List of Papers Browse by Subject Areas Author List
Abstract
Medical image registration relies on the overlapping regions between two images to calculate transformation parameters, thus posing a significant challenge for image registration with limited overlap. To overcome this challenge, this study proposes an image expansion solution by generating more overlapping regions to improve the registration performance between images with minimal overlap. As this is the first study to expand images for registration, we trained a generative network from scratch to avoid chaotic structures in the expanded regions. We proposed the Sequential Structure-Preserve Expansion (SSPE) framework to realize the expansion-based registration, where each image is present by a sliding scope and its expansion can be observed by sliding the scope. When given the current image and a sliding step, SSPE utilizes a generative network to predict the scope content of the next sliding position. Specially, we also bring in the gradient matching to maintain anatomical structures in the predicted scope. The performance of SSPE is evaluated on a public dataset of total-body CT images, which proves that our SSPE is significantly efficient in solving the registration difficulties caused by insufficient overlapping regions. The codes of our framework are made available at https://github.com/YongshengPan/Structure-Preserve-Expansion, and we will also publish software for user-friendly access and testing.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3991_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YongshengPan/Structure-Preserve-Expansion
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuZai_StructurePreserve_MICCAI2025,
author = { Liu, Zaiyuan and Pan, Yongsheng and Xia, Yong},
title = { { Structure-Preserve Expansion for Medical Image Registration with Minimal Overlap } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15963},
month = {September},
page = {539 -- 549}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of the paper is the introduction of Sequential Structure-Preserve Expansion (SSPE), a generative framework designed to improve medical image registration in cases with minimal or no overlapping regions between image pairs.
- 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 introduction of image expansion as a solution to minimal-overlap medical 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.
-
Limited Generalizability to Common Registration Scenarios. The approach is not evaluated on standard benchmarks such as the CT-to-CT registration task in the Learn2Reg challenge, which represent real-world clinical use cases. Without validation on such benchmarks, the practical relevance and general utility of the method remain unclear.
-
Heavy Dependence on Data Availability and Anatomical Continuity. The method requires training a 3D generative model to extrapolate anatomical structures across adjacent regions. This assumes (1) access to large-scale datasets with full-body or continuous CT volumes, and (2) consistent anatomical patterns between training and test subjects. However, such data is often unavailable or inconsistent, particularly in rare diseases, retrospective cohorts, or multi-center studies.
-
No Comparison with Existing Methods that Partially Address Minimal-Overlap Registration. The paper lacks comparisons with alternative approaches, such as feature-based or zero-shot registration methods, which can partially handle registration under limited overlap conditions.
-
- 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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed idea is interesting; however, it lacks validation of its applicability and clinical relevance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The authors addressed my questions regarding the training of the model. However, my major concerns, specifically model evaluation and the clinical usefulness of the proposed method, have not been fully addressed. MICCAI is a conference focused on methodological innovation with strong clinical relevance. Although the authors mentioned the HNSCC dataset in their rebuttal, they did not include this dataset in the results of their intial submission. The only registration result is shown in Figure 4.
Review #2
- Please describe the contribution of the paper
Addresses the image registration problem challenge when the overlap of input images is limited.
- 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.
Addresses a novel problem of how to register images which have minimal overlap and whether this can be solved using an out-painting GAN approach.
- 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 authors do not reference any related papers where GANs are used to improve registration, or explicitly state that this is the first such study.
- 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 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
Can you define the term “tomogram count” (page 4)? Is this simply the number of slices in the delta direction? The term “partial volume” is often used to refer to mixtures of tissue types in a single voxel. Perhaps there is an alternative term or phrase that could be using instead such as “volume of interest” VOI…? For LaResGAN, LaTransGAN, and LaResTransGAN what does “La” refer to? Figure 3 shows “sagittal and coronal views”, not “coronal and axial”. It would be informative to include details of the actual overlap (eg. percentage) of your test images or size of the non-overlapping region (in mm), to help interpret your results.
- 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?
Addresses a novel problem and demonstrates proof of concept of a method to solve this. Evaluation of registration performance is purely qualitative for four cases.
- 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 study addresses the challenge of limited overlapping regions in medical image registration. The authors propose a Sequential Structure-Preserving Expansion (SSPE) approach to extend the images, thereby enhancing the overlapping areas and facilitating more accurate registration.
- 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 authors present a clear problem statement and a detailed description of their proposed framework. All technical and data-related details are provided, ensuring the reproducibility of the experiments. The proposed method is compared against several state-of-the-art approaches, accompanied by a well-conducted analysis. Overall, the novel SSPE framework addresses a significant challenge in the field of medical 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.
There are many typographical errors in the abstract and throughout the manuscript (e.g., “TranUnet” in Section 2.2). I recommend a proofreading of the entire document to correct these issues.
- 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.
(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?
This is an excellent study that is likely to make a valuable contribution to the MICCAI community
- 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
We sincerely thank all reviewers for their valuable and encouraging feedback. We address the specific points raised below. R1Q1: Generalizability to Common Registration A1: The Learn2Reg Challenge primarily focuses on the pre-operative to post-resection registration, which is a traditional registration task mainly based on sufficiently large overlapping regions. This differs from our registration challenge with minimal overlap, which cannot be directly solved by traditional registration methods. Beyond such registration scenarios, our extrapolation strategy provides sufficiently overlapping regions for images with minimal overlap, thereby making traditional registration techniques applicable. Such minimal-overlap registration is also a common scenario but has been overlooked by previous studies. For example, in the HNSCC dataset, head and thoracico-abdominal CT images of 20 individuals were scanned separately with relative motion and minimal overlap. It was previously difficult to automatically stitch them together, but this can now be addressed by our extrapolation strategy. R1Q2: Dependence on Data Availability A2: This may be a misunderstanding here. Our extrapolation strategy iteratively expands each image and only requires its size to be larger than the input. In our experiments, the input size was 96×256×256 and the expansion step was 46. The requirement for image size for training is larger than (96+46)x 256x256, and for application, it is larger than 96x256x256, which is satisfied by most common CT images. Therefore, most common CT images can be used to train the model, rather than only whole-body or continuous CT volumes. Meanwhile, it should not undermine the demonstrated effectiveness and novelty of our study if potential limitations are also faced by existing registration approaches. Moreover, we acknowledge that consistent anatomical patterns are beneficial, but this is a common consideration for any learning-based medical image analysis technique and does not uniquely limit our methods. R1Q3: Comparison with Existing Methods A3: While feature-based or zero-shot methods can handle some specific cases with limited overlap, they are fundamentally extensions of traditional registration paradigms that rely on existing overlapping features. Their performance degrades significantly as overlap decreases. Our SSPE method is novel in tackling the absence of sufficient overlap by creating it through image expansion. Once overlap is generated, standard registration methods can then be applied effectively. As we have demonstrated that with the increased overlap provided by SSPE, even a simple linear registration method achieves successful alignment in cases where it previously failed, it is not necessary to compare it to these specific registration methods. R2: Typographical Errors A4: We appreciate your attention to detail. All typographical errors will be corrected during the proofreading process. R3Q1: Citation of Relevant Work A5: We are the first to achieve minimal-overlap registration by extending medical images. Existing GAN-based registration methods are designed mainly for cross-modal registration tasks. We will cite these studies to provide better context. R3Q2: Explanation of Certain Phrases or Terms A6: The “tomogram count” refers to “slice count”. The term “partial volume” simply refers to a portion of the CT volume and can be replaced with “volume parts”. “La” represents λ, a hyperparameter in the model that enables the model to focus more on non-overlapping regions during the expansion process. R3Q3: Result Display Issue A7: In the final result display, the overlap is 9.52%.
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.
Reject
- 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’
N/A
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’
The paper tackles a challenging, relevant and conceptually novel problem of aligning mismatched anatomies, which can be encountered in clinical scenarios. The technique of using a GAN to generate missing anatomy and then performing the alignment is new. The results show feasibility, although more extensive analysis with known datasets like Learn2Reg would further enhance validity of the approach. Nevertheless, the conceptual novelty and the technical difficulty of the selected problem with reasonable results shown in the paper, outweigh the weakness of lack of additional benchmarking with Learn2Reg datasets. Authors should acknowledge this as a weakness and also discuss the limitation of requiring datasets covering entire anatomy to train the GAN model.