Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Normative modeling characterizes heterogeneous brain pathology by quantifying individual deviations from a healthy reference population. Variational autoencoders (VAEs) are widely used for this purpose, but latent-space irregularities can compromise distance-based deviation measures and interpretability. We propose SIGMA-VAE, a Sketched Isotropic Gaussian Multi-view Autoencoder for robust and interpretable normative modeling. SIGMA-VAE combines (i) Sketched Isotropic Gaussian Regularization to enforce a well-formed latent geometry with more reliable Euclidean anomaly scores, (ii) learnable sparsity via L0 regularization for compact and interpretable representations, and (iii) an uncertainty-aware $W$-score that weights deviations by posterior variance to mitigate false positives due to image ambiguity and scanner confounds. We evaluate SIGMA-VAE on over 40,000 structural brain scans from 328 scanners, demonstrating improved sensitivity to individual-level anomalies, biologically meaningful deviations, and robust generalization to unseen scanners. In downstream discrimination tasks, SIGMA-VAE achieves strong performance for Alzheimer’s disease, with AUCs of 0.91 ±0.01(n=207) on a held-out test set and 0.87 ± 0.01 (n=585) on an external test set acquired from previously unseen scanners. These results demonstrate that SIGMA-VAE enables scalable, interpretable, and robust normative modeling across views and at scale.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3072_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/mariam186/SIGMA-VAE

Link to the Dataset(s)

ABIDE I/II, ADHD-200, COBRE, CoRR/NKI-RS: http://fcon_1000.projects.nitrc.org/indi/ HCP: https://www.humanconnectome.org/ Cam-CAN: https://www.cam-can.org/ HBN: https://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network/ PNC: dbGaP phs000607 OpenNeuro+TCP(ds005237): https://openneuro.org/ AIBL/PPMI: https://ida.loni.usc.edu/ OASIS-3: https://www.oasis-brains.org/ SRPBS-TS: syn22317076

BibTex

@InProceedings{ZabMar_SIGMAVAE_MICCAI2026,
        author = { Zabihi, Mariam AND Dai, Sicheng AND Schmidt, Andre AND Preller, Katrin AND Watrous, Andrew AND Bornemann, Klaus AND Cole, James AND Marquand, Andre AND Yang, Yanwu AND Wolfers, Thomas},
        title = { { SIGMA-VAE: Sketched Isotropic Gaussian Multi-View Autoencoder for Normative Modeling } },
        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

    The paper presents a novel methodological framework for normative modeling of brain MRI data. The authors address limitations of prior VAE-based approaches by introducing three key contributions: (1) Sketched Isotropic Gaussian Regularization (SIGReg) within the VAE training objective, promoting a compact, hole-free latent geometry that supports Euclidean deviation scoring; (2) L0 regularization via learnable Hard Concrete gates, enforcing latent sparsity by deactivating uninformative dimensions; and (3) an uncertainty-aware W-score derived from the encoder’s per-dimension, per-subject means and variances, providing calibrated subject-level deviation estimates. The methodological apparatus is well-motivated, and the authors support their claims with a robust validation pipeline covering both in- and out-of-distribution test sets. Ablation experiments support the individual contribution of each component, though no direct comparison with other VAE-based normative models from the literature is provided.

  • 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.
    • Theoretical grounded claims towards attacking VAEs limitations: The authors properly explores, discusses and provide references for the reader to better understand VAE limitations, as: chance of raising an incoherente global latent geometry (regions with roles or not sufficient represented); lack of interpretability of the learned latent representation; confusion of healthy variations given demographic factors (. e. g, age and sex) with patologies;
    • Proposer use of literature to tackle VAEs limitations: Authors adapt to their VAE techniques as SigREG (recently proposed in the LeJEPA article), L0 regularization to select important and interpretable latent dimensions, and an uncertainty-aware W-score based on epistemic uncertainty. The use of SigREG together with L0 regularization enables the authors to enable reliable distance-based deviation measures;
    • Interpretability: By individually disturbing each W-score, the authors also enable the visualization of the anatomical subsystem that the W-score represents;
    • Robust validation: Authors employ strategies to evaluate in- and out-of-distribution test sets, isolating the effect of testing in samples of the same test set or scanner. They also isolate the effects of age and sex on evaluation, which is also positive, as they employ conditioning with age and sex. Lastly, W-score Q-Q plots show that the learned scores follow a normal distribution centered at 0 and with close to unit variance.

    —— MORE DETAILS —-

    • **Theoretically grounded motivation. ** The authors provide a well-structured and referenced critique of standard VAE-based normative models, identifying three concrete limitations: the risk of an incoherent global latent geometry with redundant or sparsely populated (hole-prone) regions; limited interpretability of the learned latent representation; and the conflation of healthy demographic variation (age and sex) with pathological deviations. This framing effectively justifies each of the proposed contributions;
    • **Principled use of recent literature to address VAE limitations. ** The authors adapt three complementary techniques to directly tackle the identified limitations: SIGReg, recently proposed in the LeJEPA framework, is incorporated into the VAE training objective to promote a compact, hole-free isotropic Gaussian latent geometry that supports reliable Euclidean deviation scoring; L0 regularization via learnable Hard Concrete gates enforces latent sparsity, deactivating uninformative dimensions and improving interpretability; and an uncertainty-aware W-score is derived from the encoder’s per-dimension means and variances — grounding deviation estimates in aleatoric uncertainty. The combination of SIGReg and L0 regularization is particularly well motivated, as a compact and sparse latent space is a prerequisite for distance-based deviation measures to be geometrically meaningful;
    • **Interpretability of latent dimensions. ** By independently perturbing each latent dimension and observing the resulting changes in reconstructed anatomy, the authors enable visualization of the anatomical subsystem that each W-score dimension represents. It is a meaningful differentiator from most VAE-based normative modeling works in the literature, which typically treat the latent space as a black box;
    • **Robust and thorough validation pipeline. ** The authors employ a rigorous evaluation strategy across multiple axes. In- and out-of-distribution test sets are assessed separately, isolating performance on the same-site data from generalization to held-out scanner data. The model’s conditioning on age and sex is appropriately reflected in the evaluation, preventing demographic variation from confounding deviation scores. Finally, W-score Q-Q plots demonstrate that the learned scores follow a normal distribution centered at zero with near-unit variance across healthy subjects, providing empirical support for the Euclidean scoring assumption that underlies the deviation metric.
  • 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.
    • Use of PoE: a review of one of the cited literature (Multi-modal Variational Autoencoders for normative modelling across multiple imaging modalities, Lawry Aguila et al. ) reveals a discussion where the PoE treating all experts as equally credible, uniform contribution from every modality (in the paper under review here the contribution of erery view), and in practice noyse, and complexity may break this assumption. In Lawry Aguila et al. , they investigate using Mixture-of-Experts (MoE) or generalized Product-of-Experts (gPoE). As the authors cited Lawry Aguila et al. , it would be valuable to add one phrase discussing this limitation of PoE and how they see it;
    • Encoder/Decoder pairs for each view: authors choose to train a pair for each view (Cortical Thickness, Subcortical Volume, Surface Area), and state that Multi-View fusion formulation naturally handles missing views, but no tests are made towards it. In the future, as they expand across multiple modalities, how will they tackle it? Other works handle views within the same encoder-decoder pair but employ a separate encoder for each modality. Maybe a short phrase discussing this choice would be valuable;
    • Encoder architecture: Incomplete feature attention block description. It employs a single layer? How are attention weights computed?

    —– MORE DETAILS ——

    • **Use of Product-of-Experts for view fusion. ** The authors adopt a standard Product-of-Experts (PoE) formulation to aggregate the per-view encoder posteriors into a joint latent representation. As discussed in Lawry Aguila et al. (2023), a work cited by the authors, PoE treats all experts as equally credible, assuming uniform contributions from all views. In practice, differences in noise levels, feature complexity, and information content across views can violate this assumption, potentially biasing the joint posterior towards overconfident but miscalibrated experts. While this concern is somewhat attenuated here since all three views derive from the same acquisition, it will become increasingly relevant as the authors expand their framework to genuinely distinct modalities, as stated in their future work. Given that the authors cite Lawry Aguila et al. , who explicitly address this limitation via Mixture-of-Experts and generalized Product-of-Experts formulations, a brief discussion of why standard PoE was preferred over these alternatives would strengthen the paper.

    • **Untested claim regarding missing view robustness. ** The authors state that their multi-view fusion formulation naturally handles missing views, a known theoretical property of PoE, which allows excluding absent experts from the product. However, no empirical evaluation of this capability is provided. Given that missing data is a common practical challenge in clinical neuroimaging, this claim should either be validated experimentally or explicitly hedged as a theoretical property rather than a demonstrated one. Furthermore, as the authors plan to extend their framework to additional modalities, a concrete strategy for handling missing modalities will become essential and warrants discussion.

    • **Incomplete description of the feature attention block. ** The encoder architecture includes a feature attention block before the MLP backbone, but its implementation is insufficiently described for full reproducibility. Specifically, it is unclear whether the block consists of a single linear layer or a multi-layer MLP with a bottleneck; how the attention weights are computed and normalized (sigmoid per feature vs. softmax across features). Given that the MLP backbone dimensions [512, 256, 128] and the 32-dimensional raw latent space are clearly specified, the same level of detail should be provided for the attention block to ensure reproducibility;

    • **Incomplete references. ** The manuscript contains several instances of unresolved references appearing as [?]. While some of these may reflect intentional anonymization of self-citations for double-blind review — a common and acceptable practice — this does not account for all cases. Notably, [?] appears in contexts where already correctly cited external methods are referenced, including SIGReg, which is cited elsewhere in the manuscript with a proper reference. Authors should ensure that all non-self-referential citations are correctly resolved in the final submission, as unresolved references to key methodological sources undermine reproducibility and make it difficult for readers to trace the theoretical foundations of the work.

  • 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 provide sufficient information for 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?

    My recommendation for acceptance, independent of rebuttal, is based on the observation that the identified weaknesses are minor, concerning discussion points and reproducibility details rather than fundamental methodological flaws. Addressing them in a revised manuscript would improve clarity and completeness, but they do not undermine the core contributions or the validity of the reported results.

    The paper is theoretically well-motivated, technically grounded, and presents a coherent methodological system in which SIGReg, L0 sparsity, PoE fusion, and the uncertainty-aware W-score form mutually reinforcing components. The validation pipeline is robust, covering both in- and out-of-distribution settings with appropriate controls for demographic confounders. The work makes a meaningful contribution to the normative modeling literature and provides a strong foundation for future extensions towards multi-modality and multi-view frameworks.

  • 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

    The paper presents a variational autoencoder framework for normative modeling of neuroimaging data. Three main extension have been brought to the traditional VAE: (1) Sketched Isotropic Gaussian Regularization (SIGReg) to enforce a well-behaved latent space geometry, (2) L0-based sparsity for interpretable latent representations, and (3) an uncertainty-aware W-score for robust anomaly detection.

    The method is validated on a very large multi-site dataset (~40k scans), demonstrating improved calibration, interpretability, and generalization compared to standard VAE-based approaches.

  • 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.

    • improvement to the latent space geometry using characteristic function matching (SIGReg) - Standard VAEs enforce KL divergence per sample, not globally. The proposed work explicitly targets global latent geometry, enabling valid Euclidean distance-based anomaly scores.

    • interpretability using L1 - the latent space is reduced from 32 to 5 using L1 constraints. Is interesting how this sparsity enforces latent space dimensions to correspond to distinct anatomical regions.

    • a large dataset is used for validation with explicit out-of-distribution evaluation

    • the ablation study is well designed as well and demonstrates that neither sparsity nor geometry alone is sufficient

  • 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 ablation study is well done but the method could have been compared with other normative models to establish a baseline ex. PCNToolkit Wolfers, T., Rokicki, J., Alnæs, D., Berthet, P., Agartz, I., Kia, S. M., Kaufmann, T., Zabihi, M., Moberget, T., Melle, I., Beckmann, C. F., Andreassen, O. A., Marquand, A. F., & Westlye, L. T. (n.d.). Replicating extensive brain structural heterogeneity in individuals with schizophrenia and bipolar disorder. Human Brain Mapping

    Also: basic disentangling AE like betaVAE multi-model VAE for normative modeling https://conferences.miccai.org/2023/papers/445-Paper2360.html

    • the evaluation is focused on AUC that is a classification score. For a normative model, the focus should be anatomical and clinical interpretability (ex. correlation with clinical scores)

    • the covariates (sex, age, site …) were conditioned on the model. This does not mean that the rest of the latent space is disentangled from those covariates. The authors could consider a disentangling strategy to enforce that.

    • Reproducibility is moderate but not strong: The authors provide details on the architecture and training but the code is not available. The SIGReg implementation details are not fully specified. While the datasets are publicly available the details on data processing (e.g., normalization, harmonization) are unclear

  • 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 provide sufficient information for 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
    • pg2 - SIGReg [?] , L0 regularization [?] and more - missing references
    • fig 2 left - text is too small
    • the perturbation analysis - could be strengthened with quantitative validation (e.g., overlap with known disease patterns)
    • The uncertainty-aware W-score is interesting—could you compare it against standard Mahalanobis distance ?
  • 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 found the paper well written and the contribution sufficiently novel and practically very useful - in particular the contribution on modeling the latent space geometry and he L0 regularization. However there are some limitations, especially related to comparison to baseline and other generative normative models. Careful validation of the method is done using a large dataset from multiple sites.

  • 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 #3

  • Please describe the contribution of the paper

    This paper proposes SIGMA-VAE, a robust and interpretable deep generative framework for brain imaging normative modeling. Its core contributions include spectral regularization (SIGReg) for reliable Euclidean anomaly scoring , $L_{0}$ regularization for discovering sparse and biologically meaningful latent dimensions , and an uncertainty-aware W-score to reduce false positives. Evaluated on over 40,000 scans, SIGMA-VAE demonstrates exceptional sensitivity to individual anomalies and strong generalization to unseen scanners in clinical tasks like Alzheimer’s discrimination.

  • 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.

    Through $L_{0}$ sparsity and perturbation analysis, SIGMA-VAE uniquely maps latent dimensions to distinct, biologically meaningful neuroanatomical subsystems.

    Rigorously validated on over 40,000 scans, the model demonstrates stable, out-of-distribution generalization to completely unseen clinical environments and scanners.

    The proposed uncertainty-aware W-score effectively mitigates false positives driven by inherent scanner noise and image ambiguity.

  • 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 evaluation relies solely on self-ablations, completely lacking comparisons against established, state-of-the-art normative modeling frameworks.

    The explicit omission of standard reconstruction metrics (e.g., MSE) prevents the verification of the model’s baseline feature synthesis capability and data fidelity across different regularization strategies.

    Despite claims of processing high-dimensional data, the model heavily depends on pre-extracted 1D tabular features rather than raw 3D structural scans.

    The manuscript contains multiple uncompiled citation tags (rendered as [?]) in critical methodological sections that require immediate correction.

  • Please rate the clarity and organization of this paper

    Poor

  • 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?

    The manuscript presents several critical weaknesses that undermine its overall impact. First, the evaluation relies exclusively on self-ablations, entirely lacking comparisons against established state-of-the-art normative modeling baselines. Second, the explicit omission of standard generative metrics (such as Mean Squared Error) prevents the verification of the model’s fundamental data reconstruction fidelity. Third, the methodological contribution is primarily integrative, synthesizing established machine learning techniques instead of proposing a fundamentally new generative framework. Furthermore, despite claims of scaling to high-dimensional neuroimaging data, the model heavily depends on pre-extracted 1D tabular features rather than raw 3D structural scans. Finally, the presence of multiple unresolved citation tags ([?]) in critical methodological sections reflects poor manuscript preparation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    N/A



Author Feedback

We thank all reviewers (r) and Meta-Reviewer (MR), for their feedback. Following the MR’s guidance, we address the core concerns below. 1- Evaluation Design: Ablations vs. External SOTA (R2, R3, MR) Reviewers suggest comparisons against PCNtoolkit or β-VAE. These frameworks do not address the structural failures SIGMA-VAE targets. While PCNtoolkit is the current field standard, it models features univariately; its per-region z-scores cannot be aggregated into valid multivariate anomaly scores without the geometrically valid latent space this work provides. Univariate methods lack the multivariate geometry to resolve the calibration failures inherent in current normative modelling. On the other hand, β-VAE based approaches, including Lawry Aguila et al. (2023), inherit the standard VAE limitation: per-sample KL regularization does not constrain the aggregated posterior, leaving no guarantee of latent isotropy or hole-freeness required for calibrated Euclidean deviation scoring. The calibration failure of the KL-baseline makes this concrete: even a multivariate latent model miscalibrates by over-retaining 15.5% of healthy controls at the 80% threshold. We therefore adopted a controlled ablation against the standard β=1 VAE baseline — the appropriate experimental unit to isolate each component’s contribution and demonstrate their synergy. The calibration result (Fig. 2B) directly addresses this concern: only the full SIGMA-VAE resolves the miscalibration that persists across all ablated variants. To our knowledge, this is the first normative framework to jointly guarantee latent isotropy, sparsity, and calibrated deviation scoring. 2- Reconstruction metrics (R3, MR) Reconstruction loss was monitored throughout training and remains stable across all ablation variants, confirming that L0 sparsity prunes uninformative dimensions without degrading generative integrity (avg. MSE ~0.35). We will report the reconstruction loss weight and MSE range across variants in the camera-ready version to confirm that reconstruction quality remained controlled. That said, minimizing MSE is not the objective here: VAE training involves an explicit tradeoff between reconstruction fidelity and latent-space structure, and the model with the lowest MSE is not necessarily the best model for anomaly detection. The clinically relevant question is whether latent deviations are properly scaled for inference, which calibration error (Fig. 2B) directly measures, and where the full SIGMA-VAE demonstrably outperforms all variants. 3- Tabular features and large-scale validation (R3) FreeSurfer-derived tabular features are the field standard for large-scale normative studies (n=40,000+). Beginning with tabular features allowed us to validate the framework’s geometric and calibration properties in a controlled, interpretable setting; the architecture imposes no restriction on input dimensionality. Tested on data from entirely unseen scanners, SIGMA-VAE’s consistent W-scores demonstrated scanner-robustness that smaller benchmarks did not match, setting a standard for models that generalise across acquisition environments. 4-Citation tags (all reviewers, MR) We emphasize that the bibliography is complete; the [?] tags were a technical resolution error during the final anonymization pass: L0 (ref. [14], Louizos et al. 2018) and PoE (ref. [16], Hinton 2002) will be corrected in the camera-ready version. Minor points (R1, R2)

  • Feature attention: Implemented via channel-wise sigmoid gating.
  • PoE/Missing views: Views originate from the same acquisition, attenuating the equal-credibility concern; the missing-view claim refers to an inherent PoE property.
  • Disentanglement: Age/sex enter via conditioning c; low latent-demographic correlations (mean |r|=0.035, Sec. 3.2) confirm variation is not absorbed into the latent space.
  • Interpretability: Validated via perturbation analysis (Fig. 3), mapping dimensions to neuroanatomical subsystems.



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 are mixed, and I tend to agree with Reviewer 3.The article combines existing methods and ideas in an interesting and potentially useful way. However, similarly to Reviewer 3, I have several major concerns that should be addressed in the rebuttal.

    In particular, the authors should explain why the evaluation relies solely on self-ablations, without comparisons against established state-of-the-art normative modeling frameworks, and why they did not assess decoder performance using reconstruction metrics. Furthermore, the article cannot be accepted in its current form, as it contains multiple uncompiled citation tags (i.e., [?]).

    Please do not add new experiments or results in your rebuttal, but rather motivate and justify your choices.

  • 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 reviewers’ concerns were convincingly addressed in the rebuttal. I therefore 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.

    The study presents sufficient technical novelty. The validation is particularly strong. All reviewers converge towards 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.

    Please fix remaining citation errors.

    Reviewers agree on acceptance.



back to top