Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Computed Tomography (CT)/X-ray registration in image-guided navigation remains challenging because of its stringent requirements for high accuracy and real-time performance. Traditional “render and compare” methods, relying on iterative projection and comparison, suffer from spatial information loss and domain gap. 3D reconstruction from biplanar X-rays supplements spatial and shape information for 2D/3D registration, but current methods are limited by dense-view requirements and struggles with noisy X-rays. To address these limitations, we introduce RadGS-Reg, a novel framework for vertebral-level CT/X-ray registration through joint 3D Radiative Gaussians (RadGS) reconstruction and 3D/3D registration. Specifically, our biplanar X-rays vertebral RadGS reconstruction module explores learning-based RadGS reconstruction method with a Counterfactual Attention Learning (CAL) mechanism, focusing on vertebral regions in noisy X-rays. Additionally, a patient-specific pre-training strategy progressively adapts the RadGS-Reg from simulated to real data while simultaneously learning vertebral shape prior knowledge. Experiments on in-house datasets demonstrate the state-of-the-art performance for both tasks, surpassing existing methods. The code is available at: github.com/shenao1995/RadGS_Reg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2299_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2025/supp/2299_supp.zip

Link to the Code Repository

https://github.com/shenao1995/RadGS_Reg

Link to the Dataset(s)

VERSE ‘20 dataset: https://verse2020.grand-challenge.org

BibTex

@InProceedings{SheAo_RadGSReg_MICCAI2025,
        author = { Shen, Ao and Fu, Xueming and Jiang, Junfeng and Zeng, Qiang and Tang, Ye and Chen, Zhengming and Nong, Luming and Wang, Feng and Zhou, S. Kevin},
        title = { { RadGS-Reg: Registering Spine CT with Biplanar X-rays via Joint 3D Radiative Gaussians Reconstruction and 3D/3D Registration } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15968},
        month = {September},
        page = {458 -- 468}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces RadGS-Reg, a unified deep learning framework designed to achieve precise registration between preoperative CT volumes and intraoperative biplanar X-rays for spine imaging. Even though tested for spine cases the method can be generalized for other parts. The proposed approach performs the registration task in multiple steps. The steps start by first converting biplanar X-rays into a 3D Radiative Gaussians (RadGS) representation and then registering this reconstructed volume with the CT data. To mitigate any issues arising from overlapping vertebrae area, the authors proposed the use of Counterfactual attention Learning (CAL) to improve the reconstruction process. The authors also proposed pre-training the initial three stages. Experiments on a public vertebral dataset (VERSE ’20) and the authors in-house intraoperative dataset was used to assess the proposed methods. The proposed method (as per Table) seems to out-perform similar comparing methods. Ablation study corroborates the results on Table 1 and shows state-of-the-art performance in both reconstruction quality (with SSIM exceeding 94%) and registration accuracy (with mTRE around 1.14 mm), all while maintaining real-time runtime (approximately 0.82 seconds).

  • 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. Figure 1 shows the overall methodology clearly. This to me is a big plus.
    2. The authors have combined 3D Radiative Gaussians reconstruction with 3D registration in a unified framework is novel. These components will enable enhanced spatial information recovery for the reconstruction task compared to traditional “render-and-compare” operations.
    3. Integrating CAL is a strong point. It refines the network’s focus on the vertebral regions, addressing the challenging problem of overlapping structures in X-rays. This targeted approach is backed by both qualitative and quantitative improvements shown in the ablation studies.
    4. The three-stage pre-training makes sure operative unit which contributes jointly to the model.
    5. The method is rigorously evaluated against multiple baselines across both reconstruction and registration tasks. The combination of metrics such as SSIM, PSNR for reconstruction, and mTRE, Capture Range, and Success Rate for registration demonstrates the overall robustness of RadGS-Reg model in table 1 and 2.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. While this a strong work, I think lack of sensitivity analysis of hyperparameters would be beneficial. The manuscript can benefit from a short analysis on if inference time parameter changes help improve the final output of the model.
    2. The manuscript can benefit from including a more thorough discussion regarding potential challenges when applying the system across diverse patient populations and different clinical environments would strengthen the case for clinical translation. Maybe this can be conducted as a future task.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

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

    The paper is well-organized that clearly explains all the steps (three pre-training steps, reconstruction and registration). Visual aids are effective and results shows promising output.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    This work presents a spinal vertebra reconstruction method using radiative 3D Gaussian Splatting, which is then connected with a learning-based 3D registration network to estimate the pose to 3D CT. The key contributions include 1) applying the counterfactual attention learning (Ref.22) to extract the attention maps before the Gaussian head in the 3D reconstruction module. 2) A unified training strategy that combines the 3D reconstruction and 3D registration. 3) Use the vertebrae level segmentation and detection to predicte pose at the rigid vertebrae level in inference.

  • 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.

    This manuscript is very well written and of great clarity. This work presents a unified 3D reconstruction and 3D registration framework, which is trained using data synthesized from CT volumes. The idea of integrating counterfactual maps for enhancing the attention maps and the synergistic training is interesting. The framework is evaluated with intensive comparison and ablation experiments. The results 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.

    In Table 1., the reconstruction performance outperforms the other methods in the literature by a large margin (94.51 v.s. 57.23). The reason is likely due to the unified training framework which includes a CT volume reference in the 3D registration phase. This is not a fair comparison, because there is the 3D volume as a reference during the training and inference time. Essentially, the model encodes the 3D information with the help of the CT, leading to superior performance. This can also be observed in Table 2, where the high reconstruction scores are only affiliated with the joint training framework. The reviewer suggests adding discussions with repect to this point.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (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 novel design of the framework, the quality of this manuscript and the comprehensive evaluation experiments.

  • 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

    ​​1. Synergistic Framework​​: Introduces an end-to-end framework combining 3D reconstruction (RadGS) from biplanar X-rays and 3D/3D registration with preoperative CT, leveraging their mutual benefits to improve accuracy and efficiency. ​​2. Counterfactual Attention Learning (CAL)​​: Addresses vertebral overlap challenges by focusing reconstruction on region-specific vertebrae, enhancing spatial and structural fidelity. ​​3. Data-Aware Training Strategy​​: Implements a three-stage pre-training strategy (simulated -> real -> patient-specific data) to adapt the model to real-world clinical scenarios while incorporating shape priors. ​​4. Model-Agnostic Design​​: Demonstrates flexibility across backbone architectures (ResNet, DenseNet, ViT) while maintaining performance.

  • 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. High Accuracy​​: Achieves state-of-the-art results in both reconstruction (94.51% SSIM, 28.80 dB PSNR) and registration (1.14 mm mTRE). ​​2. Real-Time Efficiency​​: Runtime of 0.82 seconds meets clinical real-time requirements. ​​3. Domain Gap Mitigation​​: RadGS reconstruction bridges the domain gap between DRRs and noisy X-rays, avoiding traditional “render-and-compare” limitations. ​​4. Clinical Relevance​​: Focuses on vertebral-level registration, critical for spinal surgeries, and addresses pose variations between preoperative CT and intraoperative X-rays. ​​5. Innovative Training​​: Synergistic training between reconstruction and registration modules improves joint performance.
  • 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. Dependency on Preprocessing​​: Relies on external vertebral segmentation (X-ray detection and CT segmentation) for vertebral-level input, introducing potential error propagation. ​​2. Limited Data Diversity​​: Experiments are restricted to spinal data; generalizability to other anatomical regions remains unverified. ​​3. Resource Intensity​​: High GPU memory requirements (NVIDIA RTX 4090) may limit accessibility for clinical deployment.
    2. ​​Complex Pipeline​​: The multi-stage training (three pre-training phases) increases implementation complexity. ​​5. Unaddressed Motion Artifacts​​: Does not account for intraoperative patient motion, a common challenge in real-world X-ray-guided procedures.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    1. The integration of CAL, RadGS reconstruction, and synergistic training is novel and addresses critical clinical challenges.
    2. Extensive experiments on in-house and public datasets validate performance, though lack of cross-anatomy validation weakens generality claims.
    3. High accuracy and real-time efficiency align with surgical navigation needs, but dependency on preprocessing tools (e.g., segmentation) could hinder adoption. ​​4. Resource-heavy training and complex pipeline may limit scalability.
    4. Clear ablation studies and open-source commitment enhance reproducibility.
  • 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 appreciate the thoughtful feedback provided by the reviewers and value their recognition of the paper’s well-written quality.

To R#1 Q1. The manuscript refers to other existing methods (references are [9] and [26] in the manuscript) for the setting of hyperparameters. In future work, we will conduct ablation experiments to analyze the hyperparameters mentioned in the manuscript.

Q2. We will discuss the generalization of the method in the revised manuscript, and the details have been replied in R#3Q2.

To R#2 Q1. Comparison with Other Reconstruction Methods: In the current dataset, preoperative CTs are routinely acquired for spinal surgery patients. Our method fully leverages these preoperative CTs as shape priors to enhance reconstruction accuracy. To the best of our knowledge, no existing reconstruction methods explicitly utilize the geometric information from preoperative CT for vertebral reconstruction. Therefore, we can only compare our approach with other existing reconstruction techniques that do not incorporate such prior knowledge.

To R#3 Q1. Dependency on Preprocessing: The ultimate objective of this study is to achieve 2D/3D registration. The preprocessing pipeline of our method primarily involves CT vertebral segmentation and X-rays vertebral detection. The CT spine segmentation approach employed in this work achieves a Dice score exceeding 96%, while the vertebral detection method for X-rays attains an AP50 (Average Precision at 50% Intersection over Union) of over 95% (references are [19] and [14] in the manuscript). Finally, the mTRE presented in Table 2 further demonstrate that the preprocessing pipeline has a minor impact on the overall error in 2D/3D registration.

Q2. Limited Data Diversity and Unaddressed Motion Artifacts: Although the proposed method was only validated on spinal data, we believe that it has potential for reconstruction of skeletal regions characterized by rigid-body properties. As for anatomical areas involving soft tissues or vascular structures, we will incorporate deformation models in future work to accommodate more complex non-rigid transformations.

Q3. Resource Intensity and Complex Pipeline: The current implementation of our method employs ResNet-50 as the backbone network. In future work, we will explore more lightweight architectures to improve computational efficiency. For clinical deployment, existing frameworks such as ONNX, OpenVINO, and TensorRT can further optimize the model’s inference speed and resource consumption. Regarding the pipeline complexity, the primary challenges stem from synergistic training and the three-stage pretraining process.The pipeline complexity will be reduced when training on larger real-world datasets, which we are actively collecting.



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