List of Papers Browse by Subject Areas Author List
Abstract
The emergence of large-scale pre-trained visual foundation models offers a promising paradigm for medical image segmentation. However, effectively transferring their rich, generic representations to domain-specific clinical tasks remains a significant challenge. This paper presents Dino U-Net, a novel encoder-decoder framework designed to fully leverage the high-fidelity dense features from the DINOv3 foundation model for medical image segmentation. We employ a frozen DINOv3 backbone as the encoder and introduce a dual-branch DINO Adapter to bridge the feature domain gap. To mitigate the loss of fine-grained information during feature dimensionality reduction, we further propose a Fidelity-Aware Projection Module (FAPM). This module utilizes low-rank shared projection coupled with dynamic feature modulation to refine and faithfully propagate features to the decoder. Extensive experiments on seven public medical image datasets demonstrate that Dino U-Net achieves state-of-the-art performance across diverse imaging modalities, excelling in both regional segmentation accuracy and boundary delineation. Moreover, segmentation performance shows consistent improvement as the backbone scales up to 7 billion parameters, confirming the framework’s strong scalability. The code is available at https://github.com/yifangao112/DinoUNet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0995_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yifangao112/DinoUNet
Link to the Dataset(s)
See paper
BibTex
@InProceedings{LiHao_Dino_MICCAI2026,
author = { Li, Haoyue AND Gao, Yifan AND Yuan, Feng AND Wang, Xiaosong AND Cui, Jinjiang AND Gao, Xin},
title = { { Dino U-Net: Exploiting High-Fidelity Dense Features from Foundation Models for Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, the authors present a hybrid medical image segmentation framework (Dino U-Net) based on foundation model Dino-v3.They employ a dual-branch DINO Adapter and a Fidelity-Aware Projection Module (FAPM) to leverage the high-fidelity dense features of foundation model and mitigate the loss of fine-grained information. In the experiment sessions, Dino U-Net is evaluated on seven public dataset across different imaging modalities and demonstrates consisitent improvements.
- 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 an important weakness of existing segmentation foundation models (e. g. , Segment Anything Model), namely their heavy reliance on dense mask supervision, which biases them toward geometric contours and limits their ability to capture nuanced semantic relationships and ambiguous anatomical boundaries. Building on this insight, the authors introduce a Fidelity-Aware Projection Module (FAPM) grounded in the strong semantic representation capability of DINO. The proposed module effectively reduces feature dimensionality while preserving semantic fidelity, offering a theoretically sound and well-justified solution to the geometry–semantics trade-off. This design is both conceptually clear and aligned with recent advances in self-supervised representation learning. 2.The framework is evaluated across a diverse set of imaging modalities and clinical tasks, including endoscopy, ultrasound, MRI, fundus imaging, microscopy, and cardiac MR (CMR), demonstrating broad applicability in real-world medical scenarios. The inclusion of seven competitive baselines further strengthens the empirical rigor. Importantly, the authors validate the method across foundation models of varying scales (from S to 7B), consistently showing performance gains. This comprehensive evaluation convincingly demonstrates both the robustness and scalability of the proposed framework, highlighting its potential as a general-purpose solution for medical vision tasks.
- 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 does not clearly describe the structure of the DINO Adapter, including how it interacts with frozen DINOv3 features, how the dual branches are fused, and why the dual-branch design can bridge the feature domain gap. In addition, although the advantages of DINOv3 are emphasized, there is no comparison with other backbones (e. g. , SAM or other representation models). This makes it unclear whether the performance gains come from the proposed framework, from the backbone choice or just from the FAPM modules. 2.While components like Prefine and Pshortcut are partially explained, key parameters such as the scaling factor αand shifting factor β are not well motivated. Their roles in improving feature quality or preserving semantic information are unclear. Moreover, the module is relatively complex, but lacks sufficient ablation studies to justify the necessity of each component. 3.The method shows relatively small improvements on most datasets, but significantly larger gains on the BUSI dataset (~5%). The paper does not explain why BUSI benefits more. For example, whether it is due to dataset characteristics or better alignment with DINO features. This missing analysis weakens the overall claim of generalizability.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 presents a well-motivated design of the FAPM module, supported by targeted ablation studies and well-organized figures and tables. The experimental evaluation is comprehensive, covering seven public datasets, multiple imaging modalities, four model scales (from S to 7B), and seven baselines. These results convincingly demonstrate the robustness and generalizability of the framework. A notable strength is the consistent performance improvement as the backbone scales up to large models, providing useful evidence for scalability in medical imaging. However, the paper lacks sufficient justification and experimental validation regarding the backbone choice. In particular, there is no comparison with alternative architectures, making it unclear whether the performance gains are due to the backbone or the FAPM module. This limitation is the main reason for the weak recommendation.
- 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 work propose Dino U-Net to leverage a frozen DINO v3 foundation model for medical image segmentation. Dino U-Net introduce a dual-branch Adapter to bridge the domain gap and a Fidelity-Aware Projection Module to preserve fine-grained feature information.
- 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.This paper provides empirical evidence supporting the effectiveness of applying DINO v3 to medical image segmentation. 2.The proposed method is evaluated on seven medical image segmentation datasets, which provides a relatively comprehensive assessment of its effectiveness. 3.The paper provides qualitative visualizations of segmentation results.
- 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 proposed framework largely follows existing paradigms of integrating foundation models with U-Net-style decoders. The main contributions (DINO Adapter and FAPM) appear as incremental extensions rather than fundamentally new formulations. 2.The performance gains are likely dominated by the strong DINOv3 backbone rather than the proposed modules. Although ablations are provided (Table 4, page 8), the improvements from FAPM are relatively modest (e.g., <1% Dice), making it unclear how much the method itself contributes beyond the pretrained features. 3.Key components such as the dual-branch adapter (spatial vs. semantic branch) and the modulation mechanism in FAPM (α, β scaling) are intuitively motivated but lack deeper theoretical or empirical justification. It is unclear why this specific design is optimal compared to other feature adaptation strategies (e.g., attention-based fusion or standard projection layers).
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(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 shows strong empirical results, but the novelty is limited and largely builds on existing U-Net and adapter-based designs. The improvements seem to rely heavily on the pretrained backbone, with insufficient justification of key design choices.
- 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
Review #3
- Please describe the contribution of the paper
The paper introduces Dino U-Net, a hybrid architecture that utilizes a frozen DINOv3 backbone for encoding, a dual-branch DINO Adapter, and a Fidelity-Aware Projection Module (FAPM) for feature refinement. The framework aims to leverage high-fidelity dense features from foundation models for more precise medical image segmentation. The paper demonstrates state-of-the-art performance across multiple datasets and shows strong scalability as the backbone model size increases.
- 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.Broad evaluation: The paper evaluates Dino U-Net on seven public medical image datasets, spanning diverse modalities (e.g., MRI, ultrasound, endoscopy), and compares it against seven strong baselines. The large-scale evaluation and the use of statistical testing make the results robust. 2.Well-motivated and validated FAPM: The FAPM module is a key contribution, enhancing feature fidelity by using low-rank projections and dynamic affine modulation. The ablation study demonstrates that FAPM is crucial for achieving high-quality segmentation. 3.Scalability: The results demonstrate clear improvements as the model scales from the smallest (S) to the largest (7B) variant, with the largest model achieving the best performance across five of seven datasets. 4.Reproducibility baseline: The authors provide detailed implementation details, including preprocessing, augmentation strategies, and model training procedures.
- 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 evaluates FAPM in an ablation study, but the DINO Adapter, which handles feature adaptation via cross-attention, is not separately ablated. It’s unclear whether a simpler cross-attention design or a single-branch model would achieve similar results.
2.The 7B model’s inference cost, including latency, GPU memory, and FLOPs, is not reported. These metrics are essential for evaluating the practical deployability of the model, especially in clinical settings.
3.The internal architecture of the DINO Adapter (e.g., number of stages, resolution handling, deformation kernel parameters, computational footprint) is not described in enough detail for full reproducibility. These details are necessary for understanding the true impact of the DINO Adapter on model performance.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.How did you ensure patient-wise or video-wise separation in datasets where samples are correlated (e.g., m2caiSeg video frames, MRI slices from the same subject)?
2.For multi-class datasets (e.g., m2caiSeg), how are metrics aggregated (macro vs micro Dice)?
- 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 presents a novel approach to medical image segmentation using a foundation model, and the empirical results are promising. However, the concerns mentioned above need to be addressed before I can recommend acceptance. If these issues are adequately clarified in the rebuttal, I would consider increasing my recommendation.
- 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 #4
- Please describe the contribution of the paper
This paper proposes Dino U-Net, a U-Net–style segmentation framework that employs a frozen DINOv3 foundation model as the encoder, a dual-branch DINO Adapter and a Fidelity-Aware Projection Module (FAPM) to reduce channel dimensionality while preserving fine-grained details. Evaluated on seven public datasets spanning multiple modalities, the approach reports improvements over several baselines, shows average performance increases with larger DINOv3 backbones.
- 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.Integrates a self-supervised foundation encoder (DINOv3) into a U-Net architecture. Presents a scaling study across multiple backbone sizes, highlighting the potential benefits of large frozen encoders for medical segmentation. 2.Evaluates on seven datasets across diverse modalities and anatomical targets, with two metrics (Dice, HD95) and Wilcoxon signed-rank tests for significance. 3.Releases anonymized code, which is helpful for reproducibility.
- 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.Important implementation details are missing (input resolution, batch size, DINOv3 feature map resolutions/channels per stage, adapter dimensionalities). In addition, what is the implementation details for baselines? e.g. for SAM-based models, are they trained on your datasets or it is the results of zero-shot? What are prompts for SAM-based models?
2.No out-of-distribution/generalization tests are provided (e.g., train on one polyp dataset, test on another).
3.SOTA baselines comparison is missing. e.g. For polyp segmentation, comparisons omit several strong specialized methods (e.g., Polyp-SAM), which likely outperform generic baselines on Kvasir-SEG.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.What are FLOPs for S/B/L/7B variants comparing with baselines?
- 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 manuscript delivers an integration of DINOv3 dense self-supervised features into a U-Net framework, with consistent average gains over baselines across seven datasets. Overall, I view the paper as a solid step toward leveraging DINOv3 for medical segmentation with promising empirical results. With clarifications and additional experiments, this work would be strong enough for acceptance. As it stands, I lean toward a 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
Author Feedback
We sincerely thank the reviewers and meta-reviewer for their constructive comments and positive evaluation of our work. We will carefully revise the manuscript for clarity in the camera-ready version, including improving the description of the proposed Dino U-Net framework, clarifying experimental settings, and refining the presentation of results. We also appreciate the suggestions regarding future extensions and will discuss these points in the final version where appropriate.
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 Dino U-Net, a hybrid segmentation framework that integrates a frozen DINOv3 foundation model with a dual-branch Adapter and a Fidelity-Aware Projection Module (FAPM) to bridge the domain gap and preserve fine-grained semantic features. The primary strengths lie in the extensive empirical evaluation across seven diverse medical imaging datasets and the demonstrated scalability of the performance gains as the backbone size increases up to 7B parameters. However, reviewers noted concerns regarding the incremental nature of the methodological novelty, the lack of sufficient ablation studies for individual adapter components, and missing technical details regarding inference costs and baseline implementations. Despite these critiques, the consistent performance improvements across multiple modalities provide strong evidence for the framework’s utility.