List of Papers Browse by Subject Areas Author List
Abstract
Learning-based deformable registration enables fast feed-forward inference, yet robustness to limited training data and domain shifts remains challenging. We propose FSE-Reg, which reuses a shared frozen backbone from large-scale pre-trained 3D segmentation encoders as an anatomy-aware feature space for correspondence learning. Built on this fixed representation, FSE-Reg uses a standard coarse-to-fine residual pyramid decoder and a multi-scale feature-level similarity constraint in the frozen feature space, providing anatomy-aware matching cues without segmentation outputs or labels. We further design a lightweight Difference-Product Interaction Fusion module (DPI-Fuse) to explicitly model mismatch and agreement signals for flow refinement. Experiments on abdominal CT and brain MRI benchmarks demonstrate improved accuracy with plausible deformations and strong external generalization without fine-tuning, compared with intensity-driven networks, joint segmentation–registration pipelines, and foundation-feature-based matching approaches. Code will be made publicly available in a future release.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3239_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/alibabadamo-academy/FSE-Reg
Link to the Dataset(s)
AMOS22 Dataset: https://amos22.grand-challenge.org/
FLARE22 Dataset: https://flare22.grand-challenge.org/Dataset/
Abdominal-DIR-QA Dataset: https://github.com/deshanyang/Abdominal-DIR-QA
Mindboggle-101 Dataset: https://mindboggle.info/data
BibTex
@InProceedings{KanHao_FSEReg_MICCAI2026,
author = { Kang, Hao AND Ma, Tai AND Zhang, Haoyang AND Mok, Tony C. W. AND Li, Zi AND Xu, Minfeng AND He, Lianghua AND Wen, Ying},
title = { { FSE-Reg: Enhancing 3D Deformable Registration with Frozen Large-Scale Pre-trained Segmentation Encoders } },
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 FSE-Reg, a 3D deformable registration framework that reuses a shared and frozen 3D segmentation encoder as an anatomically aware feature backbone. The proposed model leverages a large-scale pretrained 3D segmentation encoder backbone in a frozen manner and learns a coarse-to-fine residual decoder for deformation estimation. In addition, by incorporating multi-scale feature-space similarity constraints and DPI-Fuse-based interaction refinement, the method improves registration accuracy by encouraging more natural deformations and achieving meaningful generalization performance without additional fine-tuning of the backbone.
- 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 overall framework and motivation of the proposed method are described in a clear and accessible manner, and the importance of the method as well as its key contributions are generally well explained.
2.In conventional segmentation-assisted deformable registration approaches, output-level cues such as masks or pseudo-labels may be directly affected by registration errors or be vulnerable to domain shift. In this regard, it is interesting that the present work addresses this issue from the perspective of multi-scale feature-level similarity constraints, and proposes a structure that progressively refines the registration field at the feature-representation level rather than directly transferring output-level cues.
3.Compared with joint segmentation-registration learning-based deformable registration methods, which often suffer from high computational cost and complex multi-stage training procedures, this paper presents a comparatively efficient learning pipeline by employing a frozen 3D backbone and introducing decoder designs such as the DPI-Fuse module.
4.The experiments are designed in a sufficiently diverse manner to support the validity of the proposed framework, and the overall performance improvements achieved by the method make the paper reasonably convincing.
- 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.Although the proposed framework provides detailed analyses and meaningful results, the claim of meaningful generalization without fine-tuning on unseen datasets could be better supported by additional experiments with a wider variety of 3D backbones. Such analyses would provide stronger evidence for the authors’ claim.
2.The main novelty of the paper appears to lie in the overall framework design, and among its components, the DPI-Fuse module seems to be one of the key novel elements. However, based on the ablation study, the contribution of DPI-Fuse to performance improvement does not appear sufficiently prominent, which leaves some room for doubt regarding the independent effectiveness of this module.
3.Although the overall results are well presented, Fig. 2 is less effective in illustrating registration quality. The small image size and boxed regions make local differences difficult to compare, and a more direct comparison with the fixed images would improve the qualitative presentation.
- 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.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
As there are many strengths in the proposed method, I have rated as weak accept.
- 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
The paper proposes FSE-Reg, a deformable registration framework that reuses a shared frozen large-scale pre-trained 3D segmentation encoder as the backbone, avoiding the need to train a registration encoder from scratch. On top of this frozen feature space, the method introduces a multi-scale feature-level similarity constraint to provide anatomy-aware matching cues, together with a lightweight Difference-Product Interaction Fusion (DPI-Fuse) module for flow refinement. Although the methodological novelty is incremental, the framework is technically sound and demonstrates strong empirical 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.The proposed framework is cleanly structured: the frozen encoder provides stable multi-scale anatomical features, while the trainable decoder performs deformation estimation on top. The DPI-Fuse module, although lightweight, is a reasonable interaction design that explicitly models both mismatch and agreement cues between fixed and moving features, which is intuitive and well integrated into coarse-to-fine flow refinement. 2.The experimental results are convincing and represent the strongest aspect of the paper. FSE-Reg consistently outperforms competitive baselines across multiple datasets, with particularly clear gains on external datasets (AMOS and DIR-QA) without fine-tuning. 3.The paper includes extensive comparisons across multiple backbones, multiple datasets, qualitative visualizations, early-stage convergence analysis, and ablation studies. The ablation experiments clearly demonstrate the contribution of each component.
- 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 paper emphasizes that only a lightweight deformation decoder is trained on top of the frozen encoder, but the actual computational complexity is not clearly quantified. The paper does not report decoder parameter counts, total trainable parameters, FLOPs, runtime, or memory usage compared with baseline registration methods. Including such analysis would better substantiate the claimed efficiency advantage. 2.Although DPI-Fuse is shown to be effective in the ablation study, its evaluation is limited to removal-based comparisons. The paper does not compare DPI-Fuse against stronger alternative interaction mechanisms such as correlation-based fusion, attention-based fusion, or simpler competitive baselines like concat-only fusion. As a result, it is difficult to determine whether the proposed fusion design is uniquely advantageous. 3.While the framework is technically sound and practically effective, the main contributions are largely based on integrating existing ideas in a well-engineered manner rather than introducing a fundamentally new registration paradigm. The paper’s value lies more in strong empirical validation than in conceptual novelty.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend Weak Accept because the paper presents a technically sound and practically useful registration framework with strong empirical results. The proposed method is well designed, and the experiments are convincing, with consistent improvements over strong baselines across multiple datasets, including robust performance in no-fine-tuning external evaluation settings. The ablation studies are also comprehensive and clearly support the contribution of each component. The main weaknesses are that the methodological novelty is somewhat incremental, the claimed lightweight design lacks quantitative complexity analysis, and DPI-Fuse is not compared against stronger alternative fusion baselines. Despite these limitations, the paper demonstrates clear practical value and solid experimental validation, which justifies acceptance.
- 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 #3
- Please describe the contribution of the paper
This paper introduces a novel image registration framework with three main contributions: using a pre-trained and frozen 3D segmentation encoder as the encoder of the registration network, deep supervision through encouraging warped feature maps to be close to the fixed feature maps, and DPI-Fuse.
- 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.
There are quite a few interesting ideas/results in this work. 1.The authors demonstrate in Fig. 3 that by using a frozen encoder, there is less parameter to train, which results in faster convergence during training. The faster convergence can also be attributed to the fact that using a frozen pre-trained encoder disentangles representation learning (encoder) from learning deformation fields (decoder). Therefore, the decoder only need to focus on predicting deformation fields from already information-rich features. 2.The idea that one can use a pre-trained 3D segmentation encoder as the encoder for the registration network is quite interesting and the motivation is for this idea is strong. The pre-trained 3D segmentation encoder is trained on a segmentation task, and is thus able to extract richer features with anatomy-related information, which is potentially helpful for image registration task. Furthermore, the authors argue that doing so also allow them to perform what they call feature-space anatomical consistency. 3.The motivation for DPI-Fuse makes sense and results in improvement in performance. Specifically, the authors claim that difference captures mismatch and product captures agreement, both of which are important for image registration task.
- 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 novelty of the work is somewhat limited. 1.The idea of using an pre-trained and frozen feature extractor (contribution 1) instead of relying on intensity information is not new. A somewhat related but old idea was introduced in HAMMER by Shen D, et al., IEEE TMI, 2002 (https://doi.org/10.1109/tmi.2002.803111). There was no pre-trained networks back then but the idea is similar, i.e., relying on higher dimensional features instead of intensity information alone. 2.As far as I know, not a lot of work explore deep supervision in the context of image registration. However, the idea of enforcing consistency in the feature-space is still somewhat standard and simple. This essentially boils down to introducing an additional loss term that computes dissimilarity between warped and fixed feature maps at different resolutions.
The clarity of the paper can be greatly improved. 1.At the end of section 1, the authors introduce DPI-Fuse and claim that it is used for “explicit difference/product interactions”. However, it is unclear at this point what difference/product interaction mean and how this interaction is potentially helpful for image registration. 2.Section 2.2 is quite verbose. The entire section can be summarized as follows. Intensity-based dissimilarity is ambiguous while feature-based dissimilarity is more informative. Therefore, an additional loss term based on cosine-similarity is included that quantifies dissimilarity between warped and fixed feature maps at different resolution levels. Consider making this section more concise. 3.The line right after equation (1). What does (\cdot)(p) mean? This notation is not standard and is hard for the readers to understand. 4.In Section 2.3, the “difference term” makes sense. However, it is unclear why the “product term” quantifies agreement. It might be helpful to elaborate on this.
The results are not convincing. 1.In Table 1, it is somewhat strange that VoxelMorph and TransMorph have very poor performance. Are they fully and properly trained to convergence? It would also be helpful to include a baseline using conventional methods (e.g., ANTs’ SyN). 2.The authors claim at the bottom of page 6 that scaling SAT backbone from Nano to Pro improves performance. However, the authors then claim that stronger pre-trained encoder yields stronger performance. This is indeed the case for SAT Nano vs. Pro but is not the case for STU-Net-S/B/L/H. A similar trend is not observed. The authors might wish to provide some explanation for why a similar pattern is not seen with STU-Net. 3.Since the percentage of non-positive Jacobian determinant is relatively large, the authors might wish to include visualization of the deformation fields so that the readers can better understand whether the model is able to produce great performance in terms of Dice/TRE while maintaining smoothness/regularity of the deformation field.
- 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.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This work is very borderline. The three main contributions are not trivial, but they are not substantial either, especially the first two contributions. However, the work presents an interesting idea, i.e., one can directly use (and freeze) the encoder of a pre-trained segmentation network as the encoder of a registration network, which I believe is worth sharing with the rest of the community doing image registration.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
This paper proposes FSE-Reg, a 3D deformable registration framework that utilizes a frozen, pre-trained segmentation encoder as an anatomically aware feature backbone. The reviewers consistently highlight the framework’s efficient design, which simplifies training by disentangling representation learning from deformation estimation, and acknowledge the effectiveness of the multi-scale feature-space similarity constraints. Extensive experiments demonstrate that the method achieves competitive performance and strong generalization across various datasets without additional fine-tuning. However, concerns were raised regarding the incremental nature of the methodological novelty, the lack of quantitative complexity analysis for the decoder, and the need for more rigorous comparisons of the DPI-Fuse module against alternative fusion mechanisms. Despite these limitations, the reviewers reached a consensus that the work is technically sound and practically valuable, providing scores of 4 (Weak Accept), 4 (Weak Accept), and 4 (Weak Accept).