Abstract

Learning-based registration treats motion as the first citizen, yet motion is inherently \emph{relational}, requiring two time points to define, and entangles learned parameters with both the fixed and moving domains. Any change in modality, dimensionality, or acquisition protocol therefore demands specialized adaptations. We approach this differently: rather than learning motion itself, we learn the \emph{motion subspace}. Unlike motion, the subspace of plausible deformations is an \emph{intrinsic} property of anatomy, potentially inferable from a single scan. This reframing leads to an asymmetric architecture in which only the fixed image couples the network, predicting patient-specific scalar \emph{motion modes} that span the subspace, instantiating motion then reduces to estimating a small set of coefficients from various observations via a closed-form solver. A mode normalization layer resolves the signed-scale ambiguity inherent in mode–coefficient decomposition, making network integration feasible. On 4D-lung CT dataset, the learned modes approach iterative pTVReg in accuracy while the coefficient solver runs in 7.5\,ms. Without retraining, the same modes unify four observation modalities (3D image, 2D image, 2D mask, surface points), all achieving Dice above 0.9, demonstrating that patient-specific motion subspaces can be inferred from static anatomy and serve as unified modes for multi-modality motion reconstruction.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1989_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhuYin_Learning_MICCAI2026,
        author = { Zhu, Yinheng AND Chen, Mingli AND Wang, Qingying AND Gu, Xuejun AND Lu, Weiguo},
        title = { { Learning What Can Move: Patient-Specific Motion Subspaces from a Single Scan } },
        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 authors present a method to predict the motion subspace of a given anatomy. From a dataset of 4D CT images, the method learns a set of modes to explain the motion observed in the images. This enables anatomically-constrained motion prediction from real-time, sparse observations such as 2D images or surface displacements. In the authors’ formulation, the modes are scalars and predicted once, and the coefficients are vectors optimized dynamically to fit each new observation. This contrasts with prior works that required a patient-specific 4D CT image for real-time motion estimation from sparse observations.

  • 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 method predicts a reusable motion subspace rather than specific deformation instances. This regularizes the predicted motion and ensures it belongs to an anatomically plausible motion subspace.
    • The method simplifies real-time motion prediction down to coefficient solving, alleviating computing constraints in real-time applications and reducing risks of overfitting to unrealistic deformations to fit the observations.
    • The use of scalar rather than vector displacement modes brings rotational invariance to network predictions, making the method more robust to non-standard image orientations.
    • The mode prediction + solver optimization enables 3D motion prediction from arbitrary observations, making the method highly flexible in its potential applications.
  • 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.

    Conditions acceptance rating:

    • The paper should absolutely discuss the RMSim paper (https://iopscience. iop. org/article/10.1088/1361-6560/acb484/pdf) which is extremely close conceptually. Cited from RMSim: “We present a novel 3D Seq2Seq deep learning respiratory motion simulator (RMSim) that learns from 4D-CT images and predicts future breathing phases given a static CT image. “ This is very closely related and highly relevant as prior work, even if the proposed approached, based on a subspace formulation, is genuinely different.

    Major:

    • The main strength of the method, the plausibility of predicted deformation, is also an important weakness. The method is dependent on the availability of motion-specific data to train the subspace prediction method: if the motion to be predicted does not exist in the dataset, predictions will fail. This is particularly relevant in interventions where the anatomy is deformed by the action of the clinician, for example due to the manipulation of the anatomy by surgical tools. These types of deformations are very unlikely to be represented in the training data, and due to the formulation of the method, will fail to be accurately predicted.
    • The paper does not evaluate the ability of the method to predict out-of-domain deformations. It claims to decouple “what deformations are plausible (modes, predicted once) from which deformation occurs (coefficients, solved per observation)”. This is a somewhat strong affirmation: the space of plausible deformation is vastly greater than the space of deformations observed in the training data, currently restricted to breathing and potentially cardiac motion 4D CTs.
    • The method was only trained and evaluated on breathing motion prediction from a single dataset. The quality of the validation would have been improved by training on one dataset and evaluating on another, ideally experimentally acquired, to evaluate the robustness of the method to deviations from the breathing patterns observed at train time. It would have been further improved by evaluating its accuracy on other types of motion, such as cardiac motion.

    Minor:

    • Linearity assumption is unexamined. The method assumes anatomical motion is always well-approximated by a low-dimensional linear subspace. This is a strong assumption that is never discussed or justified. Non-linear motion subspaces are a known alternative (see https://arxiv. org/pdf/2404.17620), and the paper does not address under which conditions linearity breaks down.
    • Choice of K is insufficiently justified. Only K=1 and K=2 modes are evaluated, with no sensitivity analysis on K. It is unclear how to choose K for a new anatomy or motion type.
    • The 2D gradient approximation lacks justification. When only a 2D slice is available, the method substitutes ∇I_fix for ∇I_mov in the Jacobian definition. The validity of this approximation is non-trivial and unexplained. A more natural formulation — restricting the problem to 2D and computing gradients directly in 2D — is not considered or compared against.
    • The solver and network details are in “Experiments and Results”, and should be moved to Methods section.
    • Abstract: “instantiating motion then reduces to estimating a small set of coefficients from various observations via a closed-form solver. “ This is unclear
    • Introduction: Need to explain the benefit of generating a reusable motion subspace rather than deformation instances.
    • Introduction: “We introduce a scalar-mode parameterization that naturally admits a closed-form coefficient solver, together with ModeNorm layer that resolves the signed-scale ambiguity in basis–coefficient decomposition” The signed-scale ambiguity should be briefly introduced before.
    • Methods: The “training” part of Fig. 1 could be made a little clearer by replacing the curved arrow with a straight arrow pointing straight down. As it is, it seems L_reg is somehow related to the solver, which it isn’t.

    Nevertheless, contributions are novel and the method has great potential to enable anatomically accurate, fast registration from sparse observations

  • 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.

    (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?

    This paper addresses a meaningful and underexplored problem: learning a patient-specific motion subspace from a single static scan, enabling anatomically-constrained real-time motion estimation from sparse observations without requiring a patient-specific 4D acquisition. The scalar mode prediction-based solution is original and well justified. The ModeNorm fix to sign ambiguity is clever : in this work, as opposed to traditional registration works, both the modes and coefficients vary: the modes are predicted by the network and the coefficients computed by the solver, and both parts could apply arbitrary scaling, if it were not for the ModeNorm layer.

    The practical implications, real-time 3D motion estimation from 2D images or surface signals, without patient-specific 4D CT, and novelty of the method are enough to justify the “accept” recommendation, despite the insufficient validation and lack of proper sensitivity/ablation study in the paper, which should be addressed in a revision, in accordance with MICCAI’s rules.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    The authors’ rebuttal did not fully address all reviewer concerns, notably the lack of clarification regarding the 2D gradient approximation and the motivation for the choice of K. These remain open points that should be addressed in the final version. More importantly, the authors’ claim that the linear combination limitation is a training data issue independent of their method is not convincing. Methods from the literature trained on unsupervised, cross-patient registration tasks do not exhibit this limitation; it is a structural consequence of the proposed framework, not an artifact of the data. The authors should acknowledge this more honestly. That said, these weaknesses do not undermine the core contribution. The method is sound and novel, the paper addresses a relevant problem, and the experimental results support the claims. The paper deserves publication, conditional on the authors revising the discussion of limitations to be more transparent about what is a method-intrinsic constraint versus a data-dependent one.



Review #2

  • Please describe the contribution of the paper

    This work propose to train a network to produce subspace of motion vector fields from static images. The papers assumes all possible motions can be derived from static anatomy.

    This network provides a pathway for performing downstream tasks such as 2D to 3D DVF estimation without retrainning.

  • 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 strength of the paper lies with its novel application of deriving DVF modes from static anatomy and applying the derived modes for various downstream tasks without requiring required model for each specific task.

  • 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 claim that all possible motion can be restricted to a low-dimensional subspace is very questionable. For example, patients with bronchus obstruction on one side of the lung will have drastically different motion from those without such obstructions. This motion difference is unlikely to be distinguishable from a single static image.

    • In my opinion, the quantitative comparison with baseline methods are not sufficient. For example, the base line is only established with 3D to 3D task. What about baseline comparison with other tasks?

    • It is true that the pipeline eliminates the need to retrain for other tasks. However, the ultimate real world clinical goal is to make registration more accurate. From the presented results, the proposed method does not outperform one of the baselines and there is lack of comprehensive information on inference speed. Again, without comprehensive baselines for other downstream tasks, the evaluation of the performance is infeasible.

  • 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 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

    Fonts on your figure 4 is too small.

  • 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 study builds on a very questionable assumption. The registration performance is not necessarily better than baselines and there is a lack of comprehensive evaluation.

  • 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.

    My previous opinion was that the method cannot generalize to every possible motion such as bronchial obstruction. After reviewing the authors’ response, I agree this might be a limitation of the training data instead of that of the method itself. Given this and combined with the clear advantage in computational speed, I now recommend acceptance.

    I still think it would be nice to demonstrate quantitative results beyond 3D to 3D but I’m not sure whether this is allowed during this stage.



Review #3

  • Please describe the contribution of the paper

    This work shifts from learning motion directly to learning a patient-specific motion subspace inherent to anatomy. By predicting scalar motion modes from a single scan and solving for coefficients via closed-form methods, it enables multi-modality motion reconstruction without retraining.

    The paper offers an entirely new approach to image registration by limiting the solution space by one image of the image. This approach offeres interesting applications.

  • 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 proposes a novel approach to image registration that creatively decouples the estimation of the deformation field from the two images. Instead of relying on pairwise observations alone, they first infer a patient-specific motion subspace from a single static scan, which defines what deformations are anatomically plausible for the subject. This decoupling of deformation plausibility from the registration process itself allows them to model the allowed motion subspace for just one of the images. By pre-defining this constrained subspace, they reduce the problem’s complexity and avoid the need to explore implausible deformations during registration. They then select the appropriate solution from the far fewer remaining degrees of freedom very quickly with the second image.

    This approach brings big gains in speed as compared to the classical image registration approaches as thousends of parameters would be have to be optimised in classical DIR where as in thie approach only a handful needs to be optimised resulting in a large speed gain.

  • 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.

    What I did not understand is on why is there an arrow in Figure 1 that points from the “Fixed image” to the “2.2 Closed-form solver”. This makes no sense to me as the authors claim to decouple the two images

  • 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.

    (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 offers a fresh new approach to solving the image registration problem with a creative approach that also offers speed advantages.

  • 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 (meta-)reviewers for constructive feedbacks.

1.R2|(1)“The claim that all possible motion can be restricted to a low-dimensional subspace is very questionable”, obstruction example, (2)”This motion difference is unlikely to be distinguishable from a single static image.”

Before addressing R2’s points, we clarify some concepts: motion subspace refers to a set of a_k(x), patient-specific and predicted by the network; a specific motion D(x,t_i)=sum^K_k(a_k(x)*w_k(t_i)).

We separate two claims R2 conflates: (1)expressiveness of a low-rank subspace vs. (2)predictability of a patient-specific subspace from a static scan.

Expressiveness is not an open question for respiratory DVFs: Li et al. (2011) and follow-ups empirically validated low-dim linear subspaces on 4D-CT. Even higher-DOF regimes (CFD flow fields) use linear subspaces (POD), with larger K. Intuitively, approximating D(x,t) by sum^K_k(a_k(x)*w_k(t)), as K grows, error diminishes.

BO is thus not a counter-example to expressiveness. The last sentence conflates “motion” and “motion subspace”. We assume the reviewer refers to whether the network can predict the patient-specific subspace when such variability is rare in training. As long as BO cases exist in training, the network can learn the subspace. We acknowledge this as a training-data diversity limitation rather than model formulation. Details in p2. 2.R1|Training-set dependence and OOD Acknowledged; we will add an explicit limitations sentence. Two conditions considered:

(i)Anatomy-conditioned variants present in clinic but rare in current training (e.g., BO in p1): a data-coverage issue rather than model limitation, addressable by curating cohorts with such phenotypes. This is common in supervised learning, not specific to our method, and data curation is a standard solution.

(ii)Tool-induced/interventional deformation (R1’s surgical example): outside our scope, which targets physiologically driven motion (respiratory, with cardiac a natural extension). Will state this explicitly.

3.R1|Benchmarks, datasets, baselines Agreed, broader benchmarks (more datasets, regions, clinical validation) would strengthen the work. The page budget was spent framing a new problem: predicting a subspace rather than a deformation. Cross-dataset/anatomy evaluation is planned for journal extension.

4.R2|“no improvement” We target real-time motion recovery: system latency budget (e.g., MR-Linac IGRT) is ~100 ms total, of which we target ≤10 ms for motion recovery, differing from DIR with no apples-to-apples baseline. We use pTVReg as a high-accuracy reference (3–5 min/pair, more dataset-robust than the learning baseline) to bound accuracy sacrificed under ~10 ms. NetADM beats the ResNet-18 baseline (Dice 0.921 vs 0.882) and approaches pTVReg (0.930) at ~24000× speed-up (7.5 ms vs 3–5 min).

6.Others

  • RMSim and Neural Modes: will cite in related work. RMSim implicitly encodes motion via 3D-ConvLSTM Seq2Seq, while we learn a linear subspace enabling closed-form solving from sparse observations at ~10 ms. Neural Modes target on computer graphics applications with elastic equilibrium(material force model), a regime distinct from ours (we model neither tissue mechanics nor excitation).
  • Decoupling and Fig 1 arrow: “Decoupling” is at the representation level: f_θ(I_fix)→modes uses only I_fix; the solver ingests whatever real-time observation is available (3D image, 2D slice, 2D mask, surface points), never I_mov as a network input.
  • 2D gradient: Follows the classical Thirion-demons static-force form (small-warp linearization). The “directly 2D” alternative computes gradients on warped images and discards out-of-plane displacements; preliminary tries showed instability.
  • K: K=2 already approaches pTVReg (diminishing returns).
  • Writing/reproducibility: We will clarify the abstract on coefficient solving, move solver/network details to Methods, enlarge Fig 4 fonts, release code upon institutional approval.




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.

    The paper introduces a novel and clinically meaningful approach for learning patient-specific motion from a single scan, enabling real-time estimation without 4D data, with a creative design such as ModeNorm. Despite its originality and practical potential, the work relies on some questionable assumptions and lacks thorough validation and analysis. With two positive and one negative rating, the AC recommends proceeding to rebuttal before making a final decision.

  • 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.

    Although some concerns remain regarding the 2D gradient approximation, the choice of (K), and the need for a more transparent discussion of method-intrinsic versus data-dependent limitations, the core contribution is sound and novel. The paper addresses an important problem, and the experimental results demonstrate clear advantages, particularly in computational speed. Therefore, the AC recommends 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.

    Authors have adequately addressed reviewers concerns in the rebuttal.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    I thank the authors and the reviewers for this insightful rebuttal.

    Authors satisfyingly addressed concerns about generalisation to new motion (data curation) and clarification of the positioning of the work with existing literature, and successfully defended their choice of baselines. 


    However, the authors acknowledged a few limitations, namely the lack of sensitivity analysis on K, missing extension to other out-of-domains datasets and organs.

    Overall, this paper introduces a highly novel registration framework which is highly promising. The experimental section lags a bit behind, but 1) it is still reasonable, and 2) it is counter-balanced by the strong emphasis on methodological aspect. Given the very interesting framework and high novelty, I recommend this paper not only for acceptance but also for oral.



back to top