Abstract

Unsupervised learning is the dominant paradigm in medical image registration, primarily due to its straightforward deployment and independence from labor-intensive manual annotations. However, its optimization objective fails to differentiate between semantically distinct anatomical structures and background regions, resulting in suboptimal registration performance. To introduce semantic priors into unsupervised registration with minimal annotation cost, we propose OPS, a One-shot Point prompt-driven Semantics-aware registration framework. Leveraging the powerful semantic clustering capabilities of the pre-trained vision foundation model DINOv3, OPS requires only extremely sparse point prompts (i.e., one point per organ) on a single reference image slice to extract global semantic anchors. These anchors are then propagated across the entire unlabeled dataset to compute feature similarities with the DINOv3 embeddings of each image volume, thereby localizing semantically consistent regions. This process enables the offline generation of dense semantic similarity maps, which serve as nearly annotation-free anatomical localization priors. Subsequently, we integrate these anatomical localization priors into the registration learning through a dual mechanism: (1) semantic feature embedding to modulate the feature representations, and (2) semantics-aware loss optimization to spatially re-weight the similarity loss, enabling semantic guidance for the registration learning. Comprehensive experiments on two datasets demonstrate that the proposed method enhances the registration performance of state-of-the-art methods. The code is available at https://github.com/xiehousheng/OPS.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0160_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/xiehousheng/OPS

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XieHou_OPS_MICCAI2026,
        author = { Xie, Housheng AND Gao, Xiaoru AND Zheng, Guoyan},
        title = { { OPS: One-Shot Point Prompted Semantics-Aware Learning for Medical Image Registration } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16889},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes OPS, a one-shot point prompt-driven framework for incorporating semantic priors into unsupervised medical image registration. Specifically, it leverages a pre-trained vision foundation model (DINOv3) to propagate extremely sparse point prompts (one point per organ on a single reference slice) into dense semantic similarity maps across the entire dataset. These maps serve as anatomical localization priors and are integrated into the registration process through both semantic feature embedding and a semantics-aware loss re-weighting mechanism. This design enables the introduction of semantic guidance into unsupervised registration with minimal annotation cost.

  • 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 proposed approach introduces semantic awareness by leveraging extremely sparse annotations (one point per organ) on a single reference slice. This is particularly interesting because it significantly reduces annotation cost while still enabling the model to capture anatomically meaningful correspondences, effectively bridging the gap between unsupervised and weakly-supervised registration. 2.Use of a pre-trained foundation model (DINOv3) to propagate sparse point prompts into dense semantic similarity maps across the dataset. This effectively leverages the semantic clustering capability of foundation models for registration, avoiding the need for dense supervision.

  • 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 method relies on extremely sparse point prompts (one point per organ) to construct semantic priors. Could the authors clarify the strategy for selecting these points (e. g. , anatomical center, boundary, or random selection)? In addition, how sensitive is the method to the spatial location and number of selected points? While a basic robustness analysis is provided, a more systematic evaluation (e. g. , varying point locations or using multiple points per structure) would help better understand the stability and generalizability of the proposed approach. 2.The method assumes that semantic anchors extracted from a single reference slice can generalize across the entire dataset. Could the authors provide further analysis or visualization to demonstrate how well these anchors transfer across subjects with anatomical variability? 3.Additional visualization of the generated semantic similarity maps would help to better assess their quality and interpretability, particularly in challenging cases.

  • 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.

    (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 proposes a novel and practical approach to incorporate semantic priors into unsupervised registration using sparse point prompts and a pre-trained foundation model. Compared to prior work, such as DINO-Reg, which directly uses DINO features for registration, the proposed method introduces an explicit semantic prior construction mechanism by propagating one-shot point prompts into dense semantic similarity maps. This design enables more targeted and interpretable semantic guidance with minimal annotation cost.

    The method is well-motivated and shows consistent improvements across datasets, demonstrating its effectiveness and generalizability. Although the sensitivity to point selection could be further analyzed, this does not significantly affect the overall contribution.

    Overall, the work provides a meaningful extension over existing foundation model-based registration approaches and supports a positive recommendation.

  • 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 semantics-aware framework for unsupervised medical image registration. Its main contribution is to use extremely sparse point prompts on a single reference slice to extract DINOv3-based semantic anchors, propagate them across the dataset to generate dense semantic similarity maps, and incorporate these priors into registration through semantic feature embedding and semantics-aware loss re-weighting.

  • 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 clearly motivated problem, namely the semantic blindness of conventional unsupervised registration methods that rely only on global image similarity. 2.The proposed one-shot prompting setup is practically appealing, as it attempts to introduce semantic priors with very limited manual annotation cost. 3.The overall framework is coherent and well organized. The generation of semantic similarity maps and their integration through semantic feature embedding and semantics-aware loss optimization form a consistent pipeline. 4.The empirical results are reasonably strong. The method improves both DSC and HD95 over the compared backbones, and the ablation study supports the contribution of each component.

  • 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 workflow is practically appealing, but the technical novelty is somewhat less convincing than the framing suggests. In particular, the method does not introduce a fundamentally new registration formulation; rather, it builds a dataset-wide semantic prior from sparse prompts using frozen DINOv3 features, and then incorporates this prior through relatively standard feature embedding and loss re-weighting steps. As a result, the main novelty seems to lie more in the prompt-to-prior workflow than in a clearly new registration mechanism.1 2.This concern is also reflected in the ablation study. On the abdominal dataset, adding SFE alone improves DSC only from 58.52 to 59.39, whereas SLO alone raises it to 65.42, which is already close to the full model at 66.54.A similar pattern is observed on the hippocampus dataset. This suggests that much of the gain may come from semantic loss weighting itself, while the additional benefit of the prompt-derived feature embedding is comparatively modest.

  • 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?

    I gave this paper a Weak Reject because, although the problem is meaningful and the workflow is practically appealing, I was not fully convinced that the technical contribution is strong enough for acceptance. The proposed one-shot prompting setup is interesting, but the main novelty seems to lie more in generating a dataset-wide semantic prior from sparse prompts than in a clearly new registration formulation. In addition, the ablation results suggest that most of the empirical gain comes from the semantics-aware loss re-weighting, while the additional contribution of semantic feature embedding is comparatively modest. Overall, the method is technically sound and reasonably effective, but I felt that the paper would be stronger with a clearer demonstration of what is uniquely enabled by the prompt-derived prior beyond semantic loss weighting.

  • 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.

    Most of the concerns have been adequately addressed, and the practicality of the proposed workflow, in particular, represents a significant contribution. Although the efficient point-prompting methodology could be further strengthened by more detailed technical analysis of the SFE and SLO components and a more comprehensive experimental design, these remaining limitations do not substantially weaken the paper’s overall contribution.



Review #3

  • Please describe the contribution of the paper

    The paper proposes OPS, a framework that introduces semantic guidance into unsupervised medical image registration with minimal annotation. Its main contributions are: One-Shot Point Prompting: Leveraging DINOv3, it generates dense semantic similarity maps for an entire dataset using only one point per organ on a single reference slice, avoiding costly full segmentation masks. Dual Integration Mechanism: It incorporates these maps via Semantic Feature Embedding and Semantics-aware Loss Optimization to explicitly guide feature learning and spatial alignment. Experiments show OPS significantly outperforms state-of-the-art unsupervised methods on abdominal CT and hippocampus MRI 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.

    The paper introduces a highly efficient data usage strategy, generating dense semantic priors for an entire dataset from just one point per organ on a single reference slice. This drastically reduces annotation costs compared to traditional weakly-supervised methods requiring full 3D masks, enabling semantic-aware registration on vast unlabeled clinical datasets.

  • 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 reliance on DINOv3 feature similarity assumes consistent feature distributions across images. This breaks down in multi-modal registration (e.g., CT-to-MRI), where intensity and texture differences cause the same organ to have vastly different feature embeddings. The method lacks a mechanism for modality-invariant alignment, severely limiting its clinical scope.

    1. In mono-modal settings, the advantage over existing pseudo-labeling strategies (e.g., using pre-trained segmentation models like nnU-Net to generate coarse masks without any user input) is unclear. The complexity of integrating a massive foundation model may not justify the marginal gain over simply running an off-the-shelf segmenter to get richer spatial priors.
  • 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?

    While the proposed method is technically sound, its practical value needs further clarification.

  • 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.

    N/A



Author Feedback

We thank the reviewers for their constructive feedback and for recognizing that our method enables more targeted and interpretable semantic guidance with minimal annotation cost.

[R2] Clarification on Technical Novelty: We thank the reviewer for acknowledging the practicality of our workflow. While we do not introduce a new registration architecture, our novelty lies in establishing a point prompted semantics-aware learning that advances unsupervised registration learning at minimal annotation cost. This is done by using extremely sparse point prompts to derive feature similarity maps, and then integrate these maps into the registration learning process via a dual mechanism comprising SFE and SLO. Leveraging the strong generalization capability of DINOv3, our method improves performance on unsupervised registration, which is a clear advantage, and the experimental results substantiate its efficacy. Moreover, our method is plug-and-play and offers users explicit controllability over the registration learning process: by simply altering the point prompts, clinicians can dynamically modulate the network’s attention toward specific anatomical targets during training, thereby enabling customized registration.

[R2] Ablation Study on SFE and SLO: The ablation results in Table 2 are consistent with our design, and the two components are complementary to each other. SLO directly reshapes the supervision signal, explicitly forcing the optimizer at the training stage to focus on aligning semantically relevant regions, thereby contributing more substantially to performance gains. SFE injects semantic awareness into feature representations at both training and inference stages. This explains why SFE continues to enhance registration performance even beyond SLO.

[R3] Advantage Over Fully Supervised Pseudo-labeling: We would like to clarify the difference between OPS and pseudo-labeling schemes. OPS targets unsupervised registration learning, where the segmentation annotations required by model nnU-Net is hard to obtain and OPS needs only a one-shot interaction. Second, segmentation models are strictly constrained by the organ categories and image domains seen during training, often failing on unseen domains or structures. In contrast, OPS builds on DINOv3’s general-purpose visual representations, requires no organ-specific or modality-specific pre-training.

[R1] Prompt Selection Strategy: For each target organ, we randomly pick a single point prompt within the organ, without requiring precise localization. Table 3 quantitatively demonstrates robust performance under varying random point locations. While multi-point prompting is a good idea, it would enable additional annotation cost, which may be not optimal in a clinical environment.

[R3] Adaptation to Multi-modal Registration: In fact, OPS naturally accommodates cross-modal scenarios. For example, in CT-MRI registration, only one point prompt targeting the same anatomical structure needs to be provided on each of the CT and MRI reference slices. The CT anchor is then propagated exclusively within CT dataset, while the MRI anchor is propagated exclusively within MRI dataset, thereby generating semantic similarity maps for the same anatomical structure in CT and MRI data to ensure compatibility with OPS.

[R1] Cross-subject Robustness and Additional Visualization: The cross-subject generalization is quantitatively validated by our existing experiments. Substantial anatomical variability (organ size, shape, position) exists among the 54 abdominal and 260 hippocampus subjects in the datasets. Supported by DINOv3’s strong cross-image semantic consistency, anchors propagate stably across subjects via feature similarity, producing reliable similarity maps under large anatomical variation. In the upper-right corner of Fig. 2, we provide examples of similarity maps, where it can be observed that similarity maps from different samples capture consistent organ-level semantic information.




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 received mixed recommendations from the reviewers, including one Weak Reject and two Weak Accepts.

    The paper addresses deformable registration using point prompts and a pre-trained vision foundation model, and demonstrates promising results on two datasets. At the same time, the reviewers identified several substantive concerns, including an unclear point selection strategy (R1), unconvincing support for the claimed semantic anchors and the reliance on DINOv3 feature similarity (R1, R3), limited qualitative results (R1), the incremental nature of the methodological contribution (R2), marginal gains from the prompt-derived feature embeddings (R2), and an unclear advantage over a nnU-Net-plus-registration pipeline (R3).

    Overall, the paper shows merit in its methodology and in its use of vision foundation models for image registration, and the meta-reviewer therefore recommends that it proceed to rebuttal. The authors are encouraged to focus their response on the major concerns outlined above.

  • 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 paper proposes a method that combines DINOv3-based semantic clustering with a single-slice, single-point prompting scheme to construct a dataset-wide dense semantic similarity map, which is subsequently integrated into an unsupervised registration framework through SFE and SLO.

    Following the rebuttal, all reviewers expressed overall positive assessments. In particular, Reviewer 2 explicitly recognized that the practicality of the annotation-efficient workflow itself represents a meaningful contribution and accordingly revised the recommendation to Accept; Reviewer 3 also upgraded the rating to Accept. While the lack of a comparison with nnU-Net remains a minor weakness, it does not substantially undermine the core technical contribution or the empirical validity of the paper.

    Based on the post-rebuttal discussion and the overall reviewer consensus, I recommend Accept.



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.

    This paper proposes to use sparse point prompts to create feature similarity maps using Dinov3 and to incorporate those maps in an unsupervised registration algorithm. After rebuttal, all reviewers agree that the method is technically sound and that the main concerns have been addressed in the rebuttal. There are some remaining concerns about the experimental design and analysis, but the strengths of the paper outweigh its weakness.



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.

    Following the rebuttal phase, two of the three reviewers issued a final decision of “Accept,” while the third reviewer did not respond. However, this reviewer’s initial score was “Weak Accept.” Consequently, the overall assessment of the paper is positive (Accept), and it is therefore recommended for acceptance.



back to top