List of Papers Browse by Subject Areas Author List
Abstract
Accurate brain segmentation across diverse atlas protocols is essential for neuroimaging, yet it is hindered by the high cost of generating large-scale ground truth labels and the prohibitive overhead of re-training models for new protocols. While vision foundation models offer powerful representations, their seamless adaptation to 3D multi-atlas segmentation in few-shot scenarios remains challenging. In this study, we propose a unified few-shot framework that adapts large-scale vision models to neuroimaging via an efficient “pre-train, fine-tune and inference” paradigm. By coupling anatomically-prioritized self-supervision with an $O(N)$ Linear Transformer, our model learns a universal, semantically consistent latent space for dense semantic label transfer. Instead of exhaustive re-training, we introduce a “Lock-and-Finetune” strategy that exclusively updates a lightweight refinement module to recover fine structural boundaries. Experiments on the Human Connectome Project (HCP) dataset show our approach achieves over 95\% of fully supervised nnU-Net performance using less than 5\% of labeled data. Notably, the model adapts to unseen atlases using only a few samples, providing an efficient solution for neuroimaging segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3573_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/GBCWORLDWALKER/FewshotBrain
Link to the Dataset(s)
HCP dataset: https://www.humanconnectomeproject.org/data/
BibTex
@InProceedings{GuoBoc_AUnified_MICCAI2026,
author = { Guo, Bocheng AND Zhang, Xiyuan AND Zhang, Wei AND Pasternak, Ofer AND O’Donnell, Lauren J. AND Zhang, Fan},
title = { { A Unified Few-Shot Framework for Multi-Atlas Neuroimaging Segmentation Leveraging Vision Foundation Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses a relevant and underexplored problem in neuroimaging: transferring learned representations across anatomical atlases with different label spaces, using only a small number of labeled examples from a target atlas. The authors propose a multi-stage framework built around a DINOv3-initialized ConvNeXt encoder coupled with a feature pyramid network (FPN), trained in a self-supervised fashion via a masked autoencoding objective that encourages reconstruction of neighbouring slices, thereby implicitly encoding inter-slice context within an inherently 2D architecture.
A key contribution is the atlas anchoring mechanism introduced during pre-training, in which a linear transformer aligns query atlas features (from 3 atlases) to key/value representations from an anchor atlas. A lightweight 3D kernel then aggregates a stack of consecutive 2D feature maps to recover volumetric coherence. At inference time, a super-resolution module — the only component fine-tuned per target atlas — adapts the predictions to the query atlas resolution using few labeled examples. The overall design philosophy, separating heavy pre-training from lightweight few-shot adaptation, is well-motivated for the low-annotation regime typical of multi-atlas neuroimaging workflows.
- 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 can be of interest for the MICCAI community for several reasons (S1-4):
S1: Practical relevance. The problem of cross-atlas transfer is of genuine interest to the neuroimaging community, where atlas ecosystems (e. g. , Desikan–Killiany, Destrieux, AAL) differ substantially in resolution, label granularity, and acquisition protocol. A method capable of leveraging representations learned on one atlas to bootstrap segmentation on another addresses a real bottleneck in atlas-based pipelines.
S2: Strong backbone choice. The use of DINOv3 as initialisation is well-motivated: its Gram anchoring mechanism mitigates the known degradation of dense feature maps during long training schedules, which is precisely the failure mode most damaging to spatially precise segmentation tasks. The combination with a ConvNeXt encoder and FPN is architecturally sound, offering multi-scale dense features while remaining compatible with volumetric inputs through the 3D aggregation kernel.
S3: Principled few-shot setup. Restricting fine-tuning to the super-resolution module with k=5 labeled atlases is a disciplined design choice that keeps the adaptation cost low and reduces the risk of overfitting. The separation between pre-training and adaptation is clearly structured.
S4: Competitive evaluation scope. The inclusion of both fully supervised baselines (FastSurfer, nnU-Net, MedSAM2) and few-shot comparators (SAM3-Adapter, DSPNet) gives a reasonable coverage of the relevant landscape.
- 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 paper presents, in my opinion, 2 major (W1-2) and 3 minor (W3-5) weaknesses:
W1: Handling of variable label spaces is not addressed. Different atlases in the neuroimaging community define between 10 and over 100 anatomical structures. The paper does not explain how the segmentation head accommodates this variability: does the output channel dimension change per atlas, and if so, how is the model initialised and fine-tuned for atlases with unseen class counts? This is not a minor detail — it is central to the claimed generalisability of the method and must be addressed explicitly.
W2: Fairness of comparison with supervised baselines. In a few-shot scenario, it is unclear whether FastSurfer, nnU-Net, and MedSAM2 are first pre-trained using the available atlases (wmparc, ribbon, aparc) and fine-tune it using k=5 (initialising the final layer dimension according to each atlas) or trained from scratch using only k=5.If the latter, the comparison is not entirely fair and the performance gap — in either direction — is difficult to interpret. The authors should clarify this and, if necessary, acknowledge this asymmetry and discuss its implications.
W3: Ambiguity in multi-slice input encoding. It is not clear from the manuscript whether neighbouring slices are provided as additional input channels to the 2D encoder, or whether only single-slice inputs are used and inter-slice context is captured solely through the masked AE reconstruction objective. This distinction affects the effective receptive field and should be clarified.
W4: Evaluation details are missing. The experimental setup does not describe the validation set: how many subjects are there? The required pre-processing is not specified, making the reproducibility assessment difficult.
W5: Evaluation limited to a single dataset. All experiments are conducted exclusively on HCP-3T data. While this is a well-established neuroimaging benchmark, it represents a specific acquisition protocol and scanner environment. The proposed method’s central claim — generalisation across atlases — would be substantially strengthened by evaluation on at least one additional dataset (e. g. , ABIDE, ADHD-200, or a clinical 1.5T cohort), which would also reveal sensitivity to scanner variability, field strength differences, and population heterogeneity.
- 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
Some other comments to improve the readability of the manuscript are the following:
-
The acronyms in Table 1 are not defined in relation to the atlases defined in Section 3.1.Which one is the aparc+aseg (148 labels)? What is aparc (gm) and aparc (wm)?
-
In the UMAP projection experiment, the authors state that “voxels belonging to the same anatomical structures are clustered into identical color manifolds without any target-task supervision. However, the model is pre-trained with an atlas-anchored segmentation head, which provides direct structural supervision. This claim should be revised to accurately reflect the role of the anchoring atlas in shaping the feature space.
-
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper tackles a meaningful and timely problem — cross-atlas transfer for neuroimaging segmentation under low annotation budgets — and proposes a technically coherent framework whose design choices are generally well-motivated. The use of robust pre-trained encoder, the atlas anchoring mechanism, and the separation between heavy pre-training and lightweight few-shot adaptation collectively constitute a contribution of sufficient interest to the MICCAI community. That said, several concerns temper a stronger recommendation. There are some critical details missing: the handling of variable label spaces across atlases is left unexplained, and the fairness of comparisons against fully supervised baselines is uncertain. Additional clarifications regarding the input encoding, validation protocol, and dataset scope would further strengthen the work. If the authors can address these points in a revision — particularly W1 and W2 — the paper would make a solid contribution to the conference. In its current form, the promise of the approach outweighs its presentational and experimental shortcomings, and acceptance is recommended conditionally.
- 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.
This paper addresses a relevant and underexplored problem in neuroimaging: transferring learned representations across anatomical atlases with different label spaces, in a few shot learning scenario. The authors committed to a list of updates from their original manuscript that tips the balance to an “accept” decision.
Review #2
- Please describe the contribution of the paper
The paper proposes a unified few-shot neuroimaging segmentation framework that leverages pre-trained vision foundation models and atlas-based priors. By combining lightweight fine-tuning with structural guidance from atlas representations, the method aims to achieve competitive segmentation performance under limited annotation settings. The framework further explores adapting 2D foundation models to 3D medical imaging through modular design.
- 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 unified framework integrating foundation models with multi-atlas priors for few-shot neuroimaging segmentation.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
While the approach is technically sound and experimentally validated, the overall contribution appears incremental, with limited methodological novelty beyond the integration of existing components. Additional analysis and stronger justification of the design choices would further strengthen the work.
1.The experimental evaluation is conducted solely on the HCP dataset, where different atlas protocols are used to simulate varying segmentation granularities. While this setup demonstrates robustness across label spaces, it does not constitute true cross-dataset or out-of-distribution generalization. As a result, the claim of “strong generalization capability,” particularly under few-shot settings, may be overstated. Additional validation on independent datasets or under more diverse acquisition conditions would strengthen the conclusions. 2.While several baselines are included, it would be beneficial to discuss or compare with methods such as SynthSeg (MedIA 2023), which are specifically designed for robust cross-protocol brain MRI segmentation without retraining. Including such comparisons would help better position the proposed method with respect to existing approaches. 3.The manuscript emphasizes foundation model adaptation and rapid clinical deployment, but the supporting evidence is limited. The contribution of the pre-trained backbone is not isolated (e.g., no comparison with training from scratch), and efficiency aspects such as adaptation time and computational cost are not quantified. Moreover, the claim of clinical deployment remains largely conceptual, as no real clinical setting or multi-center validation is provided. 4.The reported few-shot performance may benefit from favorable experimental conditions and potential inconsistencies in baseline settings. In particular, the use of a pre-trained foundation model, atlas priors, and relatively homogeneous brain imaging data suggests that the setting is closer to “strong prior + limited annotations” rather than pure few-shot learning. Furthermore, it is unclear whether the baselines (e.g., nnU-Net) are trained and evaluated under strictly identical data splits, preprocessing pipelines, and supervision settings (e.g., full training vs. few-shot fine-tuning, 5-fold optimization). Without such alignment, the claimed “near fully supervised performance” may be overestimated. 5.Fig. 2: The term “WM segm.” is not defined in the manuscript and should be clarified.
- 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 addresses an important problem and demonstrates promising empirical performance under few-shot settings. However, the experimental validation and analysis are not sufficiently comprehensive to fully support some of the stronger claims (e.g., generalization and clinical deployment), which limits the overall impact.
- 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.
I appreciate the authors’ detailed rebuttal and clarification efforts. The responses addressed several of my concerns regarding comparison fairness, the role of pretraining, efficiency analysis, and the interpretation of the few-shot setting. I also appreciate that the authors acknowledged the limitations of the current HCP-based evaluation and agreed to soften some of the stronger claims. In particular, the commitment to provide external cohort demonstrations in the public repository upon acceptance further strengthens the practical relevance of the work. While the contribution remains somewhat incremental and true cross-dataset validation is still limited in the current manuscript, I do not see any fatal methodological flaw, and I therefore slightly increase my score and support acceptance.
Review #3
- Please describe the contribution of the paper
This paper introduces a unified few-shot framework for multi-atlas 3D brain MRI segmentation by adapting 2D vision foundation models to volumetric neuroimaging, combining anatomy-aware pretraining, efficient 3D associative matching, and a lightweight lock-and-finetune adaptation strategy for rapid transfer to unseen atlases.
- 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.Clear and well-structured presentation. The paper is clearly written and generally easy to follow. The motivation, methodology, and experimental setup are presented in a structured manner, allowing the reader to understand the overall framework without excessive effort. The separation into different stages (e.g., pre-training, matching, and finetuning) further improves readability and helps clarify the role of each component within the pipeline.
2.Addresses an important problem with empirically effective results. The paper focuses on the problem of multi-atlas brain segmentation under limited annotation, which is of practical importance in neuroimaging. The proposed method targets a meaningful challenge—reducing the need for large-scale labeled data and repeated retraining across atlas protocols. Experimental results indicate that the approach is effective in this setting, demonstrating strong performance in few-shot scenarios and suggesting its potential utility for real-world applications.
- 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.
Weaknesses
1.The necessity of the “Lock-and-Finetune” strategy is not sufficiently justified. While the proposed Lock-and-Finetune strategy is presented as a key component for efficient adaptation, its necessity is not convincingly demonstrated through empirical evidence. In particular, the paper does not provide comparisons against full finetuning of the entire model under the same few-shot setting. Such comparisons are important to assess whether the proposed strategy leads to a meaningful trade-off between performance and efficiency. Additionally, quantitative analysis of computational cost (e.g., training time, GPU memory usage) is missing. Even if the expected outcome may be intuitive, these experiments are still necessary to substantiate the claimed advantages.
2.Insufficient details on how baselines are trained, leading to potential unfair comparisons. The paper does not clearly specify how the baseline methods are trained and adapted to the few-shot setting. It is unclear whether these baselines follow their original training protocols (e.g., full finetuning) or are adapted using a strategy comparable to the proposed Lock-and-Finetune approach. This ambiguity raises concerns about the fairness and validity of the comparisons. A clear and consistent description of the training and adaptation procedures for all baselines is necessary to ensure that the reported performance gains are meaningful.
3.The pretraining design is complex but lacks supporting ablation studies. The proposed pretraining pipeline involves multiple components, including a two-stage pretraining strategy and a multi-task learning objective that combines reconstruction and anatomical supervision. However, the paper does not provide sufficient ablation studies to validate the necessity of these design choices. For example, the contribution of each pretraining stage and the individual impact of different learning objectives are not systematically analyzed. Without such experiments, it is difficult to assess whether the complexity of the pretraining pipeline is justified or whether simpler alternatives could achieve similar performance.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The 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 paper addresses an important and practically relevant problem, namely few-shot multi-atlas brain segmentation under limited annotation. The motivation is well-articulated, and the overall framework is clearly presented. The proposed method appears effective in the few-shot setting, and the experimental results suggest that it can achieve competitive performance compared to both fully supervised and few-shot baselines. These aspects make the work potentially valuable for real-world neuroimaging applications.
However, several key issues limit the strength of the contribution. First, some central design choices are not sufficiently justified through experiments. In particular, the necessity of the proposed “Lock-and-Finetune” strategy is not supported by direct comparisons with full finetuning or by quantitative analysis of efficiency, making it difficult to assess its actual benefits. Second, the training procedures of the baseline methods are not clearly described, raising concerns about the fairness and validity of the reported comparisons. Third, although the pretraining pipeline is relatively complex (e.g., two-stage design and multi-task objectives), the paper lacks sufficient ablation studies to validate the contribution of these components.
In addition, the reproducibility of the work is limited. The code is not available at submission time, and important implementation details are missing, particularly regarding training configurations and data preprocessing. Given the complexity of the proposed framework, this further weakens confidence in the reported results.
Overall, while the paper presents a promising direction and shows encouraging empirical results, the lack of sufficient experimental validation and clarity in key aspects prevents a stronger recommendation. I therefore lean towards a weak reject.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and AC for their comments and for recognizing the value of our few-shot multi-atlas segmentation framework.
Varying Label Counts (R1, AC): We will clarify that, when generalizing to a new atlas, the Super-Resolution Refinement module that includes the segmentation head will be fine-tuned using a few subjects annotated with the new atlas. During this stage, we will adjust the output channel dimension to match label count in the new atlas.
Comparison Fairness (R1&2&3, AC): Our compared few-shot nnU-Net and MedSAM2 methods are trained from scratch using 5 subjects, SAM3-Adapter is under pretrain-finetune, and DSPNet does not support pretrain (FastSurfer was not compared under few-shot as it generated unfavorable results). Following R1’s suggestion, for nnU-Net and MedSAM2, we pretrain and then fine-tune the final layer according to the new atlases. The results show that 5‑shot fine‑tuning of nnU‑Net from the aparc‑pretrained model achieves Dice=0.704 on Yeo7 atlas (ours=0.797) and Dice=0.471 on HCP‑MMP1 atlas (ours=0.632). We will clarify experimental settings, update results, and note our few‑shot setting follows anatomical prior + limited annotations as suggested by R3. Input and Validation Ambiguities (R1&2, AC): We will clarify that: 1) the 2D encoder takes only a single slice as input, with neighboring slices as reconstruction targets in SA‑MAE to learn inter‑slice context, and 2) the comparison with baselines was performed on the same preprocessed data (provided in the HCP database), with identical training subjects and the same validation set (n=25).
Cross-dataset Generalization (R1&2, AC): We fully agree that cross-dataset evaluation is important. While our MICCAI submission focuses on cross-atlas generation, we have since conducted experimental evaluation on external cohorts (such as ABIDE-II, ADNI, LA5C, OASIS and MIRIAD) and found strong generalization (achieved average Dice of 0.62 and 0.83 on HCP-MMP1 and Yeo atlases). We acknowledge that including these results is not permitted; instead, we will provide a demonstration on these data in our GitHub repository upon acceptance. We will also expand our discussion and future work, and soften the “strong generalization capability” claim as pointed out by R2. Related Methods (R2): We will include discussion on related methods e.g. SynthSeg, whose focus is robust inference across contrast/resolution/domain shifts for a predefined anatomical label protocol, whereas our work focuses on few-shot adaptation to unseen atlas label spaces.
Lock-and-Finetune Strategy and Efficiency (R2&3): We will justify our “Lock-and-Finetune” strategy by reporting results from: 1) scratch‑versus‑finetuning comparisons under 5-shot setting (scratch: 0.747/0.372, versus fine‑tuning: 0.797/0.632 on Yeo/HCP‑MMP1 in Dice) and 2) the computational comparisons for model fine‑tuning (on RTX 3090, ours: ~0.3 h and 4 GB GPU memory, versus nnU‑Net: ~2 h and 22 GB GPU memory). Further, we agree that this work is at the methodology development stage, and will soften our claim regarding clinical deployment as pointed out by R2. Ablation (R3): We will report ablation results: Removing Stage‑1 anatomical supervision slightly degrades performance from Dice=0.79/0.63 to 0.76/0.55 on Yeo/HCP‑MMP1; Removing Stage‑1 reconstruction objective causes a larger decrease to Dice=0.63/0.47; Removing Stage‑2 matcher degrades performance to Dice=0.71/0.42. Code and Details (R3): We confirm that the code will be made public upon acceptance and will clarify training configurations and data preprocessing steps.
Readability and Typo (R1&2): We will: 1) clarify that we separately evaluate the GM and WM components in the “aparc+aseg” atlas, which are “aparc (gm)” and “aparc (wm)” in Tab. 1; 2) revise UMAP visualization to clarify the role of the anchoring atlas in shaping feature space; 3) fix typo “WM segm” in Fig. 2 as “wmparc” where we show subcortical structures under the wmparc atlas.
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 paper proposes a unified few-shot framework for multi-atlas brain MRI segmentation that adapts vision foundation models to 3D via anatomically-prioritized pre-training, an O(N) Linear Transformer matcher, and a Lock-and-Finetune adaptation strategy. The reviewers raise substantive concerns regarding how the segmentation head accommodates atlases with varying label counts, the fairness of comparisons against fully supervised baselines under the few-shot regime including the exact training and adaptation protocol used for each baseline, ambiguities in the multi-slice input encoding and validation protocol, and the scope of generalization claims given the single-dataset evaluation. The rebuttal should clarify these points along with the other issues raised by the reviewers.
- 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 a unified few-shot framework for multi-atlas brain MRI segmentation that adapts a vision foundation model to 3D through anatomically-prioritized pre-training, an O(N) linear transformer matcher, and a lock-and-finetune adaptation strategy. The reviewers agree that cross-atlas transfer under limited annotation is a practically important problem, that the separation of heavy pre-training from lightweight adaptation is well motivated, and that the results are competitive against both supervised and few-shot baselines. The rebuttal clarified how the segmentation head accommodates atlases with differing label counts, the training and adaptation protocol used for each baseline, the single-slice input with neighboring slices as reconstruction targets, and the validation protocol, and two reviewers confirmed these resolved their concerns. The main remaining limitation is the single-dataset evaluation, which leaves the generalization claim only partially supported, and the authors are encouraged to retain the softened framing of that claim and to fold the clarified baseline protocol, label-space handling, and readability corrections into the camera-ready version.
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.
R1 and R2 upgraded to Accept after the authors clarified the handling of varying label spaces, the fairness of few-shot baseline comparisons, the role of pretraining, and the efficiency of the proposed adaptation strategy, while concerns about limited cross-dataset validation were acknowledged and the corresponding claims were softened. Although external validation remains limited, the paper addresses an important problem with a technically sound and practically useful framework, and I recommend Accept.
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.
I am inclined to accept this paper because it addresses an important and meaningful problem. The proposed method is also reasonably innovative. Reviewer #3 raised some concerns about the details of training procedure and ablation studies, but I believe these were effectively addressed in the rebuttal. I would recommend that the authors further expand this work into a journal article.