List of Papers Browse by Subject Areas Author List
Abstract
Skeletal representations (s-reps) provide a powerful tool for anatomical shape analysis by explicitly capturing object widths and boundary curvatures. Existing s-rep construction methods follow an evolutionary framework for gradual deformation from a reference ellipsoid and remain computationally demanding and slow. In this work, we propose a unified end-to-end deep learning framework for joint volumetric segmentation and skeletal modeling directly from 3D medical images. Built upon a general segmentation backbone, the framework adopts a hybrid variational encoder–decoder architecture that integrates skeletal priors with explicit geometric constraints to promote anatomically consistent skeletal representations. Experiments on hippocampus analysis demonstrate that the proposed framework produces s-reps with accuracy comparable to state-of-the-art evolutionary fitting approaches, while substantially reducing inference time from minutes to seconds. Moreover, the learned s-reps preserve discriminative geometric information that enhances downstream Alzheimer’s disease classification and improves segmentation robustness, particularly in limited-data regimes.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4240_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WuYin_Joint_MICCAI2026,
author = { Wu, Yinhao AND Li, Haiqing AND Ma, Hehuan AND Huang, Junzhou},
title = { { Joint Segmentation and Graph-Based Skeletal Representation Learning with Geometric Priors } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes an end-to-end deep learning framework that jointly performs volumetric segmentation and skeletal representation (s-rep) prediction from 3D medical images. The architecture combines a 3D U-Net (soft) segmentation with a variational encoder to capture global anatomical context, a differentiable skeletonization for extracting medial priors, learnable positional embeddings with cross-attention for skeletal node-specific feature extraction, and a GNN decoder that outputs the predicted s-rep. Two novel geometric loss terms, specific for skeletal representations, — spoke direction consistency and spoke-boundary orthogonality — are introduced to enforce anatomically meaningful skeletal properties. The method achieves s-rep quality comparable to the established pipeline, without necessitating an existing segmentation and reducing inference time over 1000x.
- 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 problem formulation is compelling and well-motivated. The existing s-rep fitting pipeline is a major bottleneck for clinical adoption of skeletal shape analysis, as well as necessitates a prior structural segmentation. Replacing it with a learned single-forward-pass approach directly from MRI data is a practical and valuable contribution.
The architectural design is well-done. The combination of a variational encoder for global shape context, a differentiable skeletonization module for medial priors, learnable positional embeddings with cross-attention for node-specific decoding, and a graph convolutional decoder that respects fixed s-rep topology is a principled pipeline where each component serves a clear purpose. The geometric loss terms (spoke direction and orthogonality) are well-formulated and address genuine domain-specific constraints rather than generic regularizers.
The main novelty is in the overall framework, as well as the geometric loss terms that are specific for skeletal representations, spoke direction consistency and spoke-boundary orthogonality .
The ablation study is informative and shows a clear trade-off: adding geometric, skeletal constraints improves skeleton-specific metrics (medialness, angle, orthogonality) at the cost of raw coordinate RMSE. It shows that the geometric constraints push the model towards better skeletal representations even if point-wise error increases.
The downstream classification experiment (Table 2) demonstrates that predicted s-reps retain discriminative power comparable to the existing pipeline in a hippocampus study of AD vs. NC classification (ADNI dataset).
Finally, the results also show that the proposed joint segmentation & skeletal representation framework also improves segmentation accuracy, particularly in data scarce settings. Incorporating s-rep supervision improves Dice & H95 scores.
- 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.
Minor: Limited anatomical diversity. The framework is evaluated exclusively on hippocampal s-reps, which have a relatively simple slab-like geometry. It is unclear how the approach would generalize to other anatomical structures with more complex shape (e. g. , thin-walled or spreading organs, or structures with more branching medial sheets).
Missing comparisons with recent learning-based methods. The paper compares primarily against the evolutionary s-rep pipeline. However, the most relevant competitors, as discussed in the introduction— HybridGNet/HybridVNet and the GNN-based s-rep fitting method of Gaggion et al. — are discussed in the introduction but not included as experimental baselines. Since the paper’s central claim is to improve upon existing approaches, a direct comparison against these methods is essential.
- 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.
- 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.
(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?
While the novelty is iterative, the proposed losses are novel and are clearly helping with the predicted skeletal representation. That is valuable information for other groups working in this area.
The combination of segmentation and representation (improving both segmentation performance and representation performance) is really nice and worthwhile to the audience at MICCAI.
While there are experimental holes (comparison against other learning based methods), the paper does an appropriate job at ablation studies. The results show that the proposed approach is a clear step forward.
- 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 fast means of approximating skeletal representations (s-reps) from images. The integration of image segmentation and s-reps fitting is novel and important in clinical applications. While the methodology and evaluation of s-reps have flaws to be addressed, the research opens a gate to real-time clinical application of s-reps.
- 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.
S-reps plays a critical role in statistical shape analysis. The good correspondence and rich geometric features provided by s-reps allow for statistical understanding of shape data.
This research chose a good direction and proposed a computationally efficient method that connects image segmentation and s-reps fitting in a supervised fashion.
The paper introduced (1) a learnable position embedding and (2) cross-attention layer to connect image segmentation and GNN-based s-rep computation framework. These components act as a bridge that allows the two tasks (i.e., segmentation and s-rep computation) can be trained jointly. But to confirm the effectiveness, it is necessary to look deeper into the performance of each task.
The research showed with experiments that the s-rep computation improves the learning efficiency of segmentation in sparse data scenarios. The shape information encoded in s-reps provides explicit structural priors and thus helps the image segmentation 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 paper is largely well-written. But the notations are not clearly explained. For example, in section 2.2, the lack of notations such as R^{M\times d} can lead to confusions.
As the paper integrates image segmentation and s-reps fitting, it is clearer to visualize the “goodness” of each. For example, a reader may wonder how the training dynamics (e.g., learning curves) look like given the small to medium data size? More practically, how to tune the parameters in equation (1) to achieve good results from both segmentation and s-reps fitting.
It is not clear what features the authors have used from evolutionary s-reps and Liu’s s-reps regarding the experiments shown in Table 2. S-reps fitting with Cross Attention and GNN: 1) About the loss function regarding skeletal geometry: a. the loss of spoke directions should take into account the skeletal radial curvature, as detailed in Damon, Smoothness and geometry of boundaries associated to skeletal structures, II: Geometry in the Blum case. It is important to leverage Damon’s mathematical tool to produce valid spoke direction vectors. b. It should also account for the distance between spokes’ ends and the corresponding boundary, see the paper Liu et al., Fitting unbranching skeletal structures to objects. 2) The evaluation needs to show the quality of the correspondence in the context of more statistical analysis. S-reps is designed to facilitate the statistical shape analysis via good correspondence across objects. Different from Liu’s diffeomorphism-based method of s-reps, the proposed method in this paper maps image all the way to geometric primitives of s-reps driven by the loss function that takes some skeletal geometry into account. Regardless to the quality of the segmentation, it is natural to ask if the resulting s-reps can help statistical tasks such as classification, hypothesis testing, etc. The relevant evaluation of s-reps in this paper is the classification results in Table 2, showing on-par-with performance with the diffeomorphism-based fitting methods. It can be more convincing to have more relevant evaluation and metrics, see the paper as an example Liyun T. et al., Skeletal Shape Correspondence Through Entropy. 3) I wonder if the authors have tried 2-stage framework separating the segmentation and the s-rep fitting.
- 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
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.
(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 research showed significant applications of s-reps, introducing shape priors from s-reps to image segmentation and leading to an efficient segmentation algorithm. While the methodology of s-rep computation has flaws as suggested in the weakness section, the idea and the experiments have well positioned this paper as acceptance. In addition, the paper is well written and easy to follow. I expect follow-up research from this work.
- 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
The paper proposed a novel framework to combine segmentation and skeletonization based on a medical image input end-to-end. It takes advantage of the anatomical information that the specialized skeletonization method generates, and integrate into the CNN statistical feature. By combining the skeleton and the statistical prior, it construct a GNN for skeleton-representation. It also impose shape constrain with loss function combination to encourage better skeleton output.
- 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 method targets an important geometric task and combine the segmentation and skeleton extraction into a unified framework. 2.It achieves great performances compared to recent SOTA methods. 3.paper is well written, well illustrated with good literature coverage. 4.Efficiency improvement.
- 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.author could visualize example(s) of the generated skeleton overlaid with the hippocampus volume in 3D for presentation. 2.Downstream task performance didn’t indicate a necessary performance advantage over other methods. 3.The segmentation evaluation is missing from experiment.
- 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
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?
solid work with enough experiment/performance to show that it works. Experiments has shown that using shape constrain loss help with s-rep. The downstream task shows that proposed method is on-par/better then competing methods while being 100 to 1000 times more efficient. However, further efficiency comparison with Liu. et.al would be nice.
- 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
We sincerely thank the reviewers and meta-reviewer for their constructive comments and for recognizing the strengths of our work, including the joint segmentation/s-rep learning framework, domain-specific geometric losses, improved segmentation performance, and over 1000× speedup over conventional fitting pipelines. We will incorporate the feedback into the camera-ready revision. Below, we respond to the main comments.
To R#1: We thank the reviewer for the constructive suggestions. (a) We agree that the current evaluation is limited to hippocampal s-reps. While the hippocampus is an established testbed for s-rep analysis, our framework can in principle be adapted to other structures with appropriate s-rep templates, including thin-walled or sheet-like organs. Structures with branching medial sheets or variable topology are more challenging and will require topology-adaptive skeletal templates or graph generation mechanisms, which we view as important future work. (b) HybridGNet/HybridVNet are important related works using hybrid graph neural networks for landmark and surface mesh prediction. They motivate the broader direction of learning anatomical geometric representations from images, while our work directly predicts s-rep primitives jointly with segmentation. Since these methods mainly target surfaces, meshes, or landmarks rather than s-rep primitives, the settings are not directly comparable. (c) Following the meta-reviewer’s suggestion, we will also discuss CM-Rep as a continuous medial shape modeling alternative to discrete skeletonization in the revised version.
To R#2: We thank the reviewer for the insightful comments on s-rep geometry and statistical shape analysis. (a) The loss weights in Eq. (1) were selected to balance segmentation accuracy and s-rep geometric validity. We observed that overly large s-rep-related weights could degrade segmentation performance. The final weights were selected by reducing the relative contribution of s-rep losses and performing grid search. (b) For downstream classification, all s-rep-based classifiers used the same skeletal features following “Interior Object Geometry via Fitted Frames”: 61 interior skeletal points and 24 skeletal fold points per hippocampal s-rep, yielding an 8,076-dimensional Euclideanized feature tuple, reduced by PCA to 177 dimensions. (c) Regarding Damon’s radial curvature condition and valid spoke directions, we agree that these are more rigorous theoretical constraints and important future work. We also agree that spoke-end-to-boundary distance is meaningful. We experimented with such an objective, but found it less stable during backpropagation in the current training pipeline. (d) For correspondence, our fixed-topology GNN decoder, node-specific positional queries, and downstream AD/NC classification provide indirect validation that predicted s-reps preserve statistical utility. More systematic entropy/correspondence analyses will be included in follow-up work. (e) We agree that a separated two-stage learning baseline is valuable; however, such designs may introduce additional optimization and error-propagation challenges. We will explore end-to-end learned two-stage variants in future work and revise unclear notations, including dimensions and symbols in Sec. 2.2. To R#3: We thank the reviewer for the helpful suggestions. (a) Table 2 is not intended to claim superior AD classification, but to verify that predicted s-reps preserve discriminative anatomical information comparable to fitted s-reps while enabling orders-of-magnitude faster inference. (b) We clarify that segmentation evaluation is reported in Fig. 3 using Dice and HD95, including the low-data setting where s-rep supervision improves segmentation accuracy. (c) We will add a 3D qualitative visualization showing the generated skeleton overlaid with the hippocampal volume, and strengthen the efficiency discussion, including comparison with Liu et al. in Table 3.
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.
Reviewers summarized strengths of your work, novel aspects and noted that your solution may be of potential interest. Major strengths include the replacing of existing skeletal modeling by a learning-based framework without the need for prior segmentation, the novel end-to-end framework, and demonstration of improved segmentation. Reviewers also liked the combination of segmentation and shape representation with improved performance. Reviewers also mention that the paper is well written, well illustrated and good to follow. Weaknesses are mentioned as a lack of discussion of reproducibility, missing comparisons to previous work, limitations that the framework is only be applicable only to slab-like simple anatomical shape geometries, and lack of missing segmentation evaluation. Would the paper be accepted for MICCAI, authors are strongly encouraged to provide a revision that integrates suggestions for improvements that consider specific weaknesses as discussed in the reviews, e.g. limitations to specific simple types of shapes (?), lack of evaluation of segmentations etc.It is also suggested that authors may include some notes on the CM-Rep framework developed as a continuous shape modeling alternatively to discrete skeletonization.
