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

Abstract

Deformable retinal image registration is notoriously difficult due to large homogeneous regions and sparse but critical vascular features, which cause limited gradient signals in standard learning-based frameworks. In this paper, we introduce Gaussian Primitive Optimization (GPO), a novel iterative framework that performs structured message passing to overcome these challenges. After an initial coarse alignment, we extract keypoints at salient anatomical structures (e.g., major vessels) to serve as a minimal set of descriptor-based control nodes (DCN). Each node is modelled as a Gaussian primitive with trainable position, displacement, and radius, thus adapting its spatial influence to local deformation scales. A K-Nearest Neighbors (KNN) Gaussian interpolation then blends and propagates displacement signals from these information-rich nodes to construct a globally coherent displacement field; focusing interpolation on the top (K) neighbors reduces computational overhead while preserving local detail. By strategically anchoring nodes in high-gradient regions, GPO ensures robust gradient flow, mitigating vanishing gradient signal in textureless areas. The framework is optimized end-to-end via a multi-term loss that enforces both keypoint consistency and intensity alignment. Experiments on the FIRE dataset show that GPO reduces the target registration error from 6.2px to ~2.4px and increases the AUC at 25px from 0.770 to 0.938, substantially outperforming existing methods. The source code can be accessed via https://github.com/xintian-99/GPOreg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3875_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/xintian-99/GPOreg

Link to the Dataset(s)

N/A

BibTex

@InProceedings{TiaXin_Gaussian_MICCAI2025,
        author = { Tian, Xin and Wang, Jiazheng and Zhang, Yuxi and Chen, Xiang and Hu, Renjiu and Li, Gaolei and Liu, Min and Zhang, Hang},
        title = { { Gaussian Primitive Optimized Deformable Retinal Image Registration } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {218 -- 228}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The main contributions of the paper are:

    1. Descriptor-Based Control Nodes: GPO utilizes a minimal set of control nodes at key anatomical points to enhance gradient flow in sparse retinal images.
    2. KNN Gaussian Interpolation: It employs KNN-based Gaussian interpolation for adaptive blending of local deformations, improving alignment around vascular structures.
    3. End-to-End Optimization: The framework is optimized through a multi-term loss to ensure keypoint consistency and intensity alignment.
    4. Performance Improvement: GPO significantly reduces target registration error from 5.9 pixels to 2.4 pixels and increases AUC at 25 pixels from 0.760 to 0.927, outperforming existing 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.

    The major strengths of the paper on Gaussian Primitive Optimization (GPO) are as follows:

    1. Novel Framework for Deformable Registration: GPO introduces a unique iterative framework that learns a deformable displacement field from a minimal set of descriptor-based control nodes. This approach is innovative as it moves away from traditional methods that often rely heavily on global transformations or are prone to local minima. By focusing on sparse, clinically relevant anatomical features, it offers a solution tailored to the specific challenges of retinal image registration.
    2. Use of Descriptor-Based Control Nodes: The paper emphasizes anchoring control nodes at salient anatomical structures, such as major vessels, which enhances gradient signals in areas that typically lack texture. This novel application of descriptor-based nodes allows GPO to effectively preserve crucial features during the registration process, addressing the issue of inadequate alignment in textureless regions.
    3. KNN-Based Gaussian Interpolation: The introduction of K-Nearest Neighbors (KNN) Gaussian interpolation is a significant innovation, allowing for adaptive blending of local deformations. This is particularly interesting as it strategically focuses on a limited number of neighbors, reducing computational complexity while maintaining critical local detail. It demonstrates a novel method for managing the local versus global transformation balance in image registration.
  • 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. Potential Overfitting Concerns: The reliance on a minimal set of descriptor-based control nodes may lead to overfitting, especially if the dataset used for training and evaluation is not sufficiently diverse. While the authors report impressive performance metrics, it would be crucial to evaluate whether GPO generalizes effectively beyond the FIRE dataset, as it may not represent all possible variations found in real-world clinical scenarios.
    2. Dataset Partitioning: The article does not elaborate on the specific data set division method used for training and testing, such as the ratio of training and test sets, selection criteria, sampling strategy, etc., which may affect the repeatability and generalization ability of the experimental results. The paper’s qualitative results do not demonstrate alignment outcomes in cases with limited overlap between images and fail to present the model’s registration performance across different scenarios. It is suggested that A, P, and S category data should be compared independently rather than as a whole. While overall performance metrics are strong, detailing the model’s capabilities in these specific scenarios is crucial for assessing its robustness and adaptability in real-world clinical applications.
    3. Insufficient Comparison with Existing Deformable Registration Models The keyword in the article title contains deformable retinal image registration. Deformable image registration methods generally refer to non-rigid image registration, while the author’s comparative method is a rigid registration model, which can easily mislead readers and fail to verify the effectiveness of the proposed method. Therefore, the article title and research objectives should be clearer to better reflect the core contribution of the research and its comparative benchmark. On the other hand, it would be more meaningful to more comprehensively verify the effectiveness of GPO and focus on comparing the proposed method with existing non-rigid registration models (such as deep learning-based methods) in comparative experiments. This comparison will help to clearly demonstrate the advantages of GPO in handling complex deformations and enhance the understanding of its applicability in medical image registration applications.
    4. Computational Efficiency Although GPO has improved accuracy, its computational efficiency has not been fully evaluated. When comparing GPO with other existing methods in quantitative experiments, it is recommended to supplement the performance of GPO in terms of running time and computational complexity to confirm its feasibility in clinical practice. The article only analyzes how the number of control nodes, the number of nearest neighbors, and the number of optimization iterations affect model accuracy and running time in the ablation experiment, and does not compare and show the running efficiency of other methods.
  • 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.

  • 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 proposes a new method, Gaussian primitive optimization (GPO), to address the challenges in the current retinal image registration process. Compared with traditional techniques, GPO uses descriptors to control nodes, thereby improving the ability to capture local deformations. This innovative contribution deserves recognition. The framework and implementation process of GPO are described in detail, including the selection and optimization of control nodes. However, the lack of details, especially in the selection of dataset partitioning, training set and test set, reduces the transparency and reproducibility of the results. The choice of comparative experimental methods should mainly compare the model performance of non-rigid registration models, and it is meaningless to compare only rigid registration models. In terms of result presentation, different types of data situations should be evaluated separately, rather than just evaluating the pros and cons of the model overall for the entire dataset. In summary, despite some shortcomings in the article, the innovative methods and experimental results of this article provide new perspectives and foundations for future research. Based on this, I give a “weakly accepted” recommendation, indicating that this work is promising in solving related problems, but the method proposed in the article still needs further improvement and verification.

  • 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 #2

  • Please describe the contribution of the paper

    This study proposed a Gaussian Primitive Optimized (GPO) framework for a retinal image registration. Given a fixed image I_f and a moving image I_m, a descriptor-based network (GeoFormer) is utilized to obtain control nodes, as well as a coarse alignment. Next, initial Gaussian primitives, which possess three sets of learnable parameters (position, displacement vector, radius), can be obtained from control nodes, and the displacement field u(x) can be calcuated according to these Gaussian primitives. Finally, the registered result is obtained via Gaussian primitive optimization, which is guided by losses of global cross-correlation and intensity pattern. A quantitative evaluation experiment was conducted, and the proposed method obtained the best scores.

    Generally, the method can be considered as a post-processing module, which can be appended to any existing registration work, such as GeoFormer. And it achieved better scores in the evaluation experiment, compared to the existing study. The whole paper is well-organized.

    While the following limitations (weak points) in this study can be listed:

    • It missed the state-of-the-art method, RetinaRegNet, for retinal image registration. The RetinaRegNet shows competitive results with the proposed method in their paper, which is available in Jan. 2025. Comparison is recommended. https://www.sciencedirect.com/science/article/pii/S001048252401730X
    • The evaluation experiment involves one dataset “FIRE”, while it is much common to evaluate on “FIRE” and “FLoRI21”. please clarify.
    • The definitions of two losses, L_gcc and L_ncc, in Section 2.3 have not been given.
    • Please proofread the manuscript since gramma errors exist, for example, the sentence below eq. (5) …, L_ncc is aligns overall intensity patterns …
  • 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 new framework for retinal image registration based on gaussian primitive optimization
  • 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.

    While the following limitations (weak points) in this study can be listed:

    • It missed the state-of-the-art method, RetinaRegNet, for retinal image registration. The RetinaRegNet shows competitive results with the proposed method in their paper, which is available in Jan. 2025. Comparison is recommended. https://www.sciencedirect.com/science/article/pii/S001048252401730X
    • The evaluation experiment involves one dataset “FIRE”, while it is much common to evaluate on “FIRE” and “FLoRI21”. please clarify.
    • The definitions of two losses, L_gcc and L_ncc, in Section 2.3 have not been given.
    • Please proofread the manuscript since gramma errors exist, for example, the sentence below eq. (5) …, L_ncc is aligns overall intensity patterns …
  • 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.

  • 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?

    Plus points:

    • new framework based on Gaussian Primitives Optimization
    • achieved best scores in the evaluation experiment Minus points:
    • missing citation to RetinaResNet
    • limited number of datasets in evaluation experiment
  • Reviewer confidence

    Somewhat confident (2)

  • [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 work propose a key-point based framework for deformable retinal image registration. Given a set of initialized key point pairs from the fixed and moving images, the model calculates dense deformations based on Gaussian blending of K closest key points for each pixel. Each key point has optimizable parameters including location, displacement and impact radius, allowing for iterative optimization. Experiments show that the method with a powerful descriptor-based model for initial registration could produce the best performance compared to diverse baselines.

  • 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 Gaussian blending + KNN for dense deformation calculation looks very reasonable for key point based registration, in terms of rationale, design, and complexity.
    • The results cover multiple baselines of different types of strategies.
    • The ablation study provides sufficient details about effect of hyperparameters.
  • 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 method is combined with GeoFomer to produce results compared with other methods, which could not be very fair because other methods seem to be all performing registration individually. I’m not very familiar with key-point based registration, but I believe there are existing methods that directly optimize registration based on a group of initialized key point or key point pairs. If that is true, comparing the proposed method with such methods, where all methods are combined with GeoFomer for initial registration and key point initialization, should be necessary to better prove the advantages of the proposed method.

    • I don’t see the part about how the method gaurantees that the key points focus on crucial vessel structures such as bifurcations. Seems that the key points are initialized by any chosen method, and during later optimization there is no explicit loss term to force the position g_i to be close to important structures. Therefore, the proposed method is more like a general framework for iteratively optimizing key points for deformable registration, rather than specificlly handling the issue of ignoring crucial image structures. If the method can do this, please add more explanation about how to achieve this by model design.

    • It would be better to provide more details (e.g., explicit formulas) about the global cross-correlation loss L_gcc, such as what are the inputs and how the calculation is done.

  • 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.

  • 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?

    My main concerns is the fairness of the comparisons, and whether this method can explicitly gaurantee the registration considers crucial structues in the images. I anticipate to see more related explanations or clarifications.

  • Reviewer confidence

    Somewhat confident (2)

  • [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

N/A



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