List of Papers Browse by Subject Areas Author List
Abstract
Although medical image registration is an effective tool for quantifying organ motion, its application in assessing postoperative swallowing function in tongue cancer remains limited. Recognizing that existing methods often fail to fully exploit the semantic priors from segmentation tasks, we propose an end-to-end joint learning framework for segmentation and registration on dynamic cine-MRI sequences. By leveraging semantic priors from the Segment Anything Model to guide the registration, it becomes possible to precisely model the complex post‑operative motion of the tongue. Our contributions can be generalized into three aspects:(1) We construct a deeply supervised SAM feature adaptor, where the U-Net encoder generates a multi-scale semantic feature pyramid to provide hierarchical anatomical priors for the registration task. (2) We design a novel convolutional unit combining DCNv3 and Swin-transformer to capture local-to-global features; a bidirectional coordinate attention fusion module is further introduced to adaptively fuse multi-scale features from both the SAM feature adaptor and registration encoder. (3) To directly enhance registration accuracy along anatomical boundaries, we propose an EdgeWeightedIoU loss that employs a Laplacian operator to extract and refine these boundaries from the segmentation output. Ablation studies show that our SAM scheme achieves a 7.55% Dice improve-ment on the sagittal tongue MRI dataset. In SOTA experiments on the coronal tongue MRI dataset, VoxelMorph, TransMorph, and our Baseline gain 3.93%, 35.22%, and 3.32% in Dice respectively after integrating our SAM scheme. Con-sistent improvements are also observed on the ACDC cardiac MRI da-taset . Our code is available at: https://github.com/FFigo/SwallowReg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2405_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/FFigo/SwallowReg
Link to the Dataset(s)
SwallowReg Tongue Cine-MRI Dataset: https://doi.org/10.5281/zenodo.20724549
BibTex
@InProceedings{TanZhi_SwallowReg_MICCAI2026,
author = { Tang, Zhiwen AND Sun, Minghao AND Jiang, Chenghui AND Song, Xiaomeng AND Liu, Lulu AND Yu, Han},
title = { { SwallowReg: SAM-Assisted Deformable Registration with Adaptive Global-Local Features for Cine-MRI Swallowing Function Quantification } },
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
The paper proposes an end-to-end joint segmentation and registration framework for tongue motion analysis in cine-MRI, integrating SAM-based semantic priors with a hybrid CNN–Transformer architecture to improve anatomical alignment and motion estimation.
- 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 paper addresses a clinically relevant problem, namely tongue motion analysis for swallowing assessment using cine-MRI.
- The proposed method introduces a joint segmentation and registration framework, leveraging SAM-based semantic priors, which is a meaningful and relatively novel integration for improving region-specific alignment.
- The model design is technically comprehensive, combining CNN and Transformer components with multi-scale supervision, and is supported by an ablation study to analyze different modules.
- 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 performance improvements over existing methods are relatively small, and no statistical analysis is provided, making it difficult to assess their significance.
- The use of the ACDC cardiac MRI dataset is not well justified, and the stronger improvements on this dataset raise concerns about the relevance of the evaluation to the targeted clinical application.
- Parts of the methodology and result interpretation lack clarity and sufficient support, which affects readability, reproducibility, and the strength of the conclusions.
- 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.
- 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
1.Introduction:
- The claim that “Current tongue movement quantification often relies on manual annotation of tongue structures and subsequent estimation of movement parameters by clinicians, which is time-consuming, inefficient, and associated with inter-observer variability. “ is not supported by references. Please cite relevant literature.
- II recomment to briefly discuss existing semi-automatic or automatic approaches for tongue motion analysis to clarify how the proposed method advances beyond them.
2.Methodology:
- The methodology is technically rich and appears well-structured. However, parts of the description are difficult to follow, particularly regarding the interaction between the different modules (e. g. , SAM Feature Adaptor, Swin-Intern block, and BCAF). I recommend a more explicit description of the data flow to improve readability and reproducibility.
3.Experiments: 3.1.Dataset and Criteria:
- The inclusion of the ACDC cardiac MRI dataset is not sufficiently justified. Since the proposed method is designed for tongue motion analysis, it is unclear why a cardiac dataset with different anatomical and motion characteristics is used. The authors should clarify the motivation for this choice and explain how it supports the evaluation of the proposed approach.
3.2.Ablation Experiments:
- The ablation study is comprehensive and shows consistent performance improvements. However, some explanations (e. g. , why SAM performs better on coronal data or the trade-off between Dice and GNCC/SSIM) are not well supported and seem speculative. More evidence or clearer justification would strengthen this part.
3.3.Performance Comparisons with SOTA
- The reported performance improvements over existing methods are relatively small (differences are often around ~1%), but are described with very strong claims (e. g. , “most significant gains” and “optimal performance”). A more balanced discussion is needed, and statistical analysis should be provided to show whether these improvements are meaningful.
- The relevance of the ACDC cardiac dataset to the targeted tongue motion task remains unclear and should be better justified. In addition, the main performance improvements appear more pronounced on the ACDC dataset than on the tongue MRI data, which raises concerns about how well the proposed method addresses the intended clinical application.
- 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an interesting methodological approach and addresses a relevant clinical problem. However, the overall performance improvements over existing methods are relatively small and lack statistical validation, making it difficult to assess their significance. In addition, the use of an unrelated dataset (ACDC) is not well justified, and the stronger results on this dataset raise concerns about the relevance of the evaluation to the intended application. Finally, the clinical feasibility and impact of the proposed method are not sufficiently discussed, which limits the practical significance of the work.
- 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 paper addresses a clinically relevant problem; however, as a MIC paper, it lacks several important aspects, including sufficient methodological clarity, substantial performance improvements over existing methods, and appropriate statistical validation.
Review #2
- Please describe the contribution of the paper
The authors propose a framework for jointly learning segmentation and registration on dynamic cine-MRI sequences of moving tongues. By leveraging a pre-trained SAM encoder, the feature extractor generates multi-scale semantic features as priors for the registration task. The registration network then fuses the features from the feature adaptor and registration encoder via a novel convolutional unit. The authors provide ablation studies illustrating the effectiveness of various parts of the pipeline, as well as comparisons with SOTA methods, showing its superior 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.
Novelty in Architectural Design: The authors propose two novel blocks in their work: the Swin-Intern Block (for balancing local deformation and global semantics during feature extraction) and the BCAF module (for fusing features).
Strong Evaluation: The authors provide an ablation study clearly indicating improvements across various sub-models, as well as comparisons with various SOTA methods. The proposed framework outperforms all other SOTA methods. Both quantitative and visual results were provided.
- 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.
Missing limitations and future work: The paper does not include discussions of failure cases or directions for further improving the framework.
Reproducibility: No code was provided with the submission, so the paper needs to include enough details for others to reproduce the reported data. Several implementation details were missing, such as parameters used in the loss functions. It would also be interesting to report on metrics such as the training resources used and the training time.
- 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.
- 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
It would be clearer to spot the difference if ground truth were added to Fig. 3 and Fig. 4.
- 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 authors present a novel framework to model the complex post‑operative motion of the tongue. The paper clearly explains all components of the pipeline and the intuition of the overall architecture. It shows superior performance compared to other SOTA methods.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes a SAM-assisted joint segmentation and deformable registration framework (SwallowReg) for cine-MRI tongue motion analysis. It integrates semantic priors from a foundation segmentation model (SAM) into the registration process through a multi-scale feature adaptor, enabling improved anatomical alignment. Additionally, the method introduces a hybrid global-local feature extraction module (combining DCNv3 and Swin Transformer) and an edge-weighted IoU loss to enhance boundary accuracy and motion field robustness. Experimental results show consistent improvements over baseline registration methods on tongue MRI and ACDC datasets.
- 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.
- Clear problem motivation (clinical relevance)
- Well-structured architecture
- Integration of SAM is practical
- Multi-task learning design is reasonable
- 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.
- Incremental novelty (not fundamentally new)
- Limited dataset size (weak generalization)
- Lack of theoretical justification
- Missing recent SOTA comparisons
- No statistical validation
- Slight over-engineering (too many modules)
- Writing quality needs refinement
- 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.
- 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
Here are practical improvements the author can do quickly. 1.Lightweight Ablation (Remove SAM → show drop clearly, Remove BCAF → show importance) 2.Generalization Test (Train on one dataset → test on another (ACDC → Tongue))
- 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?
The paper proposes a SAM-assisted joint segmentation-registration framework with hybrid global-local feature modeling. While the approach is technically sound and shows consistent improvements over baseline models, the novelty is incremental and lacks strong theoretical justification. Experimental validation is limited by small datasets and absence of statistical analysis. The work is promising but requires stronger comparison with recent SOTA and deeper methodological insight for acceptance in top-tier venues. E.g., Add Statistical Significance Test, SOTA Comparison (MedSAM-based method)
- 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 author feedback is acceptable.
Author Feedback
We thank the reviewers for the critical comments and ratings (3,5,4) with detailed responses below. To clarify the Data Flow, the framework takes moving and fixed images as input. (1) They respectively enter the SAM Adaptor, where a frozen ViT-B encoder produces embeddings fed into a U-Net to yield four-scale features and segmentation masks. (2) In parallel, their concatenation enters the Registration Encoder (Swin-Intern Blocks), yielding same-scale features. (3) Both enter BCAF, yielding two mutually enhanced feature types per scale. (4) The Decoder uses these as skip connections to produce the deformation filed φ to warp moving images by STN. [R1; R3] On novelty and module design. (1) Paradigm: using SAM features to boost registration is a novel cross-task knowledge transfer. (2) Module/loss: BCAF bidirectionally fuses SAM and encoder features per scale; the edge-weighted IoU loss uses SAM boundary cues via adaptive weighting. Both are SAM-distilled components. Table 1 ablates each component (Row 1=our Baseline: a Swin-Intern four-scale encoder with skip connections progressively decoded to displacement field with no SAM; Row 6=full model). Sagittal: edge-weighted IoU loss +2.32% Dice (Rows 1→3); BCAF vs. concat. +2.98% (Rows 5→6); full +7.55% over Baseline (35.17→42.72), confirming necessity of each module. [R1; R3] On limited SOTA comparison and marginal gains. Rows 1–3 in Table 2 are SOTA methods (CNN/VM (2019), Transformer/TM (2022), Diffusion-assisted/LDM (2025)); Row 4 is our Baseline(separately trained). Rows 5–7 (+SAM) only replace the Registration Net with our framework. Taking VM as an example, coronal Dice+3.93%,GNCC+7.37%,SSIM+6.92%; same for TM and Baseline, confirming gains across different registration networks.[R1; R3] On the ACDC dataset and data scale. (1) ACDC records cardiac ED-ES Cine-MRI; Xiao et al. (2023) also used it for registration and cross-organ generalization. (2) However, ACDC is cardiac-only where tongue MRI spans the whole head with complex moving tissues. In the last row of Table 2, tongue Dice shows +4.75% over pre-registration (91.26→96.01), comparable to ACDC’s +5.95%. (3) Our dataset requires specialized equipment and privacy consent, with very few public dataset. [R1] On missing citations on annotation costs and tongue motion methods. Sun et al. (2024) noted substantial annotation labor in tongue Cine-MRI; Lee et al. (2014) proposed a semi-auto approach with user seeds on a few slices via random walker and deformable registration for motion analysis; Yang et al. (2020) proposed a fully auto approach using TPS robust-point-matching to register contours across Cine-MRI frames and extract deformation fields for motion quantification. They will be added to the final manuscript. [R1] On speculative explanations in the ablation study. (1) On SAM features’ coronal advantage, rows 1 vs. 2 in Table 1: SAM features surpass encoder features on all coronal metrics (Dice +0.17%, GNCC +0.36%, SSIM +1.58%) while encoder features surpass on all sagittal metrics (Dice +1.49%, GNCC +1.35%, SSIM +2.29%), showing complementarity. (2) On the Dice vs. GNCC/SSIM trade-off. BCAF boosts sagittal Dice (+2.98%, Rows 5→6) while GNCC (−0.17) and SSIM (−0.46) drop, reflecting different metric sensitivities. Dice benefits from BCAF’s boundary refinement; GNCC and SSIM measure global intensity correlation and local structural consistency-both more sensitive to boundary deformation gradients. As the primary motion axis with larger deformations, the sagittal plane amplifies these gradient effects, marginally impacting GNCC/SSIM but substantially boosting Dice. In the coronal plane (Dice +0.42%, GNCC +0.03%, SSIM +0.20%), all improve, confirming a sagittal-complexity trade-off, not a BCAF deficiency. [R1; R2; R3] Additional points. (1) Reproducibility: code/data will be released on GitHub if accepted. (2) GT contours in Figs. 3–4 are unlabeled and labels will be added. (3) Writing quality: will be carefully revised.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
Critical concerns have been raised by all the reviewers about limited improvement, unclear theoretical justification and presentation, insufficient details, and missing experimental comparisons. The authors are invited for rebuttal about the concerns raised by the reviewers.
- 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.
After carefully evaluating the authors’ rebuttal, the reviewers’ revised feedback, and re-examining the paper, I conclude that the authors have effectively addressed the key concerns raised, and the work now meets MICCAI’s publication standards.
Meta-review #2
- 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.
The rebuttal does not adequately address reviewers’ concerns. I agree with R1.The paper still lacks methodological clarity and substantial improvements over existing methods. More empirical evidence is required; with the reported results, it’s hard to justify the methodological superiority.
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.
Reviewers recognized the clinical relevance of swallowing function quantification and appreciated the integration of SAM-derived semantic priors with global-local feature modeling. Following the rebuttal, two reviewers were satisfied and maintained or changed to Accept, noting that the authors clarified the data flow, module contributions, use of the ACDC dataset, and ablation results. However, one reviewer remained unconvinced, emphasizing that the paper still lacks sufficient methodological clarity, substantial performance improvements over existing methods, and appropriate statistical validation. Overall, the rebuttal addressed several presentation and interpretation concerns and strengthened the case for the proposed modules, but some limitations remain regarding statistical evidence, reproducibility details, and the modest scale of demonstrated improvements. Given the post-rebuttal reviewer balance and the clarified methodological contribution, I recommend acceptance.