on CT-to-MR and MR-to-CT tasks.

  • 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 LCAL module assumes that rigid alignment to a common coordinate system makes landmark locations broadly comparable across CT and MR. However, rigid registration does not guarantee anatomical correspondence at the local level, and inter-subject variability may still lead to spatial mismatch. It is unclear whether such weak alignment is sufficient to support reliable “semantically consistent” local adversarial training, and whether this introduces noisy or biased supervision in LCAL. 2.For style transfer in UDA, beyond CycleGAN-based methods, there are also recent diffusion-based approaches, such as “Reliable Source Approximation: Source-free Unsupervised Domain Adaptation for Vestibular Schwannoma MRI Segmentation”, “Training-free image style alignment for domain shift on handheld ultrasound devices”, and “Zero-shot Medical Image Translation via Frequency-guided Diffusion Models”. It would be beneficial to briefly discuss and compare these methods to position the proposed work better. 3.The dataset is relatively small (70 CT and 68 MR images), and only 47 MR images are used for training. Given the reported fully supervised upper bound (MRE 3.36 mm) using 47 labeled MR images, it is unclear whether the proposed complex UDA framework is necessary compared to simply annotating few MR data and training a supervised model. 4.Relevant references of evaluation metrics (MRE and SDR) choice should be provided. 5.As shown in Fig. 3d, CycleGAN-based synthesis appears already visually plausible and close to the proposed method. A more detailed analysis or explanation is needed to support the claimed improvements.

  • 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?

    The paper presents a reasonably well-designed UDA framework combining synthesis and pseudo-label learning. Although some components (e.g., LCAL and spatial alignment assumptions) are incremental and rely on strong implicit assumptions, the overall pipeline is technically sound and demonstrates consistent empirical improvements over existing baselines.

  • 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

    The paper proposes an Unsupervised Domain Adaptation (UDA) framework built upon a CycleGAN backbone to address performance degradation in 3D pelvic landmark localization across CT and MR modalities. The core methodological contribution involves Landmark-Conditioned Adversarial Learning (LCAL), which introduces class- and laterality-conditioned local discriminators to enforce semantic consistency specifically around anatomical landmarks during image synthesis, addressing limitations of standard global adversarial learning in sparse annotation scenarios. Additionally, the framework incorporates a Pseudo-label Ensembling Mechanism (PLEM) for target supervision without manual annotations. The evaluation shows effectiveness on an unpaired hip CT-MR dataset across bidirectional adaptation tasks, with quantitative improvements over existing SOTA UDA approaches in localization accuracy.

  • 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 problem of cross-modality landmark localization is highly relevant in clinical practice, particularly given the cost and difficulty of annotating MR images compared to CT. The proposed approach addresses a realistic need in surgical planning and morphological assessment.

    • The LCAL module represents a thoughtful extension of CycleGAN by incorporating landmark-aware local discriminators. Conditioning on class and laterality to guide adversarial learning is a meaningful attempt to improve semantic consistency in anatomically critical regions.

    • The paper is generally well-structured, and the experimental section includes comparisons with state-of-the-art UDA methods (e.g., CycleGAN, ICMSC, UMSCS). Results are reported for both adaptation directions (CT→MR and MR→CT), strengthening the empirical 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 dataset is relatively small (70 CT and 68 MR scans). To ensure robustness, the method should be validated using a more rigorous cross-validation protocol to demonstrate generalizability across different cohorts.

    • The method relies on the assumption that “anatomical structures are approximately consistent” after global registration. This is a strong assumption that requires quantitative or statistical justification, particularly given the variability in pelvic anatomy due to pathology or patient-specific factors.

    • The paper does not include a sensitivity analysis for key hyperparameters (e.g., lambda weights in cycle-consistency and LCAL losses and thresholds in PLEM)​. This limits the understanding of how robust the method is and how easily it can be tuned in practice.

    • The claim that LCAL improves semantic consistency is primarily supported by qualitative visual results. Including quantitative metrics (e.g., SSIM, perceptual similarity, or region-specific evaluation around landmarks) would strengthen the evidence for the effectiveness of the proposed module.

    • The description of the three-stage training strategy is not fully clear. In particular, it is not specified how the initial pseudo-labels are obtained in Stage 1 prior to the activation of PLEM. Since LCAL relies on these pseudo-labels to define local regions, this introduces a potential circular dependency between pseudo-label generation and discriminator training. Clarifying this initialization step and the interaction between LCAL crop selection and PLEM would improve the clarity 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 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

    Some figures could benefit from clearer annotations. In particular, Figure 3(d) should explicitly highlight the landmark regions in the synthesized images to make the qualitative comparison more interpretable.

  • 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 a methodologically sound approach to a clinically relevant problem in medical imaging. The proposed LCAL and PLEM mechanisms show promise in addressing domain shift challenges in landmark localization. However, the current submission has significant concerns regarding the robustness of the experimental validation, insufficient justification of key assumptions (e.g., anatomical consistency), and lack of clarity in critical methodological components such as pseudo-label initialization. These issues, along with missing quantitative analysis of synthesis quality and hyperparameter sensitivity, weaken the overall strength of the empirical evidence.

  • 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 authors have addressed my specific technical objections through quantitative justification and clarifications. They provided evidence validating the anatomical consistency assumption and clarified that Stage 1 training avoids circular dependencies, resolving my primary concerns; furthermore, they committed to releasing the code. Regarding hyperparameters, while a more comprehensive ablation study would be preferable, the current explanation clearly justifies their selection criteria and demonstrates robustness to perturbations. While I acknowledge that a more rigorous evaluation protocol (e.g., cross-validation or multi-center data) would further strengthen claims regarding generalizability, these limitations do not undermine the methodological soundness of the proposed framework.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a synthesis-based UDA framework for cross-modality 3D landmark localization, combining image translation and localization learning in a unified pipeline.

    It introduces landmark-conditioned adversarial learning (LCAL) to improve semantic consistency in local regions around landmarks during image synthesis.

    A pseudo-label ensembling mechanism (PLEM) is designed to refine target-domain supervision by leveraging confidence and cross-style prediction consistency.

  • 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 tackles a meaningful and underexplored problem (cross-modality landmark localization), which is harder than segmentation due to sparse supervision.

    The idea of enforcing local semantic consistency around landmarks is intuitive and well-motivated.

    The pseudo-label ensembling strategy is simple but effective, and aligns well with the overall framework.

  • 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 core of the image synthesis still heavily relies on a standard CycleGAN, and the landmark detector (SCN) is a pre-existing architecture. The technical contribution is largely incremental, focusing on specific conditioning rather than a fundamental shift in UDA strategy.

    2.The success of the PLEM and LCAL modules depends on several empirically set thresholds and loss weights. There is a lack of discussion regarding how sensitive the performance is to these values across different datasets or anatomies.

    3.The method depends on anatomical priors (e.g., proxy landmark assignment), which may limit applicability in more diverse or less structured scenarios.

    4.While the results are promising, the study is validated on a relatively small dataset of 70 CT and 68 MR images. More extensive testing on diverse, multi-center data would be necessary to prove the model’s robustness against varying imaging protocols.

  • 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.

    (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?

    My recommendation is based on a balance between the relevance of the problem and the current limitations of the approach. The paper addresses an important and challenging task in cross-modality landmark localization and proposes a reasonable framework with some well-motivated components.

    However, the overall novelty appears incremental, as the method mainly builds on existing synthesis-based UDA pipelines with additional refinements. In addition, the evaluation is relatively limited and does not fully demonstrate generalization or robustness across datasets and settings.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    N/A



Author Feedback

We thank all reviewers for their comments.

R4 Incremental contribution. Unlike segmentation UDA, landmark UDA remains underexplored, and prior work [6] fails on the large CT-MR gap (Table 1). It is also intrinsically harder: segmentation UDA [19,21] preserves semantic consistency via dense masks, unavailable for sparse landmarks. LCAL addresses this through conditioned local adversarial learning, while PLEM improves pseudo-label quality based on confidence and spatial consistency.

R1,2&4 Anatomical consistency assumption. 1) LCAL does not require voxel-wise correspondence. Conditioned on the landmark class, the discriminator only requires both patches to contain the neighborhood of the same landmark, so modest misalignments do not bias the supervision, as GAN-based methods are known to tolerate such misalignment [RegGAN]. 2) The assumption is needed only when valid pseudo-labels are unavailable, mainly for early-stage stability. Once PLEM yields valid pseudo-labels, LCAL crops are driven by them. 3) Quantitatively, the pairwise distances between same-class landmarks across all training CT-MR subjects have a mean of 9 voxels, with 97.6% below 20 voxels. With a patch side length of s_c=40 voxels, the proxy-label-based patch reliably covers the same landmark. 4) Rigid alignment is a standard preprocessing step in unpaired 3D cross-modality synthesis, already applied to anatomies such as the abdomen and heart [19], and the brain [20].

R2 Three-stage training. Stage 1: only the synthesis networks are trained, while the detectors are not updated. Thus, no pseudo-labels are produced and all LCAL crop centers are set to y^s, eliminating any circular dependency. Stage 2: the synthesis networks are frozen and the detectors are trained with L_sup and L_PLEM. Stage 3: all networks are trained jointly, with LCAL crops following Eq. (1). We will release the source code.

R1 UDA necessity. For hip morphology, our 3.53 mm is comparable to the 3.36 mm upper bound, within inter-observer variability for 3D landmark annotation. Reaching it requires 47 MR volumes × 54 landmarks, exactly the burden UDA eliminates, especially as MR labeling is harder than CT [13]. R1,2&4 Dataset size & generalization. Our dataset size and splitting strategy are similar to those of prior UDA studies [19, 21] and 3D landmark detection work [8] (e.g., 100 images). Validation in both directions indicates that the effectiveness is not direction-specific. We acknowledge that cross-validation and multi-center evaluation would further validate robustness.

R1&2 Synthesis quality. Since synthesis serves localization rather than being the final goal, image quality is not our evaluation criterion. LCAL targets local semantic consistency, which global appearance metrics do not directly measure. SSIM requires voxel-aligned pairs, which are unavailable. We therefore evaluate LCAL via localization (Table 2, Fig.3(b)). Fig.3(d) shows CycleGAN distorting the acetabular-notch boundary while our method preserves it.

R1 Diffusion-based comparison. Our GAN-based method enables single-pass generation, allowing end-to-end joint training with the detector. However, diffusion-based methods have been largely restricted to 2D slices due to computational cost (e.g., TISA, RSA, FGDM). The few 3D extensions require paired data [MC-IDDPM] or segmentations [Diffuse-UDA], neither applicable to our setting.

R2&4 Hyperparameter sensitivity. All hyperparameters are set by design, not by extensive tuning: λ_cyc=10 follows CycleGAN [25]; λ_LCAL=0.5 keeps LCAL below the global adversarial term; τ_c=0.9 selects high-confidence predictions; τ_d=3 voxels (4.5 mm) matches the SDR@5mm scale. Even extreme perturbations τ_c→0/τ_d→∞ yield 3.61/3.62 mm MRE (only +0.1 mm vs. 3.53 mm; Fig.3(c)). Table 2 further shows that removing LCAL/PLEM entirely degrades MRE by only ~0.5 mm.

R1 Evaluation metrics. MRE and SDR are commonly used metrics [6,8].




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 manuscript introduces a CycleGAN-based unsupervised domain adaptation (UDA) method for 3D landmark detection in pelvic CT/MRI images. Specifically, it uses landmark-conditioned adversarial learning (LCAL) to promote semantically consistent image synthesis and adopts a pseudo-label ensembling mechanism to provide target-domain supervision. The method is evaluated on an unpaired hip CT-MR image dataset, and it produces better landmark detection performance than some other GAN-based models. However, the reviewers have raised several significant concerns as follows:

    1.The technical contribution is incremental, because the proposed method is built on existing architectures such as CycleGAN [25] and SCN [15]. It would be helpful to clarify the technical contributions.

    2.The method relies on a strong assumption, i.e., anatomical structures are approximately consistent after performing rigid registration. However, this assumption may not hold in practice because rigid registration does not guarantee anatomical correspondence at the local level.

    3.The method is evaluated on a single, small dataset with 70 CT and 68 MR volumes. Please clarify whether this dataset is sufficient to justify the effectiveness and the generalizability of the proposed method. In addition, given that fully supervised learning with a small set of images can provide much better landmark detection performance (in 3mm and 5mm SDR) in Table 1, it is not clear whether using a complex UAD framework is necessary.

    4.The method is not compared with recent diffusion-based approaches, which produce very impressive image synthesis performance. Although the authors should not provide additional experimental results in the rebuttal, it would be helpful to clarify what the advantages of the proposed GAN-based approach are in this landmark detection application, in comparison with diffusion-based models.

    5.There is no sensitivity analysis for important hyperparameters in the proposed method, such as the thresholds in Eq. (6) and the \lambda parameters in Eq. (8). Please clarify whether the proposed method is robust with respect to different hyperparameter values.

    6.The three-stage training strategy is not clearly explained. Please add more details.

    Please consider addressing the reviewers’ concerns in the rebuttal.

  • 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 manuscript introduces a CycleGAN-based unsupervised domain adaptation method for 3D landmark detection in pelvic CT/MRI images. It adopts landmark-conditioned adversarial learning (LCAL) to promote semantically consistent image synthesis and a pseudo-label ensembling mechanism to provide target-domain supervision. The method outperforms several other GAN-based approaches on CT -> MR and MR -> CT domain adaptation directions. The rebuttal has addressed most of the reviewers’ concerns, including technical contributions, anatomical consistency assumption, training details, dataset size, and hyperparameter sensitivity. Therefore, the manuscript is recommended for 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.

    Majority of reviewers are in favor of this work after rebuttal. Some concerns regarding methodology and experimental evaluation have been addressed convincingly according to two reviewers, and I agree with this assessment. While a larger evaluation would be beneficial, the idea is interesting and addresses an interesting problem area, intermodality landmark localization with few annotations. Although the methodological contribution might be a bit incremental, I still see this as interesting work to spark discussion at the conference.



Meta-review #3

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

    Reject

  • Please justify your recommendation.

    This is a borderline paper with mixed reviews. Reviewers agreed that the work addresses a clinically relevant and technically challenging problem—cross-modality 3D anatomical landmark localization under limited annotations—and that the proposed LCAL and PLEM strategies offer meaningful improvements over standard UDA baselines. However, they also raised serious concerns about the paper’s significance and soundness: first, the technical contribution is largely incremental, building on standard CycleGAN and existing detector architectures without introducing a fundamentally new UDA paradigm; second, the evaluation is limited to a relatively small single-center dataset, with insufficient validation of robustness, generalizability, and sensitivity to hyperparameters; third, key methodological assumptions—such as reliable anatomical alignment across subjects and the stability of pseudo-label generation—lack strong quantitative justification, and the dependence on empirically tuned thresholds and anatomical priors may limit broader applicability. The rebuttal clarified certain design choices and provided additional quantitative support, but it did not fully resolve concerns regarding limited novelty, dataset scale, or robustness analysis. After careful deliberation, I lean towards rejecting this manuscript.



back to top