Abstract

Most existing deep learning-based registration methods are typically constrained to dataset-specific optimization, requiring separate models for different data characteristics. In contrast, training a single model across diverse datasets presents an opportunity to create a universal registration framework capable of handling multiple domains simultaneously. However, key challenges remain in achieving effective cross-dataset adaptation while maintaining robust generalization capabilities, particularly for zero-shot registration tasks. In this work, we propose PromptReg, a universal image registration framework that incorporates prompt learning to guide the model in effectively adapting to different registration scenarios through explicit task prompts. The core of PromptReg is a Registration Prompt Generator (RPG) that generates domain-specific task prompts based on the domains of input images. Specifically, we first introduce a Static Knowledge Base (SKB) to store domain prompts and a dynamic prompt generation mechanism that projects different inputs into a shared prompt space. Then, we propose an adaptive prompt fusion strategy that combines stored domain knowledge based on the similarity between the generated dynamic prompt and the prompts in SKB, creating transferable knowledge for unseen domains. Finally, we optimize the prompt generator using domain orthogonality and task similarity losses. Our experiments show that PromptReg achieves competitive performance in universal registration and offers stronger zero-shot generalization. The code is available at https://github.com/xiehousheng/PromptReg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1233_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/PromptReg

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XieHou_PromptReg_MICCAI2025,
        author = { Xie, Housheng and Gao, Xiaoru and Zheng, Guoyan},
        title = { { PromptReg: Universal Medical Image Registration via Task Prompt Learning and Domain Knowledge Transfer } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15964},
        month = {September},
        page = {500 -- 509}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper presents PromptReg, a universal medical image registration framework that introduces prompt learning into the registration domain.

  • 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 method achieved competitive performance in universal registration and offers stronger zero-shot generalization.
    2. The motivation of paper is good, which uses prompt learning to medical image registration, guiding the understanding of various registration tasks.
    3. The framework excels in zero-shot tasks, which is highly valuable for real-world deployment in unseen clinical 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. While comparisons to uniGradICON and VoxelMorph are reasonable, recent transformer-based registration methods (e.g., TransMorph, ViTReg) are not evaluated.
    2. It is unclear what the learned prompts represent.
    3. The “dynamic template” is briefly mentioned but not described in detail.
    4. The framework illustrated in Fig. 2 is incomplete, as it omits the decoder module, which may lead to confusion and hinder readers’ understanding of the overall architecture.
    5. Relying solely on the Dice Similarity Coefficient (DSC) for evaluation is insufficient for Registration.
  • 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.

  • 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/2025/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.

    (2) Reject — should be rejected, independent of rebuttal

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

    The motivation of paper is good, which uses prompt learning to guide the understanding of various registration tasks.Yet, it is unclear what the learned prompts represent.

  • 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 PromptReg, a novel universal medical image registration framework that incorporates prompt learning to facilitate domain adaptation and enhance zero-shot generalization. The method introduces a Registration Prompt Generator (RPG) composed of a Static Knowledge Base (SKB), a dynamic prompt generation mechanism, and an adaptive prompt fusion strategy. These components collaboratively generate task-aware prompts to guide the registration process.

  • 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 overall topic of medical image registration is well aligned with MICCAI’s focus on medical image computing and computer-assisted interventions.
    2. Novel Use of Prompt Learning: The work creatively adapts the concept of prompt learning to the field of medical image registration.
    3. The work is well-organized and easy to follow.
  • 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.
    • Insufficient Baseline Comparison: The paper only compares against two methods: VoxelMorph and uniGradICON. While uniGradICON represents a class of universal registration approaches, the comparison lacks broader coverage of other mainstream or recently proposed universal and single-dataset, multi-task registration methods—such as Transformer-based models like TransMorph [1] and a relevant method from MICCAI 2024, SAMCL was not included in the comparison, despite its close relevance to the problem setting addressed in this work. This makes it limited to fully establish the advantage of PromptReg over a wider range of state-of-the-art approaches. [1] Chen, Junyu, et al. “Transmorph: Transformer for unsupervised medical image registration.” Medical image analysis 82 (2022): 102615. [2] Wang, Bomin, Xinzhe Luo, and Xiahai Zhuang. “Toward Universal Medical Image Registration via Sharpness-Aware Meta-Continual Learning.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024.
    • Limited Analysis of Prompt Semantics and Interpretability: While the proposed RPG module includes both static and dynamic prompt components, no detailed analysis is provided on what the prompts encode. There is no visualization or semantic interpretation of the SKB prompts, nor is there an exploration of whether the dynamic prompt adapts meaningfully to different anatomical tasks. This limits the interpretability and practical trustworthiness of the prompt mechanism in clinical use.
    • Questionable Zero-Shot Setup: The leave-one-domain-out setting is referred to as “zero-shot,” but some datasets are closely related anatomically or modality-wise (e.g., Brain and Hippocampus are both data of the brain), which may reduce the difficulty of domain generalization. Thus, the current “zero-shot” results might overestimate the framework’s capability under genuine domain shift.
    • Limited Ablation Analysis: The ablation study only investigates the absence of the prompt generator and regularization losses. However: There is no dissection of the three key components of the RPG (static prompt, dynamic prompt, and prompt fusion). Additionally, there is no hyperparameter sensitivity analysis, particularly for crucial design choices such as SKB size (Z), prompt dimension. This undermines the credibility of the framework’s general applicability.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/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 — could be accepted, dependent on rebuttal

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

    While the paper presents an interesting idea, several aspects limit its strength as a MICCAI publication. The novelty of introducing prompt learning into registration is noteworthy. However, the limited experimental baselines reduce the overall technical depth and experimental rigor. Given these limitations, I recommend a Weak Accept. The idea is novel and could inspire future extensions, but the current form requires stronger empirical support and deeper analysis to fully justify its claims.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper describes promptReg, which is a method to adapt the neural network based image registration for cross-domains with zero-shot. The methodology trains jointly on a set of medical domains, and during inference it can be adapted to other domains without any additional tuning.

  • 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.

    Strengths

    1. The authors motivate teh paper really well, it is easy to understand what is the impact of this work.
    2. Overall the writing is pretty solid, the methods section is clear from the get go.
    3. The evaluation is comprehensive for a conference paper and highlights the key strengths of the paper well. The datasets evaluated on includes both different anatomies and modalities, which is good.
    4. There is clarity on implementation details, especially appreciate the data normalization section.
  • 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.

    I am pretty happy with the paper so I only have minor complaints.

    Minor nitpicks:

    1. Prompt is a confusing term, after reaching the methods I realised that prompt in this case means a learned high dimentional spatial feature. It good to clarify early on, as prompt is usually used in a different context.
    2. Note on how were hyperparams selected (cross validation etc) is valuable.
    3. Additional ablation/analysis which would be valuable is to compute the similarity scores of the p_d for a new domain with each of the b_i computed.
    4. maybe in supplementary include results for individual parts in each dataset, for example I know oasis has segmentations for GM, wm and things like that.
  • 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.

  • 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/2025/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?

    The paper is very balanced with novelty in research, applied impact as well as a solid evaluation. The clarity of the paper is commendable.

  • 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




Author Feedback

We thank all reviewers for their comments.

R1,3&4 Clarification and interpretability of the term “prompt” The concept of “prompt” in this study differs from text prompts (e.g., those in CLIP) or interactive cues (points, boxes, etc) in other contexts. It is a learnable high-dimensional feature that serves as a task-specific “instruction” or “condition”. Formally, each prompt has the same shape as the input feature for RPG. It guides the model to adapt its parameters for different registration tasks, enabling task awareness in decoder. Although prompt techniques are underexplored in image registration, their introduction is well-motivated, as highlighted by R1, R3, and R4. Prompts help the network distinguish between tasks, making them naturally suited for developing universal image registration methods. In our method, static prompts in SKB store knowledge for known tasks, while dynamic prompts generate task-specific guidance and interact with SKB to produce customized prompts for both seen and unseen tasks. This allows the model to generalize to new domains by leveraging similarities in prompt space.

R1&3 Limited comparative methods. Our experiments primarily focus on: (1) evaluating the effectiveness of RPG, which is the core contribution of our method. As shown in Table 2, RPG brings significant improvements to registration performance for universal registration tasks; (2) assessing generalizability by including five diverse datasets, each covering different organs, to ensure the robustness and universality of our cross-dataset setting. While comparisons with methods such as TransMorph, ViTReg, and SAMCL would be valuable, this is restricted by space constraints.

R1 Description of dynamic template. The dynamic template is a learnable feature with the same shape as prompts in SKB. After fusing with input feature to RPG, it generates input-dependent prompt in prompt space, enabling similarity calculation with prompts in SKB for adaptive weighting.

R1 Completeness of the framework shown in Figure 2. We would like to clarify that our main contribution lies in the RPG, while the decoder adopts the one in the original RDP [18] without modification. This was pointed out in the 1st paragraph of the “Methods” section.

R1 Insufficient evaluation metrics. For evaluation, we chose the widely used DSC metric as it effectively reflects registration accuracy. Although including more metrics could provide a more comprehensive comparison, space constraint and the fact that DSC is indicative of registration performance led us to focus on DSC alone.

R3 Questionable Zero-Shot Setup. To ensure zero-shot evaluation, we used multiple training settings across five datasets for comprehensive zero-shot evaluation. Besides the brain and the hippocampus, the other datasets are from the abdomen, heart, and hip, with minimal anatomical overlap, from different centers and varying in quality and original resolution. Thus, we think our setup provides a valid assessment of zero-shot ability.

R3&4 Conduct more ablation studies. For ablation studies, we clarify: (1) all three components in RPG are essential for both in-domain and out-of-domain tasks, making single component removal infeasible; (2) under zero-shot conditions, similarity scores are typically concentrated on one or two static prompts in the SKB, rather than being evenly distributed. For example, cardiac dynamic prompt is closer to hippocampus and hip in SKB than others, and hippocampus is the closest to brain. This shows that the weights focus on the most related known domains; (3) the size of SKB equals the number of known domains, and prompt dimensions match input feature to RPG; (4) loss weights are empirically set. We appreciate all suggestions and recognize the value of broader ablation and individual organ segmentation evaluations. However, due to space limitation, our manuscript focuses on evaluating the effectiveness of RPG for universal registration.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    I think the authors did not properly address the concerns from the reviewers. For example, two of the reviewers clearly state that DSC is not a good measure for registration. Instead of truly explaining that limitation, the authors just state that it is widely used. Segmentation metrics ignore misalignments within the labeled regions, confound segmentation errors with registration errors and are heavily dependent on the segmentation method. They should be combined with other metrics for evaluation. I also think that the concern regarding the use of the word “prompt” is not adequate and does not address the real issues. Why are embeddings considered prompts and why didn’t the authors analyse what information is encoded on the prompts? Their response is essentially a retread of what is said in the paper without truly answering.

    Considering that and the lack of follow-up by the reviewers, I am basing my decision on the original overall weak scores (2 weak acceptance and one weak reject) and my concerns with the rebuttal.



back to top