Abstract

Self-supervised DINO models provide strong transferable visual representations, yet applying them directly to image segmentation remains challenging. Existing approaches commonly rely on heavy decoders with complex upsampling, introducing substantial parameter and computational overhead. We observe that introducing scale into DINO features is far more critical than increasing decoder capacity. In this work, we present SegDINO, an efficient segmentation framework that integrates a DINOv3 backbone with lightweight scale modeling. SegDINO introduces Token Pyramid Adaptation (TPA) to reorganize intermediate DINO features into a pseudo multi-scale hierarchy, and Scale-Aware Decoding (SAD) for efficient intra-scale refinement and top-down multi-scale propagation. We further curate PanCT, a new CT dataset containing 284 patients with expert-annotated pancreatic tumors, to assess SegDINO’s ability to handle difficult small-lesion cases. Extensive experiments on PanCT and three public benchmarks demonstrate that SegDINO achieves state-of-the-art results with high efficiency. The code is available at https://github.com/script-Yang/segdino_v2.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0067_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/script-Yang/segdino_v2

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YanSic_SegDINO_MICCAI2026,
        author = { Yang, Sicheng AND Wang, Hongqiu AND Xing, Zhaohu AND Chen, Sixiang AND Yang, Qiuxia AND Mao, Yize AND Yang, Guang AND Zhu, Lei},
        title = { { SegDINO: Introducing Multi-Scale Structure into DINO for Efficient Medical Image Segmentation } },
        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’s main contribution is the design of SegDINO, an efficient medical image segmentation framework that adapts self-supervised DINO representations by explicitly introducing multi-scale structure. Specifically, it proposes Token Pyramid Adaptation (TPA) to convert single-scale DINO features into a pseudo multi-scale hierarchy, and Scale-Aware Decoding (SAD) to perform lightweight yet effective multi-scale feature refinement and fusion. Together, these components demonstrate that careful scale modeling, rather than complex decoders, is key to achieving strong segmentation performance, leading to state-of-the-art accuracy with improved computational efficiency across multiple medical imaging benchmarks, including a newly curated pancreatic tumor 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.

    a) The paper addresses a key limitation of lack of explicit multi-scale structure for dense prediction tasks. b) The proposed Token Pyramid Adaptation (TPA) is novel in that it reorganizes same-resolution transformer features into a pseudo multi-scale hierarchy, instead of relying on traditional feature pyramids from CNNs. c) It bridges the gap between self-supervised vision transformers and segmentation requirements without modifying the backbone. d) The Scale-Aware Decoding (SAD) module introduces a computationally efficient alternative to heavy decoders commonly used in segmentation. Authors have incorporated intra-scale refinement (local feature enhancement) and inter-scale propagation (top-down fusion). e) The method achieves state-of-the-art or superior results across multiple datasets while maintaining low parameter count

  • 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.

    a) The paper does not compare against several recent and highly relevant SOTA models, such as:

    • DINO-based segmentation variants (e. g. , DINO-UNet-style approaches)
    • SAM-based models like SAM2-UNet, which demonstrate strong performance using foundation model encoders
    • Mamba-based architectures such as Swin-UMamba Limited novelty in decoder design (SAD resembles existing FPN/U-Net decoders)

    b) The proposed Scale-Aware Decoding (SAD) follows a top-down multi-scale fusion strategy, which is conceptually similar to Feature Pyramid Network (FPN), thus leading leading to limited novelty. c) Insufficient justification and analysis of the refinement operator (R). No ablation or theoretical justification to demonstrate why this specific design is optimal

    d) The paper claims that multi-scale representation is the key contribution. However, they do not provide ablation study on layer selection ( 3, 6, 9, 12) and scale configuration (4, 8, 16, 32).

    e) The proposed methods uses of 2D slices from 3D medical data which can lead to loss of spatial features. The PanCT dataset consists of 3D CT volumes, but the method converts them into 2D axial slices.

    f) The paper does not include ethical and clinical compliance details for the newly curated PanCT dataset. Authors should add these details in the paper.

    g) The paper lacks the discussion on curation of the newly introduced PanCT dataset e. g. number of male vs female patients, age etc. Further, it is not mentioned about the availability of the dataset.

    h) The paper does not provide the code for the implementation.

  • 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

    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?

    The paper introduces SegDINO, a well-motivated framework that adapts self-supervised DINO features for medical image segmentation through explicit multi-scale modeling.

    However, the paper lacks comparisons with several recent SOTA methods (e.g., DINO, SAM, and Mamba-based approaches), and the novelty of SAD is somewhat limited due to its similarity to FPN/U-Net-style decoders. Key ablations (e.g., layer/scale choices) and justification of design components are missing. Additionally, concerns remain regarding the use of 2D slices from 3D data, limited dataset transparency, and lack of code release.

    Overall, despite these shortcomings, the core idea is sound, effective, and practically relevant, thus the recommendation is 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 SegDINO, an efficient segmentation framework that adapts DINO v3 backbone with lightweight scale modeling. The author release PanCT, a new CT dataset with 284 patients and expert-annotated pancreatic tumor masks for small-lesion segmentation evaluation.

  • 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 identifies that the key challenge in adapting DINO-based features for segmentation lies in designing effective scale modeling mechanisms rather than relying on heavy decoders. 2.Ablation studies validate the effectiveness and primary contribution of the proposed TPA component. 3.The framework achieves superior segmentation performance while maintaining high efficiency.

  • 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.While effective, the proposed TPA and SAD modules largely extend existing ideas in multi-scale feature construction and lightweight decoding. Similar multi-scale designs are widely used in segmentation, making the contribution somewhat incremental. 2.The PanCT dataset is originally 3D CT volumes but is converted into 2D slices for training. This discards important inter-slice spatial context, yet the paper does not justify this design choice or compare with 3D or 2.5D approaches (e.g., 3D U-Net or recent volumetric methods), which are standard in medical imaging. 3.In the Scale-Aware Decoding (SAD), the use of depthwise and pointwise convolutions is not well justified (Eq. 6). It is unclear why this specific design is preferred over other lightweight alternatives or how it impacts performance beyond efficiency. 4.The model extracts features from specific DINO layers (e.g., layers 3, 6, 9, 12), but the rationale for selecting these layers is not discussed. There is no ablation on layer selection or analysis of how different feature combinations affect performance.

  • 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 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?

    This paper proposes a simple yet efficient framework to adapt DINO v3 for medical image segmentation with solid empirical performance, particularly in leveraging DINO features with a lightweight design. However, the methodological novelty is somewhat limited, and several design choices are not sufficiently justified.

  • 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 introduces SegDINO, a segmentation framework based on lightweight variant of DINOv3 with scale-based improvements for medical imaging. It evaluates the method across three public datasets (TN3K, Kvasir, ISIC) and one private dataset (PanCT)

  • 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.
    • Achieves high scores across multiple datasets and demonstrates that lightweight DINOv3 variants can be highly effective
    • The ablation study successfully validates the necessity and impact of each proposed 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.
    • While the authors compare against various SOTA models (U-Net, TransUNet, U-KAN, etc.), they omit comparisons with other medical-specific DINO-based foundation models. Furthermore, some compared methods are outdated
    • On the TN3K dataset, vanilla DINOv3 outperforms other comparisons, and the proposed improvements offer a marginal gain of less than 1%
    • The authors claim the method improves performance on small objects; however, three of the four datasets used do not primarily feature small objects, making it difficult to verify if the gains come from the proposal or the base DINO architecture
  • 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?

    If the authors can show that their SegDINO is better (in terms of performance or efficiency) than other DINO adaptations, then the decision will be a stronger Accept

  • 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



Author Feedback

We thank the reviewers for their constructive comments and for recognizing the effectiveness and practical relevance of SegDINO. We appreciate the suggestions on baseline comparison, module design, PanCT details, and reproducibility. In the revised manuscript, we will address these concerns accordingly. Below, we provide our point-by-point responses.

(1) Comparison with recent SOTA methods. Thanks for the valuable suggestions. We agree that more recent baselines would strengthen the evaluation. In the revised manuscript, we will add representative segmentation methods, especially DINO-based segmentation baselines, together with the corresponding references.

(2) Module design and ablation. Thanks. We agree that SAD is related to FPN-style designs and will add the corresponding discussion and references. However, SegDINO is not a generic FPN decoder; it adapts DINOv3 token features, which lack an explicit multi-scale hierarchy, for dense medical image segmentation. TPA builds pseudo multi-scale DINO features, while SAD performs lightweight refinement and fusion. Thus, the contribution lies in DINO-specific scale adaptation. We will also add layer-selection ablations.

(3) PanCT dataset details. Thanks. We will add more PanCT details. SegDINO focuses on adapting 2D DINOv3 representations for medical image segmentation, which enables fair comparison with other 2D segmentation methods. We acknowledge the importance of volumetric context and will discuss this limitation. Extending SegDINO to volumetric DINO-style segmentation is an important future direction.

(4) Reproducibility. We will release the implementation code and configuration files upon acceptance.




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.

    All three reviewers agree that SegDINO presents a well-motivated framework for adapting DINO features for medical image segmentation through lightweight multi-scale modeling, achieving strong performance across multiple benchmarks. The main concerns include: limited methodological novelty of TPA/SAD relative to existing FPN/U-Net decoders, missing comparisons with recent DINO/SAM/Mamba-based methods, lack of ablation on layer selection and scale configuration, unjustified use of 2D slices from 3D volumes, and missing ethical clearance details for the PanCT dataset. The authors are encouraged to address these points in the camera-ready version.



back to top