Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Large-scale 2D foundation models exhibit strong transferable representations, yet extending them to 3D data typically requires retraining, adapters, or architectural redesign. We introduce PlaneCycle, a training-free, adapter-free operator for architecture-agnostic 2D-to-3D lifting of foundation models. PlaneCycle reuses the original pretrained 2D backbone by cyclically distributing spatial aggregation across orthogonal HW, DW, and DH planes throughout network depth, enabling progressive 3D fusion while preserving pretrained inductive biases. Using pretrained DINOv3 models, we evaluate PlaneCycle on six 3D classification and three 3D segmentation benchmarks. Without any training, the lifted models exhibit intrinsic 3D fusion capability and, under linear probing, outperform slice-wise 2D baselines and strong 3D counterparts, approaching the performance of fully trained models. With full fine-tuning, PlaneCycle matches standard 3D architectures, highlighting its potential as a seamless and practical 2D-to-3D lifting operator. These results demonstrate that 3D capability can be unlocked from pretrained 2D foundation models without structural modification or retraining. Code is available at https://github.com/HINTLab/PlaneCycle.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0966_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HINTLab/PlaneCycle

Link to the Dataset(s)

MedMNIST+ dataset: https://zenodo.org/records/10519652
LIDC dataset: https://www.cancerimagingarchive.net/collection/lidc-idri/
MMWHS dataset: https://huggingface.co/datasets/YuheLiuu/DiffAtlas_Preprocessed_Data/tree/main/MMWHS

BibTex

@InProceedings{YuYin_PlaneCycle_MICCAI2026,
        author = { Yu, Yinghong AND Li, Guangyuan AND Yang, Jiancheng},
        title = { { PlaneCycle: Training-Free 2D-to-3D Lifting of Foundation Models Without Adapters } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

The paper proposes an interesting way to adapt pretrained 2D foundation models to 3D medical imaging without adding new parameters or redesigning the network. The key idea is to process the volume across different anatomical planes throughout the network, which encourages cross-slice information sharing while still reusing the original 2D backbone. Empirically, that idea looks promising: the method gives more coherent 3D features than simple slice-wise inference and performs strongly under linear probing and fine-tuning.
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 is an interesting problem that is quite timely – the motivation is well articulated (the practical and sustainability argument for reusing expensive pretrained 2D models)

It has strong empirical structure. The evaluation is broader than many MICCAI papers, although MED-MNIST makes this a bit easier to do. The inclusion of paired t-tests and feature visualizations also helps support the central claim that the lifted model has intrinsic 3D coherence. The most compelling result is that under linear probing and even zero-training feature analysis it often outperforms slice-wise 2D and naive 3D baselines. That aligns well with the paper’s main thesis that the lifting operator itself unlocks 3D capability from pretrained 2D 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.

A comparison with 2.5D methods would be valuable Much of the classification evidence comes from MedMNIST-style datasets, and the segmentation sets are relatively small (It is not clear if the center-cropped nodules were used, making it an easier task) What is the recommendation between PCm, PCg? How were the 3D models initialized for the linear probing & finetuning experiments? Some gains may come not only from the lifting operator, but from starting with a very strong pretrained DINOv3 model, while some 3D comparators do not seem to have equally strong or equally matched pretraining. The paper argues this point in its favor, but it still makes the causal contribution of PlaneCycle alone harder to isolate.

There are a couple of important 3D baselines missing from the paper https://arxiv. org/abs/2501.09001 (A 3D CNN baseline) https://arxiv. org/abs/2511.17209 (only ViT-L available as far as I know)

The method is described as applicable to both CNNs and ViTs, but the experiments are almost entirely on lifted DINOv3 ViTs. The paper itself admits that systematic CNN-vs-ViT comparisons are still missing, so one of its headline claims is only partially validated. Are CT volumes tokenized with the original DINOv3 tokenizer?
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

I think it is an interesting and valuable idea. With inclusion of stronger, pretrained baselines (3D and 2.5D) for a fair comparison, I would increase my score
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’d characterize it as a strong idea paper with compelling early evidence, where the biggest remaining questions are fairness of comparison and how well the method generalizes beyond the ViT/DINOv3 + benchmark setup actually tested.
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.

My remarks were sufficiently addressed.

Review #2

Please describe the contribution of the paper

This paper presents PlaneCycle, a simple parameter-free operator for lifting pretrained 2D foundation models to 3D by cyclically applying the original 2D layers over orthogonal planes. Experimental results are ran on multiple 3D classification and segmentation benchmarks.
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 addresses a relevant problem: how to transfer strong 2D foundation models to 3D medical imaging without costly retraining or redesign. The motivation is clear and timely. 2.The core method is simple, clean, and easy to understand. The plane-cycling mechanism is intuitive, and the operator introduces no additional parameters.
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 central claim is somewhat broader than the evidence. The method is described as architecture-agnostic, but the experiments are almost entirely on DINOv3 ViTs; even the conclusion explicitly notes that systematic CNN comparisons are still missing. 2.The comparison set is not fully convincing for a MICCAI paper. Since the paper positions itself against adapters and other 2D-to-3D transfer strategies, stronger and more direct baselines from recent 3D medical FM adaptation literature would strengthen the claim substantially. The current baselines feel somewhat limited for the breadth of the paper’s claims. 3.Some claims are a bit overstated. For example, “unlocking 3D capability” is an appealing narrative, but the evidence is still limited to a particular set of datasets, backbones, and training setups. The large-scale and broader generalization claims remain preliminary by the authors’ own admission.
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?

1.The empirical validation is not sufficiently comprehensive relative to the paper’s broad claims. 2.The comparison to existing 2D-to-3D transfer strategies and adapter-based approaches is limited. 3.while the linear probing results are promising, the gains after full fine-tuning are less consistent. 4.Some claims appear overstated relative to the evidence.
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 authors’ response has addressed my main concerns. If the manuscript removes or appropriately softens the overclaimed conclusion, I believe the paper meets the standard for acceptance at MICCAI.

Review #3

Please describe the contribution of the paper

1: Proposes PlaneCycle, a generic operator for architecture-agnostic 2D-to-3D lifting. 2: Introduces cyclic orthogonal-plane aggregation for progressive volumetric fusion. 3: Achieves zero-parameter 3D adaptation of pretrained 2D backbones. 4: Demonstrates that pretrained 2D models already contain exploitable 3D inductive capability
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.

A key novelty of this work is that it introduces plane-wise cyclic processing directly inside the encoder, allowing 3D information to be accumulated progressively across layers instead of relying on late-stage fusion across the three axes. This strategy better exploits volumetric geometric relationships and reveals that large pretrained 2D models such as DINOv3 possess a surprising ability to align heterogeneous cross-plane features into a unified representation.
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 title appears somewhat over-claiming relative to the experimental scope: The title suggests a general training-free 2D-to-3D lifting mechanism for foundation models without adapters, but the empirical study is limited to pretrained DINOv3 models with a ViT backbone. Although the proposed operator is described as architecture-agnostic, the paper does not provide evidence on other foundation model families (MAE-pretrained Model, JEPA-pretrained model) Therefore, the claimed generality is not yet fully validated experimentally.

2.The paper lacks a detailed discussion of the representational trade-off introduced by cyclic plane-wise processing. Since spatial aggregation alternates across the HW, DW, and DH planes, each individual plane receives fewer opportunities for within-plane token interaction compared with conventional slice-wise encoding, effectively reducing same-plane attention depth. While the reported results suggest that cross-plane fusion remains effective, it remains unclear whether this reduced within-plane refinement may affect deep semantic feature formation or token consistency across layers.

2.This concern is particularly relevant for segmentation tasks, where both spatial consistency and semantic abstraction are important. However, the current segmentation evaluation may not fully reflect the semantic aspect of this trade-off. In particular, whole-heart segmentation is essentially a single-class foreground segmentation task, where performance mainly depends on anatomical continuity and foreground localization rather than semantic discrimination across categories. Therefore, it would be helpful for the paper to further discuss whether the proposed cyclic plane switching is expected to behave differently in settings where deeper semantic separation is required, such as multi-class or structurally heterogeneous segmentation scenarios.

3.A further point that would benefit from clarification is the fairness of the current comparison setup. Among the compared methods, PlaneCycle is the only approach that explicitly incorporates information from multiple orthogonal directions (HW, DW, and DH) within the encoder, whereas several baselines rely primarily on single-plane processing. This makes it difficult to disentangle whether the observed gains mainly come from the proposed cyclic design itself or simply from access to richer multi-axis volumetric information. A more detailed discussion of how PlaneCycle differs from simpler multi-direction aggregation strategies (e.g., independently processing different planes followed by feature addition or concatenation) would help clarify the source of improvement.
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.

(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?

I recommend acceptance because the paper presents a technically clean and practically valuable idea for extending pretrained 2D foundation models to 3D volumetric data without introducing additional parameters, adapters, or retraining overhead. The proposed PlaneCycle design is simple yet effective: by cyclically rotating spatial aggregation across orthogonal planes inside the encoder, it enables progressive volumetric interaction while preserving compatibility with strong pretrained 2D backbones such as DINOv3. A key strength is that the work demonstrates pretrained 2D foundation models can effectively reconcile cross-plane feature inconsistency and produce competitive 3D representations across both classification and segmentation tasks.

Although the paper would benefit from more discussion on representational trade-offs, comparison fairness, and the scope of generality beyond the tested backbone, these limitations mainly concern analysis depth rather than the core contribution. Overall, the method is sufficiently novel and well motivated to merit 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.

I would keep my recommendation as Accept, mainly due to the novelty of this work. However, some of the concerns raised in the reviews have not been fully addressed:

Missing comparison with 2.5D methods: I understand that using three axes would be computationally much heavier, but such a comparison would still be useful. Representation trade-off: The paper could better discuss the trade-off introduced by the proposed representation.

Author Feedback

We appreciate the constructive reviews from R1, R2, R3, and MR2, and their acknowledgment of the method’s technical clarity, practical value, and novelty (R1, R2, R3, MR2), as well as the experimental breadth (R1) and promising results (R1, R3). We address the key concerns below. 1.R1, R2, R3: Scope and experimental coverage PlaneCycle aims to be a generic lifting operator for pretrained 2D foundation models across diverse architectures and modalities. While our experiments are broad (R1) and consistently support this goal, we acknowledge that the full scope of our claims exceeds what is currently validated. Most limitations are discussed in the future work section, and we will soften the related descriptions and clearly state the current limitations in the revision. We are actively extending PlaneCycle to broader data, tasks, architectures, and foundation model families in follow-up work. Encouragingly, we have confirmed consistent lifting gains on ConvNeXt comparable to ViTs, though per MICCAI’s rebuttal policy we cannot report detailed results here. We commit to releasing a unified operator supporting both ViT and CNN architectures. 2.R1, R2: Comparison with 3D medical foundation models CT-FM and SPECTRE are relevant works we will cite and discuss in the revision, with thorough comparisons in follow-up work. Crucially, our method is orthogonal to these approaches: performance differences largely reflect pretraining data scale. After PlaneCycle lifting, the backbone can be further pretrained with their data and methodologies. The core value of PlaneCycle is reusing large-scale pretrained 2D models simply and effectively. We have observed PlaneCycle consistently outperforms SPECTRE on classification and segmentation under both linear probing and fine-tuning; detailed results cannot be reported here per MICCAI’s rebuttal policy, but will be included in the revision and in the extension. 3.R1, R3: Comparison with 2.5D methods We admit 2.5D methods are an important comparison, though not entirely fair given their additional fusion mechanisms and computational cost. Our results showed tri-plane fusion achieves comparable classification but lower segmentation performance, and under fine-tuning requiring 3x GPU memory and 3x training time, making it difficult to scale. We will discuss, cite, and add results where policy permits. R3: source of improvement: Simple multi-direction aggregation achieves comparable performance, confirming multi-axis access contributes. However, the cyclic design achieves this more efficiently at 1/3 the cost. 4.R2: Fine-tuning performance and computational efficiency PlaneCycle’s goal is not to outperform 3D models under fine-tuning: when fully fine-tuned, PlaneCycle effectively becomes a low-rank approximation of 3D attention, so comparable performance is expected and achieved. What we pursue is efficient lifting of pretrained 2D models without adapters at 2D computational cost. 5.R3: Representational trade-off We acknowledge cyclic plane switching reduces within-plane processing opportunities, and will provide more analysis in follow-up work. (1) Semantic feature formation: Frozen PCA visualizations show consistent features despite fewer HW passes, and PlaneCycle outperforms baselines on MMWHS (5-class whole-heart substructure segmentation), suggesting semantic features are well preserved. (2) Token consistency: PCg is designed to mitigate this in the linear probing regime; under fine-tuning the backbone adapts and the issue is substantially reduced. 6.R1: Experimental setup clarification (1) Baseline initialization: All baselines share identical DINOv3 pretrained weights, with the 3D baseline using 3D RoPE. (2) CT tokenization: All methods tokenize volumes with the original DINOv3 tokenizer per slice for fairness. (3) PCm vs. PCg: PCg is recommended for classification linear probing for CLS-to-patch token consistency, and PCm otherwise given comparable performance and simplicity.

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.

Reviewers have consensus on the impact and merit of this paper, some minor concerns need to be addressed.
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.

The authors did well on rebuttal; major concerns have been clarified. Reviewers have a consensus on the acceptance ratings.

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 recommend acceptance. The paper proposes PlaneCycle, a training free operator that lifts pretrained 2D foundation models to 3D by cyclically applying the 2D layers across orthogonal planes inside the encoder. All three reviewers recommend accept post rebuttal, and the rebuttal mainly clarifies points already implicit in the submission, such as identical DINOv3 initialization across baselines, the slice wise CT tokenization, the choice between PCm and PCg, and the fine tuning behavior. Some concerns remain, including the architecture agnostic claim being validated only on ViT, missing direct comparisons with 2.5D methods and recent 3D medical foundation models like CT-FM and SPECTRE, and limited depth in the segmentation evaluation. These are limitations of analysis depth rather than fatal flaws, and the authors have committed to softening the claims and adding the relevant discussions in the revision. Given the novelty of the cyclic in encoder lifting design, the consistent linear probing results supporting the main thesis, and unanimous reviewer support, the paper meets the bar for acceptance.

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.

The rebuttal improved the assessment mainly by clarifying several experimental and interpretative points. It stated that all baselines share the same DINOv3 pretrained weights, that CT volumes are tokenized with the original DINOv3 tokenizer per slice for fairness, and that PCg is recommended for classification linear probing while PCm is otherwise a simpler option with comparable performance. It also acknowledged that the current empirical evidence does not fully validate the broad architecture-agnostic and foundation-model-general claims, and committed to softening those claims in the revision. The rebuttal further clarified that PlaneCycle is intended to be orthogonal to 3D medical pretraining and adapter methods, and that its primary goal is efficient reuse of pretrained 2D models rather than systematically outperforming fully trained 3D models under full fine-tuning. New or not fully reported results mentioned in the rebuttal were not treated as decisive evidence, but the clarifications were useful for interpreting the submitted experiments. After rebuttal, all reviewers recommended Accept. This post-rebuttal convergence is credible: R1 indicated that the remarks were sufficiently addressed, R2 changed to Accept conditional on softening overstated claims, and R3 maintained Accept while noting that some concerns about 2.5D comparisons and representation trade-offs remain. These residual issues are real and should be reflected in the final version, especially by narrowing the claims around architecture-agnostic generality, adding clearer discussion of missing baselines, and avoiding unsupported large-scale generalization statements. However, they do not invalidate the core methodological contribution or the submitted empirical evidence.

back to top

PlaneCycle: Training-Free 2D-to-3D Lifting of Foundation Models Without Adapters

Author(s):