List of Papers Browse by Subject Areas Author List
Abstract
Accurate medical image segmentation requires simultaneous modeling of robust global semantics and precise anatomical boundaries. While Vision Transformers (ViTs) exhibit strong global representation capability, they often struggle to capture fine-grained local structures and boundary details, particularly in scenarios with low contrast and blurred boundaries. To address this limitation, we propose Glasses, a novel framework that equips pre-trained Transformers with a structure-enhanced visual adaptation mechanism for medical image segmentation. Specifically, Glasses employs a frozen DINOv3 as the backbone encoder, with only the terminal layers fine-tuned. We introduce a Multi-Scale Deformable Attention Adapter that performs cross-scale interaction between lightweight convolutional structural priors and Transformer tokens extracted from specific intermediate layers of DINOv3, thereby transforming single-scale token embeddings into hierarchical representations enriched with both local structural cues and global context. This design preserves the stability of pre-trained semantic representations while improving the modeling of local anatomical structures and scale variations. Furthermore, we introduce an independent Edge-Aware Convolutional Encoder that explicitly learns boundary-enriched multi-scale representations under nested boundary supervision, providing a strong source of fine-grained local structural information that complements the adapter-enhanced Transformer features during decoding. A U-Net-style decoder then progressively fuses the adapter-enhanced Transformer features and the boundary-enriched convolutional representations to recover high-resolution segmentation outputs. Extensive experiments on the Synapse and ISIC datasets demonstrate that Glasses achieves Dice coefficients of 83.52% and 91.60%, respectively, significantly outperforming state-of-the-art (SOTA) methods. Source code is available at https://github.com/friay/Glasses.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1781_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/friay/Glasses
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhuJia_Glasses_MICCAI2026,
author = { Zhu, Jiahao AND Cui, Hengfei AND Yin, Hanlin AND Chen, Geng AND Xia, Yong},
title = { { Glasses: Adapter-Enhanced DINOv3 with Boundary-Aware Multi-Scale Representations for 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
This work introduces a multi-scale deformable attention adapter to complement convolutional structural features with DINOv3 features. Additionally, it presents a boundary-enriched convolutional encoder to capture boundary information, with the goal of providing fine-grained local structural cues for medical image segmentation.
- 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 main strengths of this study are as follows:(1) It proposes a multi-scale deformable attention adapter designed for fine-tuning DINOv3 features. (2) It adopts a UNet-like architecture with two encoders and one decoder to jointly learn and fuse global and local features.
- 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 key weaknesses of this paper are as follows:(1) The performance improvements over baseline methods are marginal. (2) Validating the method only on two public datasets is insufficient, especially since these datasets focus on relatively straightforward segmentation tasks (abdominal organs and skin lesions). (3) The input–output flow in Figure 2 is unclear and confusing. (4) The overall novelty and technical insight of the proposed method are limited. (5) The framework name “Glasses” bears no clear connection to the proposed methodology. (6) No qualitative visualizations are provided to illustrate how the proposed adapter functions or improves structural feature representation.
- 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?
My overall score is primarily based on the following factors: (1) The novelty of the proposed approach is not sufficiently strong for publication. (2) The experimental performance gains over competing baselines are minimal and not compelling.
- 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 proposed model introduces a dual-encoder architecture for medical image segmentation that combines a largely frozen DINOv3 backbone with structural representation to balance global semantics and local detail. A Multi-Scale Deformable Attention Adapter is introduced to perform cross-scale interaction between lightweight convolutional structural priors and Transformer tokens extracted from intermediate layers of DINOv3, while a parallel Edge-Aware Convolutional Encoder provides explicit boundary-enriched features through nested supervision. The model achieves SOTA performance on the Synapse and ISIC 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.
The model combines the robust global context of a largely frozen DINOv3 backbone with a dedicated convolutional stream to capture both high-level meaning and fine-grained anatomical details. The Multi-Scale Deformable Attention Adapter allows the model to learn medical-specific spatial features without retraining the entire Transformer, preserving powerful pre-trained representations while reducing computational overhead. By using an Edge-Aware Convolutional Encoder with nested supervision, the framework explicitly prioritizes contour accuracy to improve the model’s performance.
- 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 freezing DINOv3 preserves semantic stability, it may prevent the model from fully adapting its core feature extraction ability to the unique stylistic and domain-specific characteristics of medical images.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 method and the presentation clarity are the factors that lead to my overall score for this paper.
- 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
1.This paper propose a structure-enhanced adaptation framework built upon a largely frozen DINOv3 backbone. This design preserves the robust global semantic representation capabilities of the pre-trained model while preventing overfitting on limited medical datasets. 2.A novel adapter module is designed to perform cross-scale interactions between lightweight convolutional structural priors and intermediate Transformer tokens of DINOv3.This mechanism effectively transforms single-scale token embeddings into structure-aware, hierarchical multi-scale representations without modifying the frozen backbone. 3.The framework introduces an independent parallel convolutional encoder that explicitly learns boundary-enriched multi-scale representations under nested boundary supervision. This branch provides high-resolution, complementary structural information to explicitly guide the decoding stage.
- 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 precisely targets the current pain point of Vision Transformers (ViTs) in medical image segmentation: strong global semantics but weak local boundaries. The logic of using a parallel adapter for aligning semantics to structures and an independent edge encoder for explicit structural modeling is very clear, effectively forming a complementary dual-stream architecture. 2.Utilizing Deformable Attention for the sampling and residual updating of frozen features is an elegant approach to feature interaction. Furthermore, the nested boundary supervision mechanism projects features of different depths into a unified boundary space via progressive alignment operators. This design is theoretically rigorous and effectively avoids feature conflicts during multi-scale prediction. 3.Glasses achieves SOTA performance on two classic and challenging public datasets, Synapse and ISIC 2018.Particularly on the Synapse dataset, it not only reaches a mean Dice coefficient of 83.52%, but more importantly, the Hausdorff Distance (HD) is significantly reduced to 12.28 mm, directly demonstrating the model’s superiority in precise boundary prediction. 4.The ablation experiments detailed in Table 3 solidly prove the independent value and necessity of core components, including the Multi-Scale Deformable Adapter, the parallel CNN branch, and the nested boundary supervision strategy.
- 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.This paper lacks an analysis discussing the model’s sensitivity to hyperparameters. For example, the weights in the semantic loss function, alpha and beta. 2.While the proposed framework leverages DINOv3 as its backbone , the comparative experiments fail to include other state-of-the-art segmentation methods that also utilize the pre-trained DINOv3 model. 3.The experimental section lacks comparisons with methodologies that employ related pre-trained vision models (e.g., other ViT variants, SAM, or equivalent foundation models) coupled with fine-tuning strategies. 4.Although the proposed Glasses framework achieves state-of-the-art overall mean Dice and Hausdorff Distance on the Synapse dataset , the quantitative results in Table 1 reveal that it does not attain the highest scores across all individual organ categories.The manuscript currently lacks a critical discussion explaining this phenomenon.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 tackles a highly relevant and challenging problem in medical image segmentation by effectively adapting a frozen DINOv3 foundation model. My recommendation leans positive primarily due to the architectural novelty and the strong empirical validation. The proposed Glasses framework elegantly combines a Multi-Scale Deformable Attention Adapter for semantic-to-structural alignment with an Edge-Aware Convolutional Encoder for explicit, high-resolution boundary modeling. The results on the Synapse and ISIC 2018 benchmarks are compelling, achieving a notable 83.52% mean Dice and 12.28 mm HD on Synapse.
- 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 all reviewers for their detailed and constructive feedback. Here are responses to their suggestions and concerns.
R1Q1: Improvements over Baselines Thank you for your comment. While the Dice improvements over strong baselines are moderate, Glasses consistently improves both segmentation accuracy and boundary quality across two different medical segmentation tasks. On Synapse, it achieves the lowest HD (12.28 mm), reflecting improved boundary localization and anatomical continuity. On ISIC 2018, it also shows competitive performances.
R1Q2: Validation on Two Datasets Synapse and ISIC 2018 cover distinct medical segmentation scenarios with different modalities, anatomical structures, and boundary characteristics. The consistent improvements across both datasets suggest that the framework generalizes beyond dataset-specific settings.
R1Q3 & Q6: Clarity of Figure 2 and Adapter Visualization Thank you for pointing this out. In Fig. 2, lightweight convolutional structural priors serve as queries, while DINOv3 tokens provide semantic key-value representations for cross-scale interaction. Through deformable sampling and residual iterative updating, the adapter transforms single-scale tokens into structure-aware multi-scale representations. We also agree that qualitative visualization can help illustrate the adapter’s effect. In the final version, we will improve the feature-flow illustration and corresponding description.
R1Q4: Novelty and Technical Insight Glasses is not a simple CNN-Transformer fusion. Its novelty lies in structure-aware adaptation for largely frozen DINOv3 and explicit boundary-aware structural modeling. The Multi-Scale Deformable Attention Adapter performs deformable sampling and residual semantic updating between lightweight structural priors and DINOv3 tokens, transforming single-scale tokens into hierarchical structure-aware representations without modifying the frozen backbone. In addition, the nested boundary supervision projects multi-scale structural features from the parallel convolutional encoder into a unified boundary representation space, improving cross-scale edge consistency.
R1Q5: Framework Name “Glasses” is intended as a metaphor for enhanced structural perception and boundary awareness in medical image segmentation. We will clarify this in the final version.
R2Q1: Frozen DINOv3 Adaptation In our framework, we perform selective fine-tuning only on the higher-level DINOv3 layers, complemented by the proposed structure-aware adapter, which interacts with lightweight structural priors to progressively adapt the pre-trained DINOv3 features to medical segmentation tasks. This allows the backbone to retain robust global semantics while becoming task-adaptive, improving structural modeling on limited datasets.
R3Q1: Hyperparameter Sensitivity Thank you for your valuable suggestion. The weighting parameters balance the contributions of cross-entropy and Dice terms. Due to space limitations, the ablation study on the weighting parameters was not included. A more comprehensive sensitivity study remains future work.
R3Q2 & Q3: Comparison with DINOv3-/Foundation-Model-Based Methods Our primary goal is to evaluate whether the proposed strategy can enhance largely frozen foundation Transformer representations for medical image segmentation. We agree that broader comparisons with other DINOv3-based or foundation-model-based segmentation methods could provide further insights, and we plan to investigate in future work.
R3Q4: Organ-Wise Performance Differences Glasses does not achieve the highest Dice score for every individual organ category. This is largely due to the inherent characteristics of the Synapse dataset, where annotations and contrast vary. Our framework is designed to enhance overall structural consistency and boundary accuracy. Even when the per-organ Dice is not the highest, Glasses improves the overall mean Dice and significantly reduces Hausdorff Distance.
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 presents Glasses, a structure-enhanced DINOv3 adaptation framework for medical image segmentation. Overall, the work addresses an important and practical problem in medical image segmentation: how to leverage strong pre-trained semantic representations while complementing them with local structural and boundary-aware modeling. All three reviewers considered the topic relevant, the methodological design reasonably clear, and the experimental results on public benchmarks such as Synapse and ISIC competitive.
The proposed structure-enhanced DINOv3 adaptation framework is well aligned with the medical image segmentation task, and the overall empirical results are competitive. The main concerns raised by the reviewers appear to be addressable in the final version through additional discussion, clearer illustration of the framework, more qualitative analysis, and expanded experimental interpretation. Therefore, this paper is recommended for early acceptance.