List of Papers Browse by Subject Areas Author List
Abstract
Accurate aneurysm segmentation in 3D rotational angiography (3DRA) is hindered by extreme class imbalance, morphological similarity to vessels, and absent large-scale 3D pretraining. 2D vision foundation models encode dense structural priors from ~1.7 billion images, yet naïve slice-wise transfer fragments anatomical continuity and destabilises optimisation. We propose DINO-3DRA, a dual-path framework achieving effective cross-dimensional semantic transfer by injecting frozen DINOv3 features into a 3D U-Net backbone via Room-Lite spatial mixing and calibrated residual fusion. On multi-centre 3DRA data, DINO-3DRA achieves state-of-the-art aneurysm segmentation (Dice: 0.758; HD95: 2.75 mm; +13% over nnU-Net) with only 5.72M trainable parameters. Ablation studies confirm that gains arise from structured cross-dimensional transfer rather than loss design alone, with bridged foundation features improving anatomical continuity between aneurysms and parent vessels. Without fine-tuning on CADA and SHINY-ICARUS, DINO-3DRA removes all failures (Dice < 0.5) observed in baseline architectures, indicating improved stability across heterogeneous protocols. Code and weights are available at https://github.com/JiayangDS/Dino3DRA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2993_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/JiayangDS/Dino3DRA
Link to the Dataset(s)
neurIST dataset: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3262441/
CADA dataset: https://cada.grand-challenge.org/
SHINY-ICARUS dataset: https://www.synapse.org/Synapse:syn45774070/wiki/619953
BibTex
@InProceedings{LuJia_DINO3DRA_MICCAI2026,
author = { Lu, Jiayang AND Lin, Fengming AND Frangi, Alejandro F. AND Sarrami-Foroushani, Ali},
title = { { DINO-3DRA: Leveraging 2D Foundation Model Semantics for 3D Cerebral Aneurysm 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 main idea of this paper is due to the data scarcity of volumetric models
(3D stacks) from 3DRA of Intracranial Aneurysms, one cannot do a full pre- training of optimized 3D architectures. Instead, the authors leveraged DINOv3,
which is a self-supervised vision foundation model pretrained on ≈ 1.7 billion 2D natural images, through a dual-path architecture such that they take the frozen DINOv3-Small (frozen weights) and add another 3D module that learns semantic features and domain-specific context of 3D volumes, thereby achieving foundation-scale benefits with minimal overhead.
- 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.
Strength 1: This is a well-written paper that introduces a simple yet long-overdue and interesting idea with substantial benefits. The work provides convincing benchmarks against the state of the art, including statistical analysis, and presents thorough ablations across multiple cases and different components of the pipeline.
Strength 2: The idea of the Room-Lite Spatial Mixer (RLSM) is interesting, and I’m curious to know whether the authors have experimented with increasing or decreasing the number of slices and assess how the RLSM behaves in Low or High SNR (i. e. Signal-to-Noise Ratio) cases.
Possible Strength 3: Another thing I’m curious about is whether the performance of DINO-3DRA behaves linearly with respect to the scale of the backbone model? Meaning, if we use a bigger model, say DINOv3- small-plus (28.7M Parameters) or DINOv3-base (85.7M Parameters), can we expect more improvement?
- 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.
Weakness 1: The introduction does not serve the paper effectively; several sentences come across as evasive and non-rigorous. For instance, the statement A fundamental asymmetry exists between 2D and 3D visual representation learning. Self-supervised 2D vision foundation models (VFMs) acknowledges the asymmetry, but does not explain why it exists. The paper addresses how this asymmetry can be tackled, yet it leaves the underlying rationale unexplored.
Weakness 2: The paper would benefit from a more comprehensive overview of the literature to better contextualize its contributions. Currently, it is heavily centered on DINOv3, with limited consideration of other relevant approaches. While the proposed idea is interesting, distillation and fine-tuning are well-studied research domains, and citing additional over- lapping works, such as doi. org/10.1007/978-3-031-72390-2_51, would help situate the contribution more clearly.
Weakness 3: The paper mentions that post-processing was applied on overlapping patch predictions (Section 3.2 - Post Processing), yet I didn’t see any isolation of the model’s raw predictions and the post-processed ones, or an ablation at least because, as it stands, there is no indication of how much impact the post-processing has on the predictions. Also, even if we consider that it was done only during preliminary experiments, the paper never mentions whether post-processing was applied or not in the reported final results.
- 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.
(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?
Overall, this is a good paper that presents an interesting idea. My current rating is 4/6.If the authors provide more details on the post-processing step—whose impact I anticipate is likely marginal—it could raise my rating to 5/6.However, if the post-processing proves to have a significant effect, my rating would decrease to around 3/6, since similar performance might then be achievable by applying the same post-processing to other models. Generally, I would recommend this paper for acceptance.
- 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.
1) The paper provides a well-engineered approach to improve the performance on aneurysm segmentation while reducing the computational overhead.
2) My concern on Post-Processing was addressed, although a quantifiable argument would have been more convincing. However, the manuscript should be revised to explain this more clearly.
Review #2
- Please describe the contribution of the paper
This paper proposes DINO-3DRA, a dual-branch framework that transfers frozen 2D vision foundation model features from DINOv3 to 3D cerebral aneurysm segmentation. The method uses a 3D U-Net as the volumetric backbone, introduces a Room-Lite spatial mixer to mitigate the lack of inter-slice continuity in slice-wise 2D features, and applies calibrated residual fusion to inject DINOv3 semantic features into the 3D decoding process in a stable manner. Experiments on the multi-center 3DRA @neurIST dataset show that the method outperforms several classical baselines for aneurysm segmentation while adding only limited parameter overhead and achieving lower performance variance. The paper also uses ablation studies to show that the gains mainly come from structured cross-dimensional semantic transfer rather than loss design alone, and further presents cross-dataset evaluation on CADA and SHINY-ICARUS to provide initial evidence of generalization.
- 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.
This paper addresses a clinically meaningful problem. 3DRA vessel and aneurysm segmentation is a challenging task with severe class imbalance and clear relevance to clinical detection and preoperative analysis. The method is also well motivated. Rather than simply applying slice-wise transfer, the authors explicitly target two key issues: the lack of 3D continuity in 2D features and the distribution mismatch between frozen foundation features and trainable 3D representations. The proposed Room-Lite mixer and calibrated fusion modules are therefore technically well justified. The experimental section is fairly comprehensive, including comparisons with U-Net++, Dual Attention, DeepVesselNet, and nnU-Net, as well as detailed ablations such as naive DINO injection, removal of Room-Lite, removal of calibration, random-feature control, and DINOv2 substitution. This makes the claims about where the improvements come from more convincing. Empirically, the method achieves meaningful gains in aneurysm Dice and appears to have lower performance variance. The cross-dataset validation is still limited, but it is nevertheless a useful addition in a medical imaging paper.
- 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 main weakness of the paper is that the methodological novelty is moderate. Overall, the work is best viewed as a well-executed adaptation of recent foundation-model transfer ideas to the specific task of 3DRA aneurysm segmentation, rather than a major new methodological breakthrough. Compared with Knaebel et al., “DINO in the Room: Leveraging 2D Foundation Models for 3D Segmentation” (arXiv, 2025), this paper similarly focuses on injecting 2D foundation features into 3D representations. It is also closely related to recent DINO-based medical segmentation works such as Gao et al., “DINO-UNet: Exploiting High-Fidelity Dense Features from Foundation Models for Medical Image Segmentation” (arXiv, 2025) and Li et al., “MedDINOv3: How to Adapt Vision Foundation Models for Medical Image Segmentation?” (arXiv, 2025). As such, the paper’s main contribution lies more in task adaptation and empirical validation than in a fundamentally new learning paradigm. Second, the evidence for cross-dataset generalization is still somewhat limited. The abstract and conclusion make fairly strong claims such as “eliminates all catastrophic failure cases” and “robust generalisation,” but the supporting evidence in the paper appears to be mainly qualitative, with limited external quantitative analysis and no sufficiently clear definition of failure cases. Third, although the paper compares against several classical 3D segmentation baselines, it would be stronger if it included more direct comparisons to recent foundation-model-based medical segmentation methods, especially transfer-based baselines that are closer to the current setting. Fourth, the paper tends to emphasize the improvement over nnU-Net, whereas the gain over the authors’ own strong 3D U-Net baseline is more modest, so the presentation could be more balanced. Fifth, reproducibility is acceptable but still incomplete, as the paper does not provide an anonymous code repository and could describe inference-time post-processing and external evaluation protocols in more detail.
- 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?
I assign a weak accept because the paper presents a meaningful performance improvement on an important and difficult clinically relevant task, and the method is well motivated and supported by reasonably thorough ablation studies. In particular, the authors identify two real obstacles in transferring 2D foundation model features to 3D volumetric segmentation, namely the lack of inter-slice continuity and the mismatch between frozen and trainable features, and propose a coherent solution to address them. This gives the paper some methodological value beyond simple tuning. At the same time, the paper falls short of a stronger recommendation because its novelty is still moderate, its relationship to recent closely related work should be clarified more carefully, and the cross-dataset generalization claims are currently supported more by qualitative than by strong quantitative evidence. Overall, I view this work as slightly above the acceptance threshold, though my final confidence would still depend on how well the rebuttal addresses these concerns.
- Reviewer confidence
Somewhat confident (2)
- [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
This paper proposes a dual-branch framework that integrates 2D DINOv3 features into a U-Net-based 3D workflow. The method employs Room-Lite Mixer and Calibrated Fusion to make slice-wise features suitable for volumetric segmentation. And the framework has been validated on a multi-center 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.
The paper presents a well-engineered and complete system for a clinically relevant segmentation task. The experimental results show consistent improvement over the 3D U-Net baseline.
- 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 main limitation is the relatively limited methodological novelty. The core idea is largely inspired by DINO in the Room: Leveraging 2D Foundation Models for 3D Segmentation (Knaebel et al., 2025), which also explores using 2D foundation-model features for 3D segmentation. Therefore, this work seems more like an application-oriented extension and empirical validation of that idea in 3D cerebral aneurysm segmentation, rather than a fundamentally new methodological contribution.
The proposed Room-Lite mixer and calibrated fusion modules are straightforward applications of existing spatial mixing and residual fusion techniques. Although these modules may improve performance, the paper does not clearly establish how they constitute a substantial conceptual advance beyond prior 2D-to-3D feature adaptation and fusion strategies, and the comparison is mainly limited to naive concatenation-based fusion.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend rejection because this paper is more of a solid application validation paper than one with substantial methodological innovation. Its main ideas and modules seem to merely extend existing 2D foundation-model-to-3D segmentation approaches, without clearly demonstrating a significant conceptual advance beyond prior spatial mixing and residual fusion strategies.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 rebuttal clarifies the clinical motivation and demonstrates that the proposed components improve robustness and preserve vessel-aneurysm continuity in a challenging volumetric setting. I acknowledge the empirical strength and potential clinical relevance of the work. However, my overall assessment regarding novelty remains unchanged. The method appears to be primarily an incremental adaptation and integration of existing ideas for 2D-to-3D feature alignment and fusion, rather than a substantial methodological advance. Therefore, although the work is technically solid and application-motivated, I do not believe the level of innovation is sufficient for acceptance.
Author Feedback
We thank the reviewers for their valuable feedback. Our responses are included below:
Meta/R1-Postprocessing: Post-processing is fixed for all models and has limited impact. Predictions were stitched from overlapping 64^3 patches by majority voting and connected-component cleanup, with a threshold tuned solely on the baseline 3D U-Net validation set, then frozen and applied identically to all models in Tables 3-5.Thus, the reviewer’s concern has already been tested. Every baseline received the same post-processing. If it dominated performance, the baseline on which it was tuned should benefit most. Fig. 3 shows patch-stitched predictions before connected-component filtering, where DINO-3DRA’s advantage is already visible. These operations remove isolated false positives but cannot recover missing aneurysm voxels, restore vessel-aneurysm continuity, or reduce artifacts.
Meta/R4/R2-Novelty/Clinical Application: Our contribution is clinically motivated 2D-to-3D adaptation methodology for a hard volumetric task. Aneurysm masks are the geometric front-end for patient-specific hemodynamic modelling and in-silico trials. This is challenging because aneurysms occupy <1%, resemble vessels, and must remain connected; missing necks or fragmented sacs break meshing and require manual repair. DINO in the Room (DITR) and DINO-3DRA solve different 2D-to-3D problems. DITR projects calibrated RGB-D features onto 3D points, whereas 3DRA has no external views, camera geometry, visibility model, or point cloud. DINO-UNet is 2D, and MedDINOv3 remains slice-wise. DINO-3DRA instead builds a dense 3D semantic volume from slice-wise DINO tokens, restores inter-slice coherence with Room-Lite, and aligns frozen VFM and trainable 3D features via calibrated residual fusion. This adaptation methodology is essential: Table 2 shows naive DINO/FAPM/multi-directional alternatives underperform, and Table 4 shows large drops without Room-Lite or calibration. Fig. 3 shows DINO-3DRA preserves vessel-aneurysm continuity. HD95 improves from 4.20 to 2.75 mm on @neurIST and 15.52 to 3.76 / 5.16 to 3.58 on SHINY-ICARUS/CADA, with Dice<0.5 failures reduced to 0%. Thus, we claim a better starting point for downstream meshing, not CFD-ready meshes.
R1-Scaling: We do not claim linear scaling. Table 4 suggests stronger adapted VFM features help: DINOv3-Small vs. DINOv2-small improves vessel Dice 0.868→0.897 and aneurysm Dice 0.590→0.758; DINOv3 scaling remains future work.
R1-Motivation/Slices/Lit: 2D VFMs have billion-scale pretraining, unlike 3D volumes; slice-wise transfer lacks depth continuity/alignment, motivating Room-Lite+Calib. We use all 64 slices per 64^3 patch and will cite the suggested work.
R2-Generalization: Table 5 reports external results; failure is Dice<0.5.Without fine-tuning, SI Dice/HD95 improves 0.718/15.52→0.793/3.76; CADA sigma/HD95 improves 0.186/5.16→0.093/3.58; failures 11.1/3.3%→0%. We will temper “robust”.
R2-Baselines: Direct DITR/2D DINO-UNet comparison is not modality-compatible. Table 2 compares closest adapters, including naive DINO, gated/calibrated fusion, FAPM-style and multi-directional projection; all underperform Room-Lite+Calib.
R2-Framing: We will emphasize 3D U-Net comparison: aneurysm Dice 0.711→0.758 (p=.030) and HD95 4.20→2.75 with only +~120K trainable params, a modest but significant gain.
R2-Repro: Paper specifies split, resampling, patch/stride, training, external datasets/metrics; First response above clarifies inference. Code will be released after acceptance.
R4-Fusion: Room-Lite uses depth positional bias+3D depthwise conv on stacked slice tokens. Calibrated fusion normalizes frozen features and preserves the 3D path via residual injection. Table 4 shows large drops on harder target (aneurysm) when either is removed.
Meta-Fig/Mesh: We will improve figures with zoomed high-resolution views. Lower HD95/continuity indicate fewer boundary outliers and manual repairs before meshing.
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.
The reviewers showed different opinions on this article, with assessment ranging from Accept (5), Weak Accept (4), to Reject(2). The confidence on the reviews was Very confident (4), Somewhat confident (2), and Confident but not certain (3). The comments from Rv1, expecting to change opinion after the rebuttal fits better into a weak acceptance than an acceptance. Acceptance assumes making a decision independently from any rebuttal phase.
After analyzing the reviews and studying the paper, I believe that the manuscript will benefit from a rebuttal phase.
I found the comments from Rv1 interesting for the improvement of the manuscript. In addition, a satisfactory response regarding the post-processing step is relevant for the acceptance of the paper.
Rv2 and 3 commented on the low methodological novelty of the manuscript. The application is relevant and difficult so maybe more effort on providing strong applicability insights may help the reviewers to rise their initial assessment. The differences of the proposed method with respect to recent state of the art should be properly argumented.
Meta Reviewer found the resolution from the figures of low quality. The details of the obtained segmentation cannot be properly assessed through zooming. I would suggest to improve the quality to be able to perceive better the results of the proposed method. In addition, I would expect some comment regarding metrication errors, since the lack of smoothness in the segmentations may difficult obtaining meshes suitable for subsequent analysis.
- 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.
After carefully reading the author’s rebuttal and the final reviewer’s assessment, I lean toward accepting the manuscript. The main reason the paper is considered borderline is the limited novelty of the proposed methodology. However, I believe there is value in successfully applying an existing methodology to a novel and challenging problem and sharing these results with the community. Such contributions can provide useful insights and may help other researchers apply similar approaches to analogous problems.
In current research, achieving methodological novelty is increasingly difficult, mostly due to the static methodological nature of neural network architectures. In this context, I consider it more valuable to present a well-executed application of an established method to a complex problem, offering meaningful insights than to introduce minor or unnecessary architectural modifications applied to a less demanding task.
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.
I would never place a reject conclusion entirely on novelty. Reviewers pointed out that the method is an incremental extension of “DINO in the room.” I checked that against the literature, and it does not hold as a reject ground.
DINO in the room is point-cloud scene segmentation with camera geometry and projection, a different 2D-to-3D problem. 3DRA has no views, camera, or point cloud. The closest medical mechanism, the dual-branch frozen-DINOv3 adapter from Dino U-Net, is already in the paper’s own ablation and is beaten by the proposed design. No prior work applies foundation-model transfer to aneurysm segmentation on 3DRA, so this is the first attempt on this clinical problem.
The nearest volumetric-medical relatives are concurrent 2026 preprints on other anatomy. On substance the paper holds.
It targets a real clinical task with severe class imbalance, and its main claims are backed by the submitted numbers: a significant gain over the authors’ own strong 3D U-Net (Dice 0.711 to 0.758, with statistical significance analysis) with only about 120K extra parameters, and catastrophic Dice<0.5 failures driven to zero, including cross-dataset.
The experiments needed to support those claims are present, and the one genuine overclaim, “robust generalisation,” is wording the authors agree to temper. The contribution is modest at the pure technical component level but sound, well-evaluated, and first on this problem. So I recommend an accept for plain poster presentation.
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.
This is a borderline paper with mixed reviews. Reviewers agreed that the work addresses a clinically important and technically challenging problem—3DRA vessel and aneurysm segmentation—with a well-engineered system, thorough experimentation, and meaningful empirical gains over strong baselines such as nnU-Net and 3D U-Net. However, they also raised serious concerns about the paper’s significance and novelty: first, the methodological contribution is largely incremental, extending existing 2D-to-3D foundation model transfer ideas without introducing fundamentally new learning paradigms or architectural breakthroughs; second, claims regarding cross-dataset generalization and the elimination of catastrophic failures are stronger than the quantitative evidence provided, with limited external validation and unclear definitions of failure cases; third, the paper relies heavily on comparisons with classical baselines while omitting closer, more relevant foundation-model–based medical segmentation competitors, making it difficult to fully contextualize the advance. The rebuttal clarified clinical motivations and demonstrated robustness benefits, but it did not adequately resolve concerns regarding limited novelty and overstated claims. After careful deliberation, I lean towards accepting this manuscript.