Abstract

Three-dimensional medical images inherently exhibit spatial anisotropy due to non-uniform slice spacing and heterogeneous acquisition protocols. However, most 3D Transformer architectures assume isotropic voxel grids and rely on fixed resampling to standardize inputs. This mismatch introduces two key limitations: (1) forced resampling distorts native anatomical geometry and may degrade fine structural details; (2) isotropic positional encoding may overlook instance-specific spacing, limiting spatial precision. While recent anisotropy-aware designs introduce direction-specific priors, they typically depend on static, dataset-level scaling factors that require manual hyperparameter tuning and lack instance-specific adaptivity. To address these limitations, we propose Learnable Anisotropic Deformation (LAD), a geometry-aware adaptation framework for 3D Transformers. LAD learns instance-specific spatial scaling factors directly from image features and incorporates them into both positional encoding and feature modulation. We design a frequency-modulated anisotropic Rotary Positional Embedding (RoPE) to adjust spatial phase and an adaptive anisotropic feature modulation module to recalibrate semantic responses along biased axes. This design enables Transformers to respect native volumetric structure without requiring explicit spacing metadata. We integrate LAD into three representative 3D backbones and evaluate it on volumetric segmentation and multi-disease classification benchmarks. Across four datasets (17,359 3D volumes) and six tasks spanning CT and MRI, LAD consistently improves performance across backbones. Further analysis shows that the learned scales correlate with physical spacing in segmentation and adapt across classification tasks, supporting both geometric consistency and task adaptivity. The code is available at https://github.com/chengcailiu/LAD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0291_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/chengcailiu/LAD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiuChe_LAD_MICCAI2026,
        author = { Liu, Chengcai AND Sang, Haolin AND Huang, Xingyu AND Wu, Yi AND Wang, Chenghao AND Wu, Weiqiang AND Zeng, Guangjian AND Tian, Jie AND Wang, Shuo},
        title = { { LAD: Learnable Anisotropic Deformation for 3D Medical Image Analysis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16889},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    LAD operates by dynamically perceiving geometric priors directly from image features without requiring explicit physical spacing metadata. It learns instance-specific spatial scaling factors and injects them into both the spatial phase (positional encoding) and semantic amplitude (feature modulation) of the network.

  • 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.This paper introduces an Instance-aware Anisotropy Learner. It looks at the image features and guesses the right scaling factor for each specific patient.

    2.LAD doesn’t just fix position encoding, and it fixes the distorted textures caused by bad resolution.

    3.The experiment design are solid. It works on UNETR, SwinUNETR, and Primus. That covers ViT, Swin, and EVA-02 backbones.

  • 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 the introduction, the authors criticize existing methods for using forced resampling, saying it destroys anatomical geometry. However, in the experiments, the authors resized everything to a fixed $48 \times 256 \times 256$ grid. This is contradictory.

    2.Eq. (1) uses Global Mean Pooling to turn a whole 2D slice into a 1D vector to find the scale factor $s_k$. If the authors just average the whole plane, features will lose all the texture and anatomy.

    3.In Fig. 2b, the scaling factor changes based on the disease type, not the image geometry. If the factor changes with the disease, it’s not geometric scaling, i think it’s just channel attention (like SE-Net).

    4.The authors say this method doesn’t need spacing metadata from DICOM files. However, most DICOM files have perfectly accurate spacing data, especially in real clinical applications. I think this work needs a baseline where the real physical spacing is taken into account.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission 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.

    (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?

    Method novelty and application potential

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper introduces a novel learnable anisotropic deformation (LAD) approach for addressing the spatial biases (e.g., fixed resampling and isotropic voxel grids, etc.) in 3D Transformer architectures.

  • 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.Well-motivated problem setting and thoughtful design: The method is well motivated by the mismatch between anisotropic data and presumed isotropic models, which results in spatial bias, and is highly relevant to various tasks (e.g., classification and segmentation). The proposed components—instance-aware anisotropy learner, adaptive anisotropic feature modulation, and frequency-modulated anisotropic RoPE within the LAD Attention Block—are all well designed for the purpose of learning scaling factors, feature amplitude (semantics), and spatial phase, respectively. 2.Extensive empirical evaluation: The authors have conducted sufficient experiments and demonstrated superior performance when incorporating LAD into three Transformer models (UNETR, SwinUNETR, and Primus) across several benchmarks, including one multi-organ abdomen segmentation task and three downstream tasks: abnormality prediction for atelectasis, lung opacity, and bronchiectasis; brain IDH mutation prediction; and multi-center lung cancer subtype prediction. 3.Solid comparative and ablation studies: The method is also well supported through comparison with an anisotropy-aware baseline, thorough ablation of several proposed components (isotropic RoPE, anisotropic RoPE without AAFM, and full LAD), and analysis of the correlation between physical and learned spatial scale ratios.

  • 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.

    Although the paper is well motivated and the reported performance gains are promising (Tables 1 and 2), a few additional analyses could further enhance the discussion of its potential clinical relevance.

    1.Performance analysis: While sufficient quantitative analysis is provided for the overall performance gains of LAD (Tables 1 and 2), it may also be informative to examine segmentation improvements for individual organs, both quantitatively and qualitatively. This could offer better insight into the performance gains across different organs and provide additional context for downstream tasks, such as abnormality prediction for different organs (Fig. 2b). 2.Statistical analysis: The results are primarily presented in terms of mean performance differences across methods. Including statistical significance analysis could further contextualize the observed improvements of LAD over the baselines.

  • 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

    1.Notation clarification: The notation Δz in Section 2.3 is not explicitly defined and could be clarified. 2.Definition of “Single Axis Learner”: The “Single Axis Learner” shown in Fig. 1 is somewhat difficult to interpret before reading Section 2.2.However, Section 2.2 does not explicitly define this term. A more explicit definition may improve readability.

  • 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 problem is well-motivated, and the method is well-designed. However, the results (e.g., quantitative performance) and the analysis (statistical testing) are limited or absent, raising concerns about the performance gain and clinical applicability.

  • 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.

    The authors have clarified that they will release the code and data to better understand segmentation performance across organs and the downstream classification task, which addresses my concerns.



Review #3

  • Please describe the contribution of the paper

    This paper proposes Learnable Anisotropic Deformation (LAD), a framework that adapts Transformer representations to anisotropic 3D medical data without relying on fixed resampling or metadata.

    The key idea is to learn instance-specific scaling factors from image features and use them to jointly modulate:

    spatial encoding (via an anisotropic variant of RoPE), and feature representations (via channel-wise modulation).

    This results in a unified mechanism that adjusts both where attention operates and how features respond under anisotropic conditions.

  • 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 real and often overlooked issue—mismatch between isotropic model assumptions and anisotropic medical data—which is clearly articulated and practically important.

    2.The combination of learned scaling with both positional encoding and feature modulation is well thought out. The dual-pathway design is not just additive but complementary.

    3.Learning anisotropy directly from features (rather than using dataset-level priors) is a meaningful step forward and aligns well with real-world variability.

    4.Improvements are observed across multiple backbones and tasks (e.g., segmentation and classification), which supports generality rather than task-specific tuning.

    5.The correlation between learned scaling and physical spacing provides evidence that the model captures meaningful geometric structure, rather than acting as a black-box adjustment.

    6.The incremental ablation study convincingly shows that both spatial (RoPE) and semantic (feature modulation) components contribute to the final performance.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The method builds on existing ideas (anisotropic encoding, RoPE extensions, feature gating). The contribution is mainly in how these components are combined and learned, rather than introducing a fundamentally new formulation.

    2.The evaluation would be more convincing with comparisons to newer geometry-aware or hybrid architectures beyond standard baselines and AFPE.

    3.In the classification setting, volumes are resized to a fixed grid. While the method is claimed to adapt to this, the effect of such preprocessing on learned anisotropy is not deeply analyzed.

    4.The added modules introduce overhead, but there is no discussion of runtime, parameter increase, or training stability.

    5.Certain implementation aspects—particularly how anisotropic scaling interacts with channel partitioning in positional encoding—could be described more precisely.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper tackles a relevant and practical problem and proposes a technically sound and generalizable solution. The method is well-designed and shows consistent improvements across tasks and architectures.

    However, the overall contribution is somewhat incremental, and the experimental validation could be strengthened with broader comparisons and additional analysis.

    Overall, the work sits slightly above the acceptance threshold.

  • 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 author has addressed the concern regarding the the mismatch on the paper.



Author Feedback

We thank all reviewers. According to policy, we are not allowed to add new experiments. R1: Q1 (fixed-grid resizing contradicts the critique of resampling): For segmentation, we don’t resample but use native spacings to directly validate LAD’s geometry awareness. For classification, we use resizing to ensure a standardized and fair input pipeline; however, LAD does not aim to recover physical spacing here, but learns task-related morphology via the Anisotropy Learner, as suggested by Fig. 2b. Q2 (global pooling discards texture): This branch is used to estimate anisotropy-related scaling information, not to encode texture details. As it only predicts a per-axis scalar, global pooling is sufficient here. Importantly, the original features are preserved through residual connections around this branch, so texture details are not discarded. Q3 (vs. SE-Net): LAD differs in both the learned information and the modulation target. SE-Net learns generic channel weights from global channel statistics and only modulates feature amplitudes. In contrast, LAD learns instance-specific axis-wise scales from structural profiles, which capture anisotropy-related information that SE-Net does not explicitly model. These learned scales further modulate both anisotropic RoPE’s spatial phase and AAFM’s feature amplitude. Q4 (no baseline using real spacing): AFPE, which uses dataset-level spacing, serves as the closest available geometry-aware PE baseline in Table 1.However, we agree that sample-level real-spacing baselines would provide a more direct comparison, and we will explore them in future work. R2: Q1-2 (Performance across organs and statistical analysis): We performed individual-organ comparisons and observed overall improvements across organs, but these results were not included in the main text due to space limits. Our code and data will be publicly released, making these analyses easy to reproduce and verify. For statistical analysis, Fig. 2 already includes Pearson correlation and Mann-Whitney U tests to support the learned-scale behavior and we plan to explore more statistical analyses in future work. R3: Q1 (contribution mainly combines existing components): While LAD leverages existing techniques, its key contribution is not simply combining modules, but learning an instance-level axis-wise scale prior that jointly drives positional phase via RoPE and feature modulation via AAFM in a plug-and-play manner. This joint design enables LAD to adapt to anisotropic spatial patterns across different backbones, which is difficult for traditional static methods to achieve. Q2-3 (lack of comparison with newer geometry-aware baselines and analysis of resizing effect): For resizing, our experiments suggest that LAD can adapt to both native-spacing and fixed-grid settings. In classification, although fixed-grid resizing is used, LAD can still learn task-related scaling information through the Anisotropy Learner and improves over the compared methods. We agree that more comparisons would further strengthen the evaluation, and we will explore them in the future. Q4 (no runtime, parameter cost, stability analysis): LAD introduces limited overhead, increasing GPU memory and training time by ~10% and FLOPs by ~5% on average. Training stability is supported by zero initialization in the Anisotropy Learner and AAFM, which allows the model to start from an isotropic state; empirically, we did not observe training instability in our experiments. Q5 (anisotropic scaling, channel partitioning unclear): The scale vector is learned by the Anisotropy Learner and produces three values per attention head. Within each head, channels are equally split into three groups, and each group is rotated by the RoPE frequency derived from its corresponding scale. AAFM uses the same per-head scales to generate axis-wise gates aligned with these groups, ensuring that phase and amplitude layouts are geometrically consistent. We will clarify this in the revision.




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 addresses the mismatch between the anisotropic input data and the isotropic input expected by the trained network. This work conducted a solid comparison between various backbones.

    The reviewers have shared concerns about the validity of the proposed design (e.g., global pooling over the features, disease-dependent scaling factor), the correctness of the experiment setting (e.g., the missing evaluation with inputs that have varying resolutions), and the missing comparison to baselines that take real physical spacing into account or geometry-aware architectures.

  • 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 Learnable Anisotropic Deformation (LAD), a plug-and-play framework that adapts 3D Transformer representations to anisotropic medical data by learning instance-specific scaling factors from image features and jointly modulating spatial positional encoding and feature amplitude.

    Reviewers broadly agree that the problem is well-motivated and practically relevant, the mismatch between anisotropic medical data and isotropic model assumptions is a real and under-addressed issue. The dual-pathway design is well thought out, the ablation study convincingly supports both components, and consistent improvements across multiple backbones and tasks demonstrate meaningful generalizability.

    The shared concerns are incremental novelty, missing stronger baselines, and insufficient analyses. On novelty, the contribution is mainly combinatorial rather than fundamentally new, though the rebuttal argues convincingly that the joint instance-level learning of spatial phase and feature amplitude is a meaningful integration beyond simply combining existing modules. On the apparent contradiction between criticizing forced resampling and using fixed-grid resizing in classification, the rebuttal clarifies that the two settings serve different purposes, which R1 and R3 find satisfactory. The rebuttal provides reassuring runtime numbers (~10% memory, ~5% FLOPs) and promises code release for reproducibility.

    Two out of three reviewers accept post-rebuttal and the core contribution is sound. I recommend acceptance.



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.

    There are 2 Accepts after rebuttal, while R1 did not update the final recommendation and was initially borderline negative. Reviewers generally agreed on the importance of the problem and the broad empirical validation across tasks and backbones. The rebuttal clarified the motivation behind the anisotropy learner, its distinction from generic attention mechanisms, and the role of fixed-grid preprocessing. It also addressed concerns regarding efficiency, implementation details, and statistical analysis. While some questions remain regarding comparisons with additional geometry-aware baselines, the overall contribution is well supported. The presentation could also be further improved, including figures, tables, and general formatting. I recommend 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 authors have provided fair rebuttal materials, and 2 reviewers read and adjusted their ratings to accept. Major concern has been justified.



back to top