Abstract

Ultrasound (US) is widely used for surgical navigation, and real-time intraoperative 2D US to preoperative 3D US registration is crucial. However, existing methods either lack accuracy, suffer from low efficiency, or are highly prone to overfitting. To address these challenges, we propose a novel and efficient end-to-end real-time 2D-3D US registration framework (EUReg). Specifically, we introduce a cross dimension flow estimator (CDFE) that is both learn-free and differentiable, along with a decoupled transformation prediction (DTP) network. Furthermore, we design a flow loss to supervise the coarse deformation field, effectively decoupling the entire registration process into four sequential steps: feature extraction, coarse deformation field estimation, translation estimation, and rotation estimation. In addition, we improve the differentiable 2D-3D sampling process. We evaluate our framework through comparative, ablation, and exploratory experiments on two public datasets for cardiac and prostate US. Experimental results demonstrate that our method achieves a registration speed exceeding 100 frames per second (FPS) while maintaining high accuracy, meeting the requirements for clinical interventional procedures. Moreover, our exploration reveals that registration accuracy improves when each frame within the volume is larger than the target frame. Our code is publicly available at https://github.com/ZAX130/EUReg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1387_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ZAX130/EUReg

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanHai_EUReg_MICCAI2025,
        author = { Wang, Haiqiao and Wang, Yi},
        title = { { EUReg: End-to-end Framework for Efficient 2D-3D Ultrasound Registration } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {174 -- 184}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    Study proposes an approach for 2D-to-3D registration in ultrasound that can run in real time to rigidly align a 2D frame with a 3D volume. The authors study the utility of the approach in two applications heart and prostate.

  • 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 study is relevant for clinical care and presents a relatively simple framework for identifying first the translations and then the rotations.

  • 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.

    This study is interesting, but is a bit confusing, and oversimplies the problem.

    1. the over simplification comes from registering one frame from to the entire volume, in reality that frame is somewhat different that the frames in the volume, also acquired at ta different angle. while this approach will work in this training setup will it be working in the real scenario, is not clear.
    2. to my point above, the images will have different fields of view, and likely different instances of noise and missing data. maybe this are experiments to consider
    3. The evaluation metrics are quite oversimplifed relative to registration methods too, e.g., using the center of the image and corners to compute DE, that tell us very little about whether the organs, heart or prostate align.
  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

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

    N/A

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

    (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?

    See listed strengths and weaknesses.

  • 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 paper proposes a new method for performing 2D/3D ultrasound registration. Building upon [18], this paper optimizes the sampling method to allow for sub-pixel accuracy, and uses an additional network to convert sampling grid into rigid transformation parameters. The tesulting method achieves significantly better performance compared to previous works.

  • 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.
    • A novel and sensible sampling grid method
    • A novel framework that revolutionizes deep learning-based 2D/3D registration
    • Extensive benchmarking on different datasets
  • 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.
    • It is unclear how the sampling process could become differentiable simply by changing the grid shape. The process of sampling a slice from 3D volume should be undifferentiable. The authors might have meant to say that the prediction process of transformation parameters/grid is differentiable, but still unclear.
  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

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

    N/A

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

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    This is a solid improvement based on previous methods. Its acceptance into MICCAI benefits the research community.

  • 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 rebuttal has addressed my question.

    My only other concern is what R1 commented about the discrepancy between the 3D volume and 2D image. Also, the sampling process itself could create random borders in the image, different from the frame. Not sure how this problem was addressed.



Review #3

  • Please describe the contribution of the paper

    This paper proposed an end-to-end framework for 2D-3D US registration, with four novel parts: cross dimension flow estimator, decoupled transformation prediction, sampling method and flow loss. The extensive experiments with ablation study and comparison results show the effectiveness of the 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.

    This paper proposed two novel modules and one sampling method, together with the flow loss, for 2D-3D US registration. The presentation is clear, and the experiment is extensive.

  • 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.

    Abolition study of the two network modules demonstrate the effectiveness of the proposed network architecture. Without ablation studies of comparing with previous sampling approach in [9] and [18], the effectiveness of the proposed sampling method cannot be guaranteed. The same applies to the validation of flow loss.

    1. “Finally, Td is upsampled to the original resolution to derive T.” - how the “upsampled” is performed?
    2. Does T and R has the same dimension with the image or volume? Better to add this information in the paper.
    3. In Section 2.5, “our approach generates a grid matching the size of Is using πθ and directly samples Iw from Iv” - The approach of generating a grid using πθ is confusing. Do you mean the grid is generated in 3D space (rather than in 2D space?) by translating and rotating the 2D frame into 3D space? More detailed information could be added here.
    4. In section 2.6, “where ϕ∗ is derived from T∗ and R∗”. - better add equations to illustrate this.
    5. What is the “down scaled version”?
    6. The definition of “Initial” in Table. 1 should be added in the paper.
    7. Could you explain the reason why there are no ablation study for experiment iii, iv, v, and vi
    8. Better to add more explanation for Fig. 3.
  • 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.

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    This paper is clearly presented with novel method and extensive experiments. That would be nice if the author could address the comments above and release the code afterwards.

  • 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 authors addressed most concerns raised by reviewers.




Author Feedback

R1: Thank you for your review and for raising the challenges of 2D-3D US registration task. However, there might be some misunderstandings about our work and related research. First, we believe our work should be categorized as MIC rather than CAI, and using MIC’s criteria to evaluate our work would be more appropriate since we propose novel computational modules, a sampling method and a loss for 2D-3D US registration that address limitations of previous methods. The experimental settings i & ii (see Section 3, paragraph 1) follow the configurations established in previous work [9,18] (MICCAI 2021&2024). And building upon this foundation, our model demonstrates substantial improvements over prior methods, as well as generalizability across two distinct organs. From the perspective of methodological innovation, we think these advancements are sufficient to constitute a new work. 1) Simplified problem: Real-time intraoperative pose estimation (translation & rotation) of US frames within 3D volumes has been a longstanding challenge, as detailed in our Introduction section. This fundamental step must be both fast and accurate to serve as the basis for subsequent fine registration. 2) Experimental configuration: Related works [9,13,18,27] also perform registration between 3D volumes and 2D slices sampled from the 3D volume (see Section 1, paragraph 3). The results from [Xu et al., MICCAI2022] demonstrate model consistency across simulated and real data. Besides, the sampled frames always differ from the original frames in the volume, as they are generated by interpolating multiple frames within the volume based on randomly sampled 3D transformation parameters, rather than by selecting a single frame at random and applying a transformation to it. Moreover, experiments iii-vi (Tables 1&2) with varying FOVs and angles still show excellent performance. 3) Evaluation metrics: works [9,18] both use the same computation of DE. For rigid registration, the distance of corresponding points remains the most intuitive alignment accuracy measure. Moreover, we use 7 typical metrics to evaluate distance, consistency, intensity & structure similarity, and speed; see Section 3, paragraph 3.

R2: We appreciate your recognition of our work. The sampling strategy proposed in [9], which extracts 2D slices from 3D volumes, is already differentiable. Our contribution lies in improving the grid generation strategy of [9] while maintaining differentiability, resulting in superior efficiency and accuracy (see Section 2.5 and Fig. 2).

R3: Thank you for your positive feedback. 1) Flow loss: It is exclusively applicable to our CDFE module and, therefore, is not used in the ablation study without CDFE. 2) Upsampling and downsampling: they are used to adjust T to fit the different volume resolutions. Taking t_y from T = {t_x, t_y, t_z} as an example, t_d_y is the downsampled (down scaled) version of t_y, where t_y = t_d_y×(H−1)/(H_v−1). Here, H is the second dimension of the original-resolution volume I_v, and H_v is the counter part of the low-resolution volume feature F_v. 3) T and R: Their definitions are in Section 2.1, both containing 3 parameters. 4) Sampling grid: It is generated simply by multiplying the transformation matrix generated by πθ with an identity grid. Previous works used identity grid of the same size as I_v, while we prove that matching the size of I_s suffices, apparently improving efficiency. 5) Φ*: Its calculation follows Equation 4. 6) “Initial”: It refers to initializing I_w at the center of I_v (the position of I_s in Fig. 2), which means T=R=0. 7) Experiments iii-vi: They don’t include ablation studies because they share identical network architecture with i-ii, primarily testing performance under varied inputs. 8) We will include more explanations of Fig. 3 to show our favorable registration results. 9) Our code will be made publicly available, and all datasets used are publicly accessible, ensuring the reproducibility of results.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    I think the rebuttal answers the main comments raised by Rev 1. Rev2-3 are in favor of acceptance. The code will be released for reproducibility.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top