Abstract

Segment Anything Model (SAM) adaptation has shown remarkable performance in medical image segmentation, but typically relies on large and precisely annotated datasets. However, acquiring such dense annotation is a labor-intensive and time-consuming task that requires significant expertise. An effective direction is to focus on sparse annotation, where only a few slices are annotated. However, sparse annotations are insufficient for capturing the complete 3D anatomical structure. To address this limitation, we innovatively leverage point cloud completion to generate robust volumetric shape from sparse annotation, offering a promising solution to this challenge. In this paper, we propose a novel Geometry-Aware SAM adaptation framework (namely GA-SAM) that integrates point cloud shape generation module with cross-view segmentation supervision mechanism. Specifically, we train a point cloud completion network to infer the 3D structure of the target anatomy. The generated point cloud shapes are then used to produce pseudo-labels, guiding the adaptation of SAM via a geometry-aware shape constraints. Furthermore, we incorporate a cross-view supervision mechanism, leveraging multi-view consistency to ensure reliable segmentation across different planes. We demonstrate the effectiveness of our method on Pancreas-CT dataset, surpassing the state-of-the-art SAM adaptation method by a Dice score of 15.25% and significantly improving segmentation robustness. Our code is available at https://github.com/ShumengLI/GA-SAM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5332_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ShumengLI/GA-SAM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiShu_GASAM_MICCAI2025,
        author = { Li, Shumeng and Zhang, Jian and Qi, Lei and Shi, Yinghuan},
        title = { { GA-SAM: Geometry-Aware SAM Adaptation with Sparse Annotation-Driven Point Cloud Completion } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    1.) A point cloud completion netwotk is used to complete the partial point cloud generated from annotated three CT slices. 2.) SAM is applied to generate segmentation masks and supervised by the annotated three CT slices and completed shape (as pseudo mask).

  • 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 point cloud completion is a good way to generate pseudo mask by completing the partial point cloud produced from partially annotated data.

  • 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.

    Limitations: 1.) This method replies on the point cloud completion network to generate pseudo masks, while the network should be pre-trained on a shape completion dataset. Then, if we want to apply the method to a new organ or multi organs, we should collect corresponding shape data to train the completion network. This significantly limits the applicability of the proposed method. 2.) The completion also limit the shape of target organs. In other words, the organ should have simple and clear shape/boundary. If the shape is thin (eg., vessel) or complicated (eg., brain, or multiple organs), it would be difficult for the completion network to produce a good pseudo mask for the supervision during the adaptation.

    Weaknesses: 1.) Only experiments on pancreas dataset are conducted. Then, it would be unknown that how this method will perform on other shapes. This limits the contribution of this work. 2.) It is unfair to compare the methods that only use labeled data for training. In the setting of this work, both labeled and unlabeled slices are used for training, and labeling ratio is 3/128=2.23%. Hence, comparing with more semi-supervised methods (not just SAM-based) is necessary. For 2D methods, it is ok to use the annotation way mentioned in this work, ie., only labeling three slices for each CT. For 3D methods, they should annotate some of CT data such that the labeling ratio is close to 3/128. 3.) It seems that they did not mention which slice is labeled in each axis. This raises questions that a.) if the annotations are different for CT cases, how do they decide which slice should be annotated in each axis? And will this introduce more annotation time or require priors? b.) if the annotations are the same for different cases (for example, only the middle one, ie., 64th slice is annotated), how do they guarantee each slice contains the target shape? and if some slices do not contain the target shape, I do not think the completion network can perform well to generate high-quality pseudo masks.

  • 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.

    (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?

    Limitations of the completion network and concerns on the comparing methods.

  • Reviewer confidence

    Very confident (4)

  • [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.

    Concerns are addressed in the rebuttal.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a novel framework, GA-SAM, for adapting the Segment Anything Model (SAM) to medical image segmentation with sparse annotations. The authors use point cloud completion to generate robust volumetric shapes from a few annotated slices. Then they combine with a cross-view SAM adaptation module. The point cloud completion network is trained to infer the 3D structure from triple-slice annotation, and the generated shapes are used to produce pseudo-labels, guiding SAM adaptation with geometry-aware shape constraints. A cross-view supervision mechanism is incorporated to enhance multi-view consistency. Effectiveness of GA-SAM is demonstrated on the Pancreas-CT dataset.

  • 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 paper addresses a critical problem in medical image segmentation: how to effectively train models with limited and sparse annotations.

    2. The proposed framework is technically sound and well-motivated.

    3. The paper presents strong experimental results on a challenging dataset.

  • 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. A more detailed explanation of the point cloud completion network architecture and training process.

    2. The computational cost of the proposed framework, especially the point cloud completion step, is not clearly discussed.

    3. The evaluation is limited to one dataset.

  • 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
    1. Provide more details on the point cloud completion network.

    2. Discuss the computational complexity.

    3. Evaluate the method on a more diverse set of medical image datasets.

  • 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?

    Overall, the paper is well organized; method is sound and interesting. My major concerns: Computational cost.

  • 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
    1. The authors introduced the point cloud completion model to generate shapes for sparse annotation (three slices).
    2. They proposed a geometry-aware guidance strategy to leverage soft shape priors to boost shape awareness.
    3. They also introduced a cross-view supervision mechanism to support reliable and robust segmentation in 3D medical data based on multi-view consistency.
  • 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 motivation and novelty of this paper is good. The proposed method can well address the 3D medical segmentation task using sparse annotation and foundation model like SAM.
    2. This paper is well-written and easy-to-follow.
  • 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. Some descriptions and method details are missing. Details refer to Q10
  • 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
    1. The proposed method is highly dependent on shape priors. If the organs do not share similar anatomical shapes, this method may fail. How to handle these situations well? - -Besides, how can shape similarity be measured quantifiably?
    2. How to select the three slices for sparse annotation? Will the selected slice seriously affect the model’s performance?
    3. “For inference, we retain the segmentation model trained on transverse plane, which is often used in clinical diagnostics, and point completion network and segmentation models from other views are eventually discarded.”- -it is confusing for the inference stage: a) During testing, is only the transverse plane information required? Is the information from the other two planes (sagittal and coronal) not needed? b) Is the point completion network not required during testing? c) The detailed testing process/pipeline should be clearly provided for better understanding.
    4. SAM2 or MedSAM2 should be compared, since they share a similar problem setting (annotating one/limited frame for video/volume segmentation).
  • 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?

    Good motivation and novelty

  • Reviewer confidence

    Very confident (4)

  • [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 authors have adequately addressed most of my concerns. I support the acceptance of this paper.




Author Feedback

We appreciate the reviewers’ thoughtful feedback. We are encouraged that our work is considered “novel” (R3), “well-motivated” (R1,3), “easy to follow” (R3), with a “good method” (R2), and showing “strong results” (R1). We have well-documented code ready for release and will share it upon paper acceptance.

  1. Method details (R1): Thanks. Our point completion network adopts an encoder-decoder structure. The encoder has MLP and Transformer-based branches for global and local features, and a PointNet-style decoder outputs detailed point clouds. Training settings are described in Section 3.1 and we will add details in the final version.
  2. Computational cost (R1): Our framework introduces low overhead, as the point completion network (~83M params) is only used during training. It is trained for 50 epochs (~1.5h on a V100 GPU) and used once per volume (~1s/case) to generate pseudo-labels. For segmentation, only ~4M parameters in SAM (ViT-B, 91M) are tuned via LoRA. At inference, the model has the same size as other SAM-based methods.
  3. Dataset & Limitations (R1, R2): Our key contribution is a new paradigm for integrating 3D structural information into sparse-annotation segmentation, not an organ-specific pipeline. With MedShapeNet as a public shape source, the method can be extended to many organs (e.g., liver, spleen) without collecting extra data. Considering the application for other organs, first, our method has been validated on pancreas, which is a challenging organ with high variability [31]. The experiments on pancreas have shown the effectiveness of GA-SAM. Second, prior work [34] shows that the complex shape can also be generated with point completion, thus vessel segmentation may also be feasible. To this end, we also plan to develop a generalizable shape model across diverse anatomies.
  4. Comparisons (R2): Thanks. We compare not only with fully supervised but also semi-supervised methods, including 3D2DCT, SemiSAM, and CPC-SAM (Table1), using the same labeled/unlabeled slices as ours. And we also evaluated IC [32] and CML [33], which achieved 42.03% and 47.80% Dice—both lower than ours—highlighting the benefits of geometry-aware guidance.
  5. Slice Selection (R2, R3): The general principle is to select slices with visible targets [1,21], and we randomly sampling one from each axis. The model is trained to handle spatial variation by simulating diverse sparse slice inputs, which allow to generate reliably shapes.
  6. Shape similarity (R3): Thanks. Most organs exhibit statistically regular shapes across populations, but some diseases may alter morphology. To address this, we plan to compare the predicted mask with the learned prior and quantify shape similarity by metrics like Chamfer Distance. Samples with low similarity could be flagged for manual annotation to enhance robustness.
  7. Inference (R3): The point completion and cross-view supervision are only in the training phase. During inference, the input volume is sliced along the transverse plane (a clinically common view) and segmented using the fine-tuned SAM model trained on this plane. This keeps inference lightweight and efficient.
  8. SAM2 (R3): SAM2 and MedSAM2 enable zero-shot segmentation with a mask prompt, but their performance under sparse annotation is limited without adaptation. Also, [35] reports Dice scores below 30% for SAM2 on pancreas. We instead focus on SAM adaptation and semi-supervised comparisons. Thank you again for the constructive comments.

[31] Abdominal multi-organ segmentation with organ-attention networks and statistical fusion. MedIA. 2019. [32] Exploring inherent consistency for semi-supervised anatomical structure segmentation in medical imaging. IEEE TMI. 2024. [33] Cross-view mutual learning for semi-supervised medical image segmentation. ACM MM. 2024. [34] CarveNet: Carving Point-Block for Complex 3D Shape Completion. IEEE TMM. 2024. [35] A Short Review and Evaluation of SAM2’s Performance in 3D CT Image Segmentation. Arxiv. 2024.




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.

    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 #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’

    All reviewers recommend accepting this work due to its strong performance and well-motivated designs. Most concerns raised have been addressed in the rebuttal.



back to top