Abstract

Precise and efficient spinal Computed Tomography (CT)/X-ray registration remains a long-standing challenge for image-guided navigation. Methodologies have evolved from intensity-based optimization to learning-based acceleration, transitioning from the traditional “render and compare” paradigm toward “reconstruction and compare” strategies. However, current learning-based methods still face difficult convergence and limited generalization, and a versatile framework capable of accommodating both single- and multi-vertebrae registration is yet to be established. To resolve these issues, we propose FlowReg, a versatile end-to-end framework that leverages CT volume reconstruction from intraoperative biplanar X-rays as a robust spatial prior to guide precise registration. FlowReg utilizes a SE(3)-aware flow matching branch to learn deterministic pose gradients in se(3) while learning shared representations across vertebrae. To bridge the domain gap, we introduce a Difference-Aware Feature Fusion module to reconcile cross-modal disparities. Extensive experiments on synthetic and real clinical datasets demonstrate that FlowReg achieves state-of-the-art performance in both reconstruction quality and registration accuracy, successfully satisfying real-time intraoperative clinical demands. The code is available at: github.com/shenao1995/FlowReg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0758_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/shenao1995/FlowReg

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SheAo_FlowReg_MICCAI2026,
        author = { Shen, Ao AND Jiang, Junfeng AND Fu, Xueming AND Zeng, Qiang AND Tang, Ye AND Chen, Zhengming AND Nong, Luming AND Wang, Feng AND Zhou, Shaohua Kevin},
        title = { { FlowReg: Reconstruction-Guided Flow Matching for Universal Spinal CT/X-Ray 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 proposes an end-to-end framework, FlowReg, for spinal CT/X-ray registration that combines a 3D reconstruction branch with pose estimation. The method uses reconstructed volumes as spatial priors to guide registration and adopts a flow-based formulation for predicting vertebral poses. The framework is designed to handle both single- and multi-vertebrae scenarios within a unified pipeline.

  • 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.Leveraging 3D reconstruction as a spatial prior for 2D/3D registration is a reasonable and potentially effective direction to mitigate the ill-posed nature of projection-based matching. 2.The framework is designed to support multiple vertebrae within a single model, which is relevant for practical clinical workflows. 3.The integration of reconstruction and registration into a single training framework is well-motivated and may provide regularization benefits.

  • 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.In section of “Riemannian Flow Matching Registration”, the formulation appears inconsistent. While the paper claims to operate on SE(3), the actual method performs linear interpolation directly in se(3) without clearly mapping back to the manifold. This suggests a potential misunderstanding of the distinction between SE(3) and its Lie algebra.

    2.The method mixes Euclidean optimization in se(3) with a geodesic constraint in SE(3) without a consistent formulation. Since the trajectory is defined via linear interpolation in R^6 rather than geodesics on SE(3), the claimed Riemannian formulation is not technically valid, and the added geodesic loss appears to be temporary rather than principled.

    3.The DAFF module performs feature difference, element-wise product, concatenation, and MLP-based fusion (Sec. 2), which are standard operations without clear novelty or theoretical justification. Additionally, flow matching for pose estimation has been explored in prior work (as also cited in [12,19]), making the methodological contribution incremental.

    3.The paper claims that flow matching “fundamentally resolves” issues such as local minima and convergence (Sec. 1), but does not provide theoretical analysis or controlled experiments isolating these effects. Similarly, the claim that DAFF “reconciles cross-modal disparities” is not directly validated beyond overall performance improvements.

    4.Although the framework outputs multiple vertebrae poses, it does not explicitly model inter-vertebral constraints or relative motion. The dataset includes only 11 three-level samples (Sec. 3.1), which is insufficient to demonstrate robustness under realistic multi-vertebra motion scenarios. And there is a lack of clear experimental results regarding multi-vertebral registration.

    5.The model is pre-trained on 37,950 synthetic samples generated from 253 CT scans (Sec. 3.1) and then fine-tuned on the real dataset (Sec. 3.2). However, the paper does not clarify whether subject-level splits are enforced, raising potential data leakage concerns. Furthermore, fine-tuning on the target dataset weakens claims of generalization to unseen data.

    6.The real dataset contains only 62 samples (30 single-, 21 two-, 11 three-level cases) (Sec. 3.1), yet five-fold cross-validation is applied (Sec. 3.2). Given the small sample size and imbalance, this evaluation protocol may not provide reliable estimates of performance.

    7.The ablation study (Table 2) combines multiple factors such as DAFF and sim-to-real training (D2), making it difficult to attribute improvements to individual components. For example, improvements from (2) to (3) may stem from domain adaptation rather than the DAFF module itself.

    8.The ground-truth poses are obtained via C-arm cone-beam acquisition (Sec. 3.1), but the relationship between the preoperative CT and intraoperative imaging is not clearly specified. In particular, it is unclear whether there exists any significant inter-vertebral motion between the preoperative CT and intraoperative scans. If the vertebrae remain largely rigid, the task may reduce to a near global rigid alignment rather than true multi-vertebra registration. Furthermore, potential spatial inconsistencies between preoperative CT and CBCT (e. g. , differences in patient positioning, deformation, or calibration) are not discussed. This raises concerns about the accuracy and consistency of the supervision signal, and whether the provided “ground-truth” poses are reliable for evaluating fine-grained multi-vertebra registration performance.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

    (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?

    I recommend reject due to fundamental issues in the technical formulation, limited methodological novelty, and insufficient experimental validation. In particular, the claimed Riemannian formulation on SE(3) is inconsistent with the actual implementation in Euclidean space, and key claims are not well supported by analysis or controlled experiments. Additionally, concerns regarding dataset scale, evaluation protocol, and ground-truth reliability further weaken the credibility of the reported results.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    While the rebuttal clarifies several implementation details, my overall recommendation remains Reject because the main methodological and experimental concerns are not fully resolved.

    The authors acknowledge that the original manuscript conflated optimization in se(3) with geometric modeling on SE(3). The rebuttal clarifies that the method performs linear interpolation in se(3) and only applies exponential mapping and geodesic loss afterward. Therefore, the proposed trajectory is not a true Riemannian/geodesic flow on SE(3), meaning the method description and abstract would require substantial correction. The rebuttal also implicitly admits that several claims in the paper were overstated.

    Second, regarding the DAFF module, I acknowledge that Table 2 demonstrates positive effects in isolated direct-regression and reconstruction settings. However, these experiments do not directly establish DAFF’s contribution within the complete FlowReg framework. Improvements observed in simplified sub-tasks cannot automatically be interpreted as evidence that the module remains essential in the full system, where dominant components may overshadow or subsume its effect.

    Third, concerns regarding data scale and evaluation robustness also remain unresolved. Although the authors state that patient-level splitting was enforced, the rebuttal still does not clearly explain how fine-tuning was performed across folds, nor whether any tuning decisions could indirectly bias evaluation. In addition, the real dataset contains only 62 samples, among which only 11 involve three-level vertebrae cases. Given this limited and imbalanced dataset, five-fold cross-validation may not provide reliable estimates of generalization performance. More importantly, the current experiments still do not convincingly demonstrate robustness in realistic multi-vertebra registration scenarios.

    Therefore, I do not believe the current evidence is strong enough to support acceptance.



Review #2

  • Please describe the contribution of the paper

    This work presents a 2D/3D vertebrae registration approach named FlowReg, which firstly applies the flow matching idea in diffusion model to the X-ray/CT 2D/3D registration problem, and handles both single and multiple vertebrae spinal registrations. Specifially, it learns the deterministic pose gradients directly on the SE(3) manifold and the cross-modal disparities between CT and X-rays. Quantitative studies presented superior registration accuracy over the other methods.

  • 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 application of the Riemannian Flow Matching method to the SE(3) 2D/3D registration is novel. The learning objective is targeted at the key ill-posed registration pose field and mapped from the data-driven manner. This idea is brilliant and from the qualitative presentation (Fig.1), the FlowReg loss field is much more smooth than the other methods. 2.The other two contributions, the DAFF module and the reconstruction branch also presented superior performance to the overall pipeline success and the reconstruction quality from its ablation study. Plus, the framework handles the multiple vertebral motion, up to three vertebrae. This is a significant contribution in a unified end-to-end learning paradigm by enforcing the correspondence learning from the cross-domain feature fusion and the reconstruction overlay check. 3.The presentation and the writing of this manuscript are clear. The experiments are designed with great rigour, and the results presented are solid.

  • 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 CT resolution is limited to 64 voxel per dimension. Given a reasonable volumetric physical dimension of 300 mm, the voxel size is about 5 mm / voxel, which is very coarse. It would be interesting if the manuscript can discuss how this work can generalize to 3D or 2D domain with higher resolution. 2.This method is specifically demonstrated on N <= 3 vertebrae. The authors should clarify if the cropped vertebrae are always at the same level or they are a mix of multiple levels, and how this framework can generalize to or recognize mismatched 3D volume and 2D images. 3.Robustness to image artifacts or other objects. In real clinical scenario, the image quality can be impacted by various factors, such as additional surgical instruments (cables, K-wires, screws) in the scene, it would be interesting to know how the DAFF module can generalize to impaired images. 4.It’s worth noting that the real testing results are performed on an in-house, limited number of samples (<100 samples). This small evaluation subset posts the clinical validation quality at risk.

  • 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?

    The methodology contribution is novel. The presentation is great and the results are solid.

  • 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 author’s rebuttal letter responded to the reviewer’s concerns well. I keep my original recommendation as Accept.



Review #3

  • Please describe the contribution of the paper

    The main contributions of this paper are the unique usage of the Riemannnian Flow Matching Registration as well as the combination and integration of well known techniques into a single framework for fast single and multi vertebrae 2D-3D registration. No papers until now have used 3d generation from AP and lateral views for registering multiple vertebrae. Only single vertebrae reconstruction and subsequent registration exists (Rad-GS). The method was evaluated using 5-fold cross validation on a real-world dataset while using a 253 patient synthetic dataset for pre-training. The paper shows increased performance for reconstruction and registration against different techniques.

  • 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.

    One major strength is that the paper is well written and it’s evaluation is done cleanly, also the ablations are done well. Another major strength is the usage of Riemannian Flows during the registration, it makes sense to formulate the registration of the different vertebrae as a flow problem, which is empirically validated when looking at the results. Furthermore this technique is very fast with only using 0.42s.

  • 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 following minor weaknesses need to be addressed:

    • give some details about the population of the real world dataset the method was tested on i.e. age, gender etc.
    • the details about semantic segmentation and localization of the vertebrae on the real x-rays should be given a bit more context and given some more results as it’s one of the fundaments for the method to work, as it’s not obvious that they work as well as reported on the real dataset this paper is using.
    • in the introduction one could give some more references to feature-based and contour-based approaches as these also have shown to be promising for 2d-3d registration.
  • 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?

    This paper presents a strong new method for multi-vertebral 2d-3d registration that is accurate and fast and marks a strong step towards using this technology in real spine surgeries resp. spine surgery navigation systems.

  • 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.

    The paper introduces a novel and promising frame for multi-vertebrae 2D-3D registration with a strong evaluation and ablation section and should therefore be accepted.



Author Feedback

We sincerely thank all reviewers and the Area Chair for their valuable time, constructive comments, and positive assessment of our work.

1.SE(3)/se(3) formulation and wording [Meta, R1W1-R1W3]: We agree that our wording may have conflated the parameterization space with the supervision space. Following standard RFM practice [12], we parameterize conditional paths in se(3), where linear interpolation defines a tractable vector field for stable pose learning; we do not claim this path is an exact SE(3) geodesic. Predicted velocities/poses are mapped back to SE(3) via the exponential map before rendering, warping, and geodesic-loss computation. Thus, RFM learns pose evolution in se(3), while the geodesic loss enforces SE(3) manifold consistency. We will rephrase “Riemannian flow on SE(3)” as “RFM with se(3) parameterization and SE(3) geodesic supervision,” and soften “fundamentally resolves” to an empirical claim, e.g., “substantially mitigates.”

2.DAFF design rationale [R1W3, R2W3]: DAFF uses standard operations, but is tailored to X-ray/DRR fusion: difference maps capture artifacts/misalignment and Hadamard maps capture shared anatomy. Table 2 rows (2)-(3),(6)-(7) share D2, isolating DAFF gains. Severe occlusions remain a limitation.

3.Inter-vertebral motion and data split [Meta, R1W4, R2W4]: FlowReg does not treat multiple vertebrae as a single rigid body; it predicts semantically indexed vertebra-specific 6-DoF poses. To compensate for limited real multi-vertebra data, synthetic samples from 253 public CTs include local inter-vertebral perturbations following [21] (rotations x±12°, y±6°, z±2°; translations ±2 mm), approximating lateral bending, axial rotation, and flexion/extension. RFM jointly learns pose velocities and implicitly captures inter-pose correlations; explicit biomechanical constraints remain future work. Patient-level splitting was enforced in synthetic pre-training and real-data five-fold validation; public CTs and private clinical data are independent, avoiding real-subject exposure. Each clinical case is evaluated once.

4.GT pose and evaluation details [Meta, R1W8]: Preoperative CT is acquired supine, while intraoperative X-ray/CBCT is acquired prone. This posture change can cause spinal curvature changes and inter-vertebral motion, so the task is not near-global rigid registration. The intraoperative GT pose was established by a hybrid protocol combining optical tracking and manual refinement: reference frames on the C-arm detector and patient ROI were synchronously captured by an NDI Polaris tracker to estimate the initial CT-to-X-ray relationship. Intraoperative C-arm CBCT was used as the surgical 3D reference, and vertebra-by-vertebra CT-to-CBCT registration yielded vertebra-specific transforms to mitigate supine/prone deformation. mTRE averages Euclidean distances between six CT landmarks transformed by GT and predicted poses: the centers of both pedicles and the four outermost vertebral-body corners. Residual errors from skin-motion artifacts and calibration uncertainty may remain, and will be discussed as a limitation.

5.CT spatial resolution and artifact robustness [Meta, R2-W1, R2-W3]: The 64^3 volume is not the original CT resolution or a 300-mm whole-spine volume. VerSe’20 CTs have voxel spacing from 0.19×0.19×0.40 to 1.67×1.67×5.00 mm^3, with median matrix size 512×512×292.Our private CTs are 512×512×401 with 0.3×0.3×0.625 mm^3 spacing. CTs are resampled to a unified physical coordinate system; each vertebra is localized/segmented, cropped with margin, and resized to 64^3 as a vertebra-level spatial prior. Higher-resolution variants can use coarse-to-fine refinement. For artifacts, CLAHE and local masks suppress background; DAFF models DRR/X-ray differences and shared responses, while the 3D reconstruction branch provides a geometric prior, making the model less sensitive to projection noise than render-and-compare methods. Severe tool/implant occlusions remain a limitation.




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 addresses 2d/3d registration for X-ray/CT, proposing a network with a 3D reconstruction branch that informs pose estimation. Strong evaluation demonstrates the advantage of the results. There are potential issues with the presentation of mathematical details (se(3) vs SE(3)) and overly strong claims (fundamentally resolving issues of local minima), but these are minor issues which can be addressed in the rebuttal. R1 raised further concerns inter-vertebral motion. The authors should clarify whether their method treats multiple vertebrae as a single rigid body or if it accounts for inter-vertebral motion. Concerns about subject-level data leakage should be addressed. The real dataset size is sufficient for CAI. Further discussion of CT spatial resolution, evaluation details, and robustness to image artifacts is warranted.

    Given the strengths of the paper, I look forward to the rebuttal and the response to the concerns raised.

  • 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.

    While R1 maintains some concerns regarding mathematical details and evaluation robustness, these are not fundamental issues that would prevent acceptance, and the overall strength of the paper significantly outweighs these observed weaknesses. In light of the strong support from R2 and R3, I recommend acceptance of this paper.

    I will address R1’s remaining concerns here, as I do not consider them to be critical issues that would prevent acceptance:

    1.Mathematical formulation: R1’s concerns about the mathematical formulation of the problem can be addressed in the CR. While some claims were softened in the rebuttal because a fundamental theoretical guarantee is out of scope, the method still represents a significant advancement in the field with strong empirical results. Because the question of “true Riemannian/geodesic flow” is not fundamental to the contribution of the paper, I do not consider this a critical issue that would prevent acceptance.

    2.DAFF effect isolation: While more substantial ablations would further isolate the effect of DAFF within the complete FlowReg framework, other reviewers have not identified this as a critical issue, stating “the studies are performed with great rigor and solid results” (R2) and specifically mentions the strong ablations (R3).

    3.Evaluation: The rebuttal has clarified questions about patient-level splitting during cross-evaluation. While rebuttal does not explicitly address the question of fine-tuning on the real data test set, and this would be a critical flaw if it were the case, I believe it is reasonable to assume that fine-tuning was on each training set during cross-validation, ensuring patient-level separation between training and test sets. The authors should clarify this point in the CR, but I do not consider this a critical issue that would prevent acceptance. If there was fine-tuning on the test set, I strongly recommend the authors to withdraw the paper based on flawed evaluation. However, I do not have evidence that this is the case, and I would give the authors the benefit of the doubt on this point.



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.

    A few issues remain unresolved after rebuttal: 1) Strictly not a Riemannian parametrisation for optimisation. While it’s true that the wording of “we formulate the registration … on the SE(3) manifold, …” needs to be clarified to make it explicit that riemannian constraints are enforced via loss rather than explicit parametrisation - this seems a small adjustment that is fairly acceptable to fix in the final version - and it doesn’t take away from the main results and contributions of the paper

    2) Ablations appear reasonable to me

    3) Cross-validation also appears sound and clear to me. A model is pre-trained on VERSE’20 and then fine-tuning is performed in cross-validation on a proprietary dataset. Main issue is that dataset is small (mitigated by cross validation) and that results are not fully reproducible due to use of proprietary data (which goes beyond the rebuttal discussion).

    I am therefore leaning towards accepting the paper, while urging the authors to clarify the optimisation framework and in which sense riemannian consraints are applied.



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 will align with the accept decisions for this work. Although some points were not fully addressed during the rebuttal process, the response space was limited and the authors appropriately focused on the major concerns raised by the reviewers. In addition, two reviewers were strongly in favor of acceptance, further supporting the overall positive assessment of the paper.



back to top