Abstract

This paper presents a fast, learning-free registration method tailored to clinical constraints. While shape registration is fundamental to medical imaging, existing techniques often force a trade-off between computational scalability and strict topology preservation. Current diffeomorphic methods are either computationally expensive or demand extensive pre-processing, such as training neural networks on large datasets. This renders them impractical for clinical studies, which often rely on limited cohorts. Image-based methods also require large memory banks to process high-resolution data, while surface-based approaches have received less attention in the deep learning era. In this context, we propose a principled registration algorithm that alternates between a feature-aware matching scheme, a conjugate gradient solver and a diffeomorphic shooting step. Our pipeline can handle point clouds, curves and surface meshes. It is compatible with clinical workflows and can register a pair of shapes sampled with 10k points each in less than a second, without landmarks or supervision. We demonstrate consistent results on several anatomical structures and showcase the efficiency of our method by computing an atlas from hundreds of complex vertebra scans in a few minutes. Upon publication, our code will be released under the permissive MIT license at: scikit-shapes.github.io.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3666_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)

VerSe dataset: https://github.com/anjany/verse Lung250M-4B dataset: https://github.com/multimodallearning/Lung250M-4B

BibTex

@InProceedings{BigHad_Optimal_MICCAI2026,
        author = { Bigo-Balland, Hadrien AND Boeken, Tom AND Feydy, Jean},
        title = { { Optimal Steps for Fast Diffeomorphic Shape 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

    This paper presents a fast, learning-free method for diffeomorphic registration of point clouds, curves, and surface meshes. The method combines feature-aware matching, regularized displacement estimation with a conjugate-gradient solver, and an LDDMM-like diffeomorphic shooting step. The stated goal is to achieve topology-preserving registration at a speed compatible with clinical workflows.

  • 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 addresses a highly practical problem: fast, robust, topology-preserving shape registration without requiring large training datasets. This is very relevant to clinical shape modeling workflows, especially for anatomical surfaces where volumetric deep-learning registration may be memory-intensive or require substantial training data.

    The method is technically motivated. It combines ideas from non-rigid ICP, coherent point drift, and LDDMM . The regularization step is described clearly, including the kernel normalization, local loss matrices, conjugate-gradient formulation, and preconditioning. The empirical results are strong. On the VerSe vertebra experiments, the method achieves better Chamfer distance and HD95 than both VoxelMorph and Deformetrica while being around 40× faster than Deformetrica.

  • 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 baseline comparison is limited. VoxelMorph is a volumetric learning-based method and may not be the strongest competitor for surface/point-cloud registration. Deformetrica is a relevant diffeomorphic baseline, but the paper would be stronger with comparisons against CPD, robust optimal transport point-cloud registration, functional maps, non-rigid ICP variants, and recent neural or geometric surface-registration methods.

    The paper claims strict topological consistency, but the evidence is not fully sufficient. LDDMM-style shooting can define diffeomorphic flows in continuous settings, but the implementation is on sampled point clouds/meshes, and the paper does not provide a formal proof or a direct measurement of self-intersections, triangle inversions, or Jacobian positivity. The reported log-Jacobian variance is useful but does not by itself establish topology preservation. There is a slight inconsistency in the presentation: the abstract says the method works “without landmarks, pre-alignment, or supervision,” yet the method section includes a RANSAC-ICP affine pre-alignment and anisotropic bounding-box scaling. This is not necessarily a problem, but the claim should be revised. The experiments are largely focused on vertebrae. Vertebrae are challenging and clinically relevant, but additional anatomical structures with different topology and partial-overlap conditions would strengthen the generality claim. The authors acknowledge partial overlap as future work, which is important because many clinical surfaces are truncated or contain segmentation defects. The computational complexity remains quadratic due to kernel matrix-vector products. The paper reports good runtimes for 10k and 40k points, but it is unclear how the method scales to hundreds of thousands of points or to dense multi-organ atlases.

  • 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.

    (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?

    This is a practical, and well-written contribution. The speed and accuracy results are very good, and the method is attractive for clinical shape analysis. The main concerns are the limited baseline set and the need for stronger evidence supporting the topology-preservation claim.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper proposes a new algorithm for shape registration, where the points are represented as point clouds.

    • First point correspondences are found by creating high-dimensional features followed by nearest-neighbors.
    • Then an optimization problem is solved to smooth out the map
  • 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.
    • Excellent speed up compared to the state of the art and slightly better accuracy
    • Elegant algorithm, robust
    • No neural networks, a big plus for robustness and efficiency.
  • 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.

    None

  • 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.

  • 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.

    (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?

    Excellent paper. Well written, nice algorithms, GPU implementation, good results, smart ideas. A breath of fresh air.

  • 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

    This paper proposes a learning-free registration method for anatomical structures, addressing a key gap in the literature: the need for computationally efficient approaches that do not require large training datasets, a constraint that aligns well with clinical settings. Rather than operating on image-based representations, which tend to be memory-intensive, the method works with point cloud and mesh representations of anatomical shapes. Registration is achieved through an iterative pipeline consisting of a matching step, a linear solver, and a diffeomorphic shooting step, which progressively aligns a source shape to a target. Experiments on two datasets demonstrate that the method achieves registration performance comparable to classical LDDMM while matching the speed typically associated with deep learning approaches. The authors further validate the method’s versatility by demonstrating its applicability to atlas estimation.

  • 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 most notable strength of the paper is its demonstrated combination of high registration accuracy and computational speed, which is particularly valuable in clinical settings where data is scarce, annotation is expensive, and precision is critical. Unlike deep learning methods that require large training datasets, rarely available in medical imaging, this learning-free approach sidesteps that requirement entirely while still matching the speed advantages typically associated with learned methods. Beyond raw performance, the paper makes a compelling case that thoughtfully assembling well-established techniques from the shape optimisation literature, such as normalized kernel matrix from diffusion operators and the LDDMM-like diffeomorphic shooting, an close the speed gap with learned methods without sacrificing rigour or interpretability. The paper is also well-written and the method is easy to follow, which makes it straightforward to understand and build upon.

  • 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 most significant weakness of the paper is its reliance on spatial pre-alignment, which surfaces in two places: the affine initialisation that depends on extrinsic FPFH features, and the exclusive use of 3D coordinates in the two finest scales of the multiscale scheme.

    This raises concerns about the method’s sensitivity to the initial alignment and scale of the input shapes, even when anisotropic normalisation is applied. An ablation study removing the pre-alignment step would have strengthened the paper considerably, as it would clarify how much of the registration performance is attributable to the core method versus the initialisation.

    Related to this, the reliance on 3D coordinates as features implicitly assumes that the source and target shapes share similar topology — an assumption that does not always hold in clinical practice. A straightforward counterexample is a vascular tree where one branch is absent in the source but present in the target.

    The authors do acknowledge partial overlaps as a limitation, but given how frequently topological mismatches arise in anatomical shape analysis, this deserves more explicit treatment.

    Finally, a more minor but noteworthy issue concerns the atlas estimation experiment: since the constructed atlas inherits the topology of the chosen template, the result is inherently biased by that choice. This is not discussed in the paper and would be worth at least acknowledging in the limitations section.

  • 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.

    (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?

    Despite some sensitivity to spatial pre-alignment, this paper addresses a genuine and practically important gap in the literature, a registration method that is both fast and accurate without requiring training data, making it well-suited to the realities of clinical practice. It would be a valuable contribution to the MICCAI community.

  • 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

We want to thank the reviewers and AC for their reviews. We address the main points below, the camera-ready will reflect these clarifications.

Topology preservation (R1). Our shooting Eq. (8) is the LDDMM Hamiltonian flow of Beg-Miller-Trouvé-Younes [4]. With our exponential kernel k(x,y) = exp(-||x−y||/r), the function k(·, q_i) is globally Lipschitz on R^3 with constant 1/r, so the velocity field v_t(x) = Σ_i k(x, q_i(t)) p_i(t) is Lipschitz uniformly in t ∈ [0,1]. By Cauchy-Lipschitz, the flow ODE has unique solutions and φ_t : R^3 → R^3 is a topology-preserving homeomorphism. The kernel normalization Eq. (4) plays an essential role: the spectrum of the normalized kernel matrix is bounded in [0,1] (Kessler-Magnet-Feydy [18, Thm 4.4]), so the kinetic energy is controlled, ||v||² = p^⊤K p ≤ ||p||². At the discrete level, our adaptive Euler step constrains dt so that ||q_i(t+dt) − q_i(t)|| ≤ ℓ, where ℓ is the local sampling length: each step is too small to invert a triangle.

Pre-alignment (R1). The RANSAC-ICP + anisotropic bounding-box scaling is fully automatic and requires no landmarks, supervision, or template-specific tuning, which is what the abstract claims. It is standard procedure in image registration and LDDMM literature where the data can come from different scanners and be aligned differently. This step is applied identically to every baseline. Deformetrica gives very poor results without it. We will modify the abstract wording regarding this aspect.

Baselines (R1, Meta). VoxelMorph is the canonical learning-based reference for medical image registration and Deformetrica is the canonical LDDMM reference. A primary objective of our paper is to demonstrate the massive computational speedup our method offers over classical LDDMM approaches. As shown in our experiments, our approach achieves approximately a 40x speedup over Deformetrica. Furthermore, it accomplishes this while delivering strictly better surface alignment metrics (Chamfer and Hausdorff distances) and preserving better topological integrity, as evidenced by a lower log-Jacobian variance. Comparisons with CPD, robust OT, and non-rigid ICP variants are highly relevant and are intended for extensions of this work.

Anatomical scope (R1, R3). The paper already evaluates two topologically distinct and complex families: 757 vertebrae and 54 lung vascular trees. We are currently partnering with clinicians on ongoing experiments on different datasets to extend the applications of our work.

FPFH and 3D-coordinate matching (R3). The coarse scales (r_1, r_2) use 3D coordinates and FPFH features for robust long-range matching, and the fine scales (r_3, r_4) turn off FPFH and keep 3D coordinates only because by that stage the source already lies within the local convergence basin of the target. This is a multiscale schedule, not a structural assumption of similar topology.

Partial overlap (R1, R3). We identified this limitation in the relevant section of the article and are currently working on it.

Atlas template bias (R3). We will add this to the limitations.

Quadratic complexity (R1). The O(N^2) bottleneck is the kernel mat-vec inside CG and the Hamiltonian shooting. FGT-like schemes and the Giga-scale Kernel Matrix-Vector Multiplication on GPU of Hu et al. are potential lines of inquiry but lack well maintained GPU implementations to date.

Novelty (Meta). A ~40x speedup over Deformetrica at strictly better CD/HD95 is the main claim of our paper. The enabling ingredients are, to our knowledge, new in point-based LDDMM: (i) a Gauss-Newton formulation on a normalized kernel, instead of the quasi-Newton iteration of Deformetrica [5], (ii) the analytic 3x3 Block-Jacobi preconditioner Eq. (7), with closed-form eigenvectors, (iii) CG early termination at tolerance r/10.Combined, these reduce the inner CG loop from 20-30 iterations to 1-5.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 three reviewers accepted the paper with ratings ranging from Strong Accept to Weak Accept. The confidence from two of the accept reviewers is “Confident but not absolutely certain (3) “ while the Weak Accept reviewer is “Very Confident (4)”.

    While the reviewers are overall very positive, I have some reservations regarding the level of methodological novelty. The proposed approach appears to combine previously published components rather than introducing fundamentally new techniques. In addition, the empirical comparison, particularly against VoxelMorph, may not be entirely appropriate due to the need to convert point sets into segmentations, which may result into the shown low accuracy.

    That said, reviewers highlight the solid experimental validation and competitive performance with respect to state-of-the-art methods such as Deformetrica, and find the contribution valuable for the community. I lean towards acceptance, considering the overall positive evaluations and the soundness of the approach, although I agree more with Rv1 comments and Rv3 concerns.

    I would suggest the authors to improve the camera ready reflecting in the manuscript responses to the main comments provided in this review.



back to top