Abstract

Deformable medical image registration is an essential task in computer-assisted interventions. This problem is particularly relevant to oncological treatments, where precise image alignment is necessary for tracking tumor growth, assessing treatment response, and ensuring accurate delivery of therapies. Recent AI methods can outperform traditional techniques in accuracy and speed, yet they often produce unreliable deformations that limit their clinical adoption. In this work, we address this challenge and introduce a novel implicit registration framework that can predict accurate and reliable deformations. Our insight is to reformulate image registration as a signal reconstruction problem: we learn a kernel function that can recover the dense displacement field from sparse keypoint correspondences. We integrate our method in a novel hierarchical architecture, and estimate the displacement field in a coarse-to-fine manner. Our formulation also allows for efficient refinement at test time, permitting clinicians to easily adjust registrations when needed. We validate our method on challenging intra-patient thoracic and abdominal zero-shot registration tasks, using public as well as internal datasets from the Innsbruck University Hospital. Our method not only shows competitive accuracy to state-of-the-art approaches, but also bridges the generalization gap between implicit and explicit registration techniques. In particular, our method generates deformations that better preserve anatomical relationships and matches the performance of specialized commercial systems, underscoring its potential for clinical adoption.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1761_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2025/supp/1761_supp.zip

Link to the Code Repository

https://git.uibk.ac.at/informatik/igs/open/msm

Link to the Dataset(s)

N/A

BibTex

@InProceedings{FogSte_Implicit_MICCAI2025,
        author = { Fogarollo, Stefano and Laimer, Gregor and Bale, Reto and Harders, Matthias},
        title = { { Implicit Deformable Medical Image Registration with Learnable Kernels } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {247 -- 257}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a new implicit registration framework that reconstructs a dense displacement field from sparse keypoint correspondences using a learnable kernel.

  • 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.

    As mentioned in the contributions section. Additionally, the comparison with uniGradICON uses instance optimization.

  • 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 method’s description is too vague, and the overall method architecture is unclear. For example, how the dual-stream attention mechanism is integrated with architectures like UNet to construct the complete network is not explained, nor is the implementation of the multi-scale incremental prediction strategy across five scales. This makes it difficult for readers to reproduce the method and grasp this new registration approach.

    2. The comparison methods listed in Table 1, such as H-ViT, CorrMLP, ModelT, NICE-Trans, and RDP, are all brain registration models, which is unfair when tested on the lung dataset NLST. Lung deformations are much larger than brain deformations. Brain registration models generally use single-level networks, which cannot register large deformations, resulting in TREs typically above 3mm. Furthermore, the image resolution used in this paper, 112 × 96 × 112, is too small, which affects the performance of multi-scale cascaded methods like uniGradICON. The first level of such networks takes downsampled images (1/4 resolution) as input to capture large global deformations. Large deformation features will be severely lost when the deformation span is smaller than the downsampling stride. Although the authors mention this is due to memory limitations, it does indeed lower the performance of methods that include additional downsampling operations, especially for lung registration tasks with large deformations. Therefore, I do not believe the comparisons in Table 1 are fair or effective, which is the main reason for my weak rejection.

    3. Other methods lack speed metrics. The proposed method’s 1.6 s is insufficient for real-time registration, especially when the image resolution is still relatively small at 112 × 96 × 112. The proposed method is keypoint-based and uses a two-stage registration approach. The keypoint extraction should also be included in the algorithm’s runtime.

    4. The correctness of keypoint detection should be discussed. After all, in Fig. 2 (b), some keypoints are placed outside the anatomical structures. The authors should discuss the impact of this on the generated deformation field. Keypoint detection methods may be unstable in low-contrast images, noisy environments, or across-modal data (e.g., MRI to CT), which could lead to registration failure in complex scenarios.

  • 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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    The introduction of the methodology is unclear, and the experimental comparisons are unfair.

  • 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 #2

  • Please describe the contribution of the paper

    The authors combine sparse keypoint correspondences with implicitly learned (across the population) deformation models to extrapolate the displacements densely and achieve very high accuracy for lung and liver registration.

  • 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 idea is novel to my knowledge. The experiments are extensive (for a MICCAI paper) and the obtained results high-quality. Using sparse keypoints in combination with implicit deformation field modelling is well motivated. The TTA strategy is interesting.

  • 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.

    More closely related approaches such as CorrField, which also use sparse keypoint correspondences and were a baseline for NLST and reached lower TREs are only cited but not numerically compared (in fact few of the Learn2Reg challenge contestants are included - LapIRN achieved better scores as well). The description is not sufficiently clear. Especially the feature extraction and supervision of this part is neither visualised nor mathematically explained well. What is the size of the correlation volume? What is the difference of semantic and geometric feature encoder? The mentioning of AI instead of deep learning is confusing. The results for liver registration would be more appreciated with ASSD/HD95 for all structures (including tumour) not just liver. Some details about the convolutional implicit deformation field modelling remain also unclear - which part is learned for the population and which is adapted to a specific image pair?

  • 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.

    (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?

    overall the contribution is strong despite some shortcomings in description and inclusion of SOTA results

  • 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 #3

  • Please describe the contribution of the paper

    In this paper, the authors reframe the deformable image registration problem as signal reconstruction, namely to predict a kernel function that reconstructs the deformation from sparse key point correspondences. In order to capture both the large deformations and the fine details, a hierarchical coarse-to-fine architecture is used. Comparing the other deformable registration approaches, the proposed method allows efficient test-time interactive refinement without requiring an additional full registration. The authors also show that the proposed method achieves commercial system-level accuracy but with about 50 times faster inference. The evaluation is also performed on various clinical data.

  • 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 authors proposed a fresh perspective of solving the deformable registration problem by reframing registration as signal reconstruction. Instead of using classical extrapolation, they learn how to extrapolate the sparse key points to dense fields, leading to more flexible and accurate deformations. Comparing the the state-of-the-art method, the proposed method outperforms in accuracy. The evaluation is performed on both public-available clinical datasets as well as local hospital clinical datasets. Ablation studies were also conducted using the public NLST dataset. It demonstrates the reproducibility and the clinical feasibility

  • 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.

    There are parts that could be improved: the use of implicit neural representations conditioned on learned features follows prior work such as Wolterink et al. [46] and Zimmer et al. [48], the authors did not state the clear difference to the prior work. Same to the learnable kernel for dense displacement recovery, seems to be only incremental improvement based on GraphRegNet [17]. As the author already stated, the high GPU memory footprint (~17.8 GB during training) limit real-world clinical deployment, especially in resource-constrained settings.

  • 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.

  • 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?

    I recommend weak accept to this paper, because it proposed a fresh perspective to solving the deformable registration problem, as a practical contribution to the field of medical image registration. The evaluation based on real patient data shows the clinical relevance. The conceptual innovation is somehow incremental.

  • 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 #4

  • Please describe the contribution of the paper

    The paper introduces a novel learnable registration approach. A set of sparse correspondences on specific points of interest are obtained by computing a dense cost volume from dense multi-scale predicted descriptors. From this sparse set of correspondences/displacements, the dense displacement field is obtained via kernel interpolation. The kernel is a sum of 3 kernels, a traditional “handcrafted” spatial regularization kernel, a learned spatial regularization kernel derived from predicted embeddings of the spatial coordinates and a learned intensity based kernel derived from predicted embedding of the image content. In addition, this method supports in a straight forward way, test time interactive refinement by incorporating manually inputted correspondences into the kernel regression. All components are jointly optimized/learned by minimizing a standard registration objective (NCC + isotropic diffusion regularization + Dice and TRE when possible). The method is validated on 2 datasets showcasing long range complex deformations: NLST data featured in the L2R 2023 challenge and an in-house pre-op post-op liver dataset. The method is evaluated against a 8 other SOTA very recent models and compares favorably. In addition, for the liver dataset, an additional clinically relevant metric, the SMA is also computed and compared to a commercially available solution. Finally, the authors mention the possibility of deriving some uncertainty signal from the kernel values.

  • 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.

    I found the paper very convincing on many aspects. First, the method is original and sound. It leverages elegantly learning capabilities into a very understandable and meaningful traditional modeling. It is clearly explained and contextualized within the current literature. It also offers straightforward interaction possibilities which is an important consideration in my opinion. The experiment section is also very convincing as the 2 datasets are both clinically relevant and challenging. In addition, the authors compare against very recent and performing baselines which ensures the value of this contribution. The final conclusion and discussion section is insightful.

  • 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.

    Again, I found the paper excellent and it does not have, in my opinion any major weakness. I only found 2 elements a bit confusing when looking at the result tables. First, I am not sure it is said what the final setting for the proposed method is in the experiments in terms of key-points extraction. Similarly for the ablation study, it is not 100% clear if the full proposed kernel is used when comparing the different key points extractors and what key point extractor is used when comparing the addition of the proposed kernel components.

  • 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.

    (6) Strong Accept — must be accepted due to excellence

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

    I found the paper excellent.

  • 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




Author Feedback

N/A




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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



back to top