List of Papers Browse by Subject Areas Author List
Abstract
Although DINOv3 has demonstrated remarkable semantic discrimination in natural imagery, its direct application to volumetric medical segmentation is hindered by inherent dimension and domain disparities. To resolve these issues, we propose DINO-Med3D, a two-stage progressive framework that repurpose the pre-trained DINOv3 encoder for 3D medical tasks. In the first stage, we mitigate the dimension gap by introducing a multi-slice embedding module that incorporates pseudo-3D context, while simultaneously employing a segmentation proxy task to adapt representations learned from natural scenes to the medical domain. Subsequently, we further enhance volumetric understanding by adding lightweight 3D adapters into the frozen backbone to enforce global inter-slice continuity. Finally, to compensate for the spatial information loss inherent in the embedding process, we design a parallel detail recovery stream to explicitly preserve high-frequency boundary cues. Extensive experiments on five public datasets demonstrate that our approach successfully adapts DINOv3 to the medical domain and significantly outperforms state-of-the-art baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3353_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HaoyuHu1/DINO-Med3D/
Link to the Dataset(s)
AISD: https://github.com/GriffinLiang/AISD
MSD-Pancreas: http://medicaldecathlon.com/
MSD-Colon: http://medicaldecathlon.com/
BraTS 2020: https://www.med.upenn.edu/cbica/brats2020/data.html
ACDC: https://www.creatis.insa-lyon.fr/Challenge/acdc/databases.html
BibTex
@InProceedings{HuHao_DINOMed3D_MICCAI2026,
author = { Hu, Haoyu AND Ma, Xiyao AND Liu, Shiqi AND Zhang, Linsen AND Xie, Xiaoliang AND Zhou, Xiaohu AND Hou, Zeng-Guang},
title = { { DINO-Med3D: Bridging Dimension and Domain Gaps in Volumetric Segmentation via Progressive Adaptation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes DINO-Med3D, a two-stage progressive framework that applies 2D DINOv3 to 3D medical segmentation by progressively bridging dimensional and domain gaps. The authors conducted experiments on five public datasets spanning multiple modalities, anatomical structures, and pathological textures to verify the effectiveness of the proposed method in 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 authors achieve the alignment of the 2D DINOv3 backbone with volumetric medical data via the Inter-slice Context Embedding (ICE) module and a pseudo-3D proxy task. Furthermore, in the second stage, a dual-stream architecture is adopted to combine semantic abstraction with spatial precision.
- 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.In Stage I, it is not clearly stated which part of the network corresponds to the Proxy Task: Domain Adaption. The description in the “methods” section is not clear enough. 2.It is hoped that sub-modules such as HRSE and ICE can be illustrated with diagrams. 3.In the experimental results section, the absolute values of the evaluation metrics for the proposed method and the compared algorithms are lower than expected. Higher performance has been reported in the original paper that introduced the AISD dataset. The authors are encouraged to provide an explanation for the results.
- 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.
(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 author improves the original DINOv3 model and adapt it to 3D medical images, and propose a two-stage strategy to recover fine-grained details, which shows certain methodological novelty and practical value in the target application domain. However, the methodological descriptions are not sufficiently detailed, and some key designs and rationales lack in-depth elaboration. And the experimental results and analyses need to be further supplemented and improved to better validate the effectiveness and superiority of the proposed method.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The author’s responses to the first two questions are reasonable. However, the explanation for the question that the absolute values of the evaluation metrics for the proposed method and the compared algorithms are lower than expected is still not sufficiently persuasive. This flaw undermines the credibility of the research.
Review #2
- Please describe the contribution of the paper
The paper proposes a two-stage framework to adapt DINOv3 for 3D medical image segmentation by addressing both domain and dimensional gaps. The framework introduces a multi-slice embedding module (ICE), a volumetric context stream with lightweight 3D adapters, and a parallel detail-recovery stream. Experiments on five public CT and MRI datasets show consistent improvements over several CNN-, Transformer-, Mamba-, and DINO-based baselines.
- 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 presents a technically feasible and empirically competitive solution to the problem of adapting a 2D foundation model to 3D medical segmentation. 2.The experimental validation is fairly solid, with evaluations conducted across multiple CT and MRI datasets, anatomical structures, and segmentation scenarios. 3.The ablation studies are relatively systematic and help validate the role of the major components.
- 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.Since Stage I already fully fine-tunes the DINOv3 backbone on the target medical task, it is unclear why Stage II then freezes this adapted backbone and switches to LoRA-based updates. 2.The role of the 3D adapter remains unclear. Based on the current description, it appears to simply upsample a coarse stacked feature tensor derived from the DINO outputs, rather than construct genuinely new hierarchical features. This makes the claims of producing “hierarchical features” or “genuine volumetric representation” seem overstated. 3.Although the paper motivates the method in terms of annotation scarcity, domain gap, and transferability, the empirical evidence remains largely limited to in-domain random splits. It therefore remains unclear whether the method truly improves broader transfer/generalization beyond standard supervised benchmark settings. 4.The same K may correspond to very different anatomical context across datasets, yet this issue is not explicitly discussed. This is particularly concerning given the sharp drop on MSD-Colon, where increasing K from 3 to 5 reduces DSC from 53.28% to 33.10%. 5.Several symbols introduced in Section 2 do not map clearly to the components shown in Fig. 1, making the methodology difficult to follow visually.
- 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.
(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?
While this work tackles an important problem and shows competitive results, its methodological novelty is somewhat diffuse and does not yet provide sufficiently strong novel insights. Several core design choices are not sufficiently justified. In addition, although the method is motivated by annotation scarcity, domain gap, and transferability, the empirical evidence remains mainly limited to in-domain random splits. Consequently, the extent to which the method improves robustness beyond conventional in-domain supervised benchmarks is still not fully established.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 rebuttal has addressed my main clarification concerns, I recommend acceptance.
Review #3
- Please describe the contribution of the paper
This paper proposes DINO-Med3D, a progressive adaptation framework that repurposes the 2D DINOv3 encoder for volumetric medical image segmentation. The paper explicitly formulates the challenge as bridging both a dimension gap and a domain gap. To address this, the authors design a two-stage pipeline: Stage I uses an inter-slice context embedding (ICE) module and a proxy segmentation task to adapt DINOv3 to medical volumetric data, while Stage II freezes the aligned backbone and introduces a dual-stream architecture with a semantic volumetric context stream and a detail-recovery stream for dense prediction. Experiments on five public datasets show consistent improvements over several strong baselines.
- 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 clearly identifies the challenge of adapting a strong 2D foundation model to 3D medical segmentation, and frames it in terms of both dimension and domain gaps. 2.A key strength is that the paper recognizes a common issue of DINO-style features for dense prediction: they are semantically strong but often weak in fine-grained spatial precision. The proposed semantic/detail dual-stream design is therefore well aligned with the backbone characteristics. 3.The two-stage design is intuitive and sensible. The ablation study supports that Stage I adaptation and Stage II volumetric/detail refinement both contribute meaningfully. 4.The method is evaluated on five public datasets across CT and MRI, with comparisons against several competitive baselines. The reported gains are consistent and make the empirical results convincing overall.
- 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 main contribution is the integration and adaptation around DINOv3, making the paper largely an incremental work.
2.The claim of medical-domain alignment in Stage I is not fully verified at the representation level. The paper argues that Stage I aligns DINOv3 features to the medical domain, but the current evidence is mostly indirect and limited to downstream segmentation gains. More representation-level analysis, such as feature visualization, linear probing, or attention inspection, would strengthen this central claim.
3.The explanation for why K=3 works best remains largely empirical. The ablation on ICE input slices is useful, but the reason why performance peaks at K=3 is not well explained. Additional feature-level or attention-based analysis would help clarify whether ICE truly captures meaningful inter-slice context and why larger K degrades performance.
4.Inference efficiency is a practical limitation. It would be helpful to provide runtime or memory analysis.
- 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 paper addresses an important problem and proposes a coherent two-stage framework for adapting a 2D foundation model to volumetric medical segmentation. The method is well motivated, the semantic/detail dual-stream design is sensible, and the empirical evaluation across five datasets is strong enough to support the practical value of the approach.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
N/A
Author Feedback
We thank the reviewers for their constructive comments and for recognizing the importance (R2, R3), novelty (R1), practical value (R1, R3), experimental design (R3), and results (R2, R3) of our work. We address the main concerns below.
Stage I proxy task & domain adaptation (R1). Stage I does not contain a standalone module for domain adaptation. Instead, it is achieved by training the proxy task on medical data with full backbone fine-tuning: taking multiple adjacent slices as the input and outputting the dense segmentation of the central one, thereby aligning natural-image features with medical ones.
Details of the HRSE/ICE modules (R1). As adding new figures is prohibited, we clarify their structures here. ICE replaces DINOv3’s 2D patch embedding using a C×K×16×16 3D convolution, mapping K×H×W inputs to a C×H/16×W/16 pseudo-3D embedding. HRSE cascades three 3D convolution layers (residual 3DConv-ReLU-Norm), outputting a multi-scale hierarchy {C1, C2, C3} for the decoder or AGF (Fig.1, top-right).
Lower AISD score than the original paper (R1). This stems from our new data split. The original split yielded higher test scores than validation ones, implying distributional imbalance. Thus, we adopt random split and apply it to all baselines. It ensures consistent validation/test behavior and fair comparisons despite lower absolute scores.
Stage II LoRA & Backbone Freezing (R2). Stage I performs initial dimensional and domain alignment, taking multiple slices as input and performing 2D segmentation. Stage II turns to construct a volumetric segmentation model, of which the input and output are both 3D. The tasks are different, thus requiring further fine-tuning in Stage II. Full fine-tuning in Stage II risks overwriting the representation learned in Stage I, whereas frozen-backbone-with-LoRA preserves it while allowing 3D adaptation. It is also practically necessary, as full fine-tuning with 16-slice inputs is memory-prohibitive.
Role of the 3D adapter (R2). In Stage II, features from ICE+DINOv3 are first restacked along z-axis, then obtain inter-slice dependencies and restore resolution through the 3D adapter (with Dx256xHixWi deconv kernels). Thus, the 3D adapter is not simply for per-slice 2D upsampling. “Hierarchical features” denotes multi-scale features obtained from DINOv3 ViT layers. “Genuine volumetric representation” denotes outputting features as integral 3D entities, distinguishing from Stage I. For accuracy, we will revise the term to “volumetrically coupled representation”.
Sensitivity to K (R2, R3). K defines the adjacent-slice context for central slice prediction. Gains from K=1 to K=3 reflect anatomical continuity, while larger K may introduce off-target anatomy. MSD-Colon is highly sensitive due to a lower adjacent-slice foreground IoU (0.578 vs. 0.701/0.705 for BraTS/Panc.) and fewer slices per lesion(10 vs. 17/17). Thus larger K more often includes irrelevant background or heterogeneous anatomy, explaining the K=5 drop.
Transferability claims (R2). Our goal is to adapt a 2D natural-image foundation model to 3D medical segmentation, not to transfer among datasets. Results across five heterogeneous CT/MR datasets support robustness across tasks, not cross-dataset generalization. We will soften claims implying zero-shot or cross-dataset transfer.
Symbols & Fig.1 Corresponding(R2). We will add the annotations in the final version of Fig.1. Stage I representational analysis (R3). UMAP/t-SNE or attention analyses would strengthen this point, but rebuttal rules preclude new experiments. As indirect evidence, Table 3 row 2 uses only ICE, fine-tuned DINOv3, and an MLP head for predicting final results. With such limited decoding capacity, Stage I still improves Panc. DSC by 26.03.This gain confirms that the backbone has been adapted to medical domain.
Memory and runtime (R3). On AISD (mean input 512×512×21, RTX Pro 6000), DINO-Med3D Large/Base/Small require 7.10/6.23/5.89 GB and 1475/668/441 ms per case.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
Scores are 3/3/4, with the most confident reviewer recommending weak reject primarily on grounds of insufficient methodological detail and underexplained results. The paper is borderline-leaning-negative, but the substantive concerns are largely addressable through clarification, additional analysis, and targeted experiments rather than redesign. The paper is invited for rebuttal; authors should prioritize (a) a principled justification for the Stage I full-fine-tuning vs. Stage II frozen-backbone-with-LoRA design and a concrete account of what the 3D adapter contributes beyond upsampling, (b) representation-level evidence that Stage I produces medical-aligned features (e.g., linear probing, t-SNE/UMAP, or attention visualizations), (c) at least one cross-dataset or low-data transfer experiment that directly tests the transferability claim, (d) an explanation for the K-sensitivity on MSD-Colon and the discrepancy between reported AISD numbers and the dataset’s original paper, and (e) clarification of the Stage I proxy task, HRSE/ICE module diagrams, and runtime/memory figures.
- 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.
This paper proposes DINO-Med3D, a two-stage progressive framework that adapts the 2D DINOv3 encoder for 3D medical image segmentation. Experiments on five public CT and MRI datasets show improvements over CNN-, Transformer-, Mamba-, and DINO-based baselines. Reviewers agreed that the problem of adapting a strong 2D foundation model to 3D medical segmentation is well-motivated, the two-stage progressive design is intuitive, the semantic/detail dual-stream is well-aligned with DINO’s strengths and weaknesses, and the evaluation across five datasets with consistent gains is reasonably solid. They raised several concerns: (1) methodological clarity is limited — the proxy task location, HRSE/ICE structures, and symbol-to-figure mapping are under-described; (2) the absolute Dice values are lower than those reported in original AISD-dataset paper, raising credibility concerns; (3) Stage II’s switch from full fine-tuning to LoRA on a frozen backbone is not justified; (4) the 3D adapter appears to mainly upsample stacked features rather than build genuine hierarchical volumetric representations; (5) the K parameter (number of input slices) shows sharp degradation on MSD-Colon (K=3 vs K=5: 53.28→33.10), but the underlying mechanism is not analyzed; (6) the transferability/generalization claims rest on in-domain random splits without cross-dataset evaluation; (7) Stage I’s “medical-domain alignment” is not verified at the representation level (no UMAP/linear-probing/attention analysis); and (8) inference runtime and memory cost are not initially reported. After rebuttal, the authors clarified Stage I’s proxy task implementation, described HRSE/ICE structures in detail, explained the AISD random-split choice for fair comparisons, justified Stage II’s LoRA + frozen backbone (preserves Stage-I representation and avoids memory blow-up on 16-slice inputs), explained the 3D adapter’s role beyond simple upsampling, analyzed K-sensitivity via per-dataset foreground IoU and slices-per-lesion statistics, softened transferability claims, and provided memory/runtime numbers. Two reviewers raise their scores to Accept, while the most expert reviewer maintains Reject — the explanation for the unexpectedly low AISD metrics was deemed insufficiently persuasive and is seen as undermining credibility. Overall, the merits outweigh the drawbacks and I recommend accept
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper provided limited insights and has a method ambiguity. Though reviewers agreed on the positive merits. The paper provides limited contribution to the segmentation and pretraining studies.
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.
Even after the rebuttal, this paper remains borderline in my view. Although the authors addressed several of the raised concerns, the overall contribution still feels incremental, and some claims are not supported empirically (e.g., cross-domain generalization, which authors commit to smooth in the revised paper). Overall, I am not particularly enthusiastic about the submission, but considering all the concerns and how authors have addressed them, I recommend Accept.