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

Abstract

Cortical surface analysis is a specialized 3D shape modeling paradigm of fundamental importance in neuroscience. Despite the rapid advancement of pre-trained generic 3D models, their direct application to cortical data is challenged by the highly convoluted topology of the cortex and the presence of fine-scale anatomical features absent in generic 3D objects. To bridge this domain gap, we propose CortexAdapt3D, a parameter-efficient fine-tuning framework that adapts geometric priors from large-scale 3D pre-training to cortical surface analysis. First, to respect intrinsic manifold topology, we adopt a geodesic-aware patch sampling strategy on the cortical icosahedral (ICO) sphere. Second, we inject relative positional bias to capture the fine-grained local geometric dependencies inherent in highly folded cortical morphology. This enables effective fusion of geometric priors from large-scale generic 3D data with domain-specific cortical features. Finally, recognizing that cortical surfaces encode complementary neuroanatomical signals beyond spatial coordinates, we propose an Attribute-aware Spectral Modulation Adapter (ASMA). This spectral-domain adapter leverages attention to dynamically fuse generic 3D priors with domain-specific attributes like sulcal depth and curvature. Extensive experiments across classification, regression, and segmentation tasks demonstrate that CortexAdapt3D consistently outperforms both representative 3D fine-tuning strategies and dedicated cortex-specific architectures. Our findings demonstrate the effectiveness of cortical-specific adaptation for transferring generic 3D priors to the tested infant cortical analysis tasks, and suggest a promising direction for broader cortical representation learning. The code will be available at https://github.com/ladderlab-xjtu/CortexAdapt3D.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3433_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ladderlab-xjtu/CortexAdapt3D

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiKeh_CortexAdapt3D_MICCAI2026,
        author = { Li, Kehan AND Shen, Chen AND Hao, Shuang AND Wang, Fan AND Ma, Jianhua AND Lian, Chunfeng},
        title = { { CortexAdapt3D: Parameter-Efficient Fine-Tuning of General 3D Foundation Models for Cortical Surface Analysis } },
        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

    This work presents CortexAdapt3D, a parameter-efficient fine-tuning framework that adapts pre-trained 3D geometric priors to cortical surface analysis. It introduces geodesic-aware patch construction, relative positional bias, and an attribute-aware spectral modulation adapter to better capture intrinsic geometry and domain-specific cortical features.

  • 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 CortexAdapt3D framework introduces a learnable relative positional bias and a cross attention module in the spectral domain, enabling efficient fine-tuning. The proposed framework is thoroughly evaluated and achieves consistent improvements across classification, regression, and parcellation tasks in cortical surface analysis.

  • 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.

    For the geodesic-aware patch construction, the similar ideas have been proposed in Surface Vision Transformer [1] and Spherical Transformer [2]. The related prior work should be discussed and compared.

    [1] Dahan, Simon, et al. Surface vision transformers: Attention-based modelling applied to cortical analysis. MIDL (2022). [2] Cheng, Jiale, et al. STF: A spherical transformer for versatile cortical surfaces applications. NeuroImage (2025).

  • Please rate the clarity and organization of this paper

    Good

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

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

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

    N/A

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

    N/A

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

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

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

    Relevant previous work should be properly reviewed and compared.

  • 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 have addressed my concerns during rebuttal.



Review #2

  • Please describe the contribution of the paper

    This paper proposes CortexAdapt3D, a parameter-efficient fine-tuning (PEFT) framework that adapts generic 3D point cloud pre-trained models (Point-MAE, PointDif) to cortical surface analysis. The framework introduces three components: (1) geodesic-aware patch construction on an icosahedral sphere to respect cortical topology, (2) learnable relative positional bias (RPB) injection into the pre-trained Transformer’s attention mechanism, and (3) an Attribute-aware Spectral Modulation Adapter (ASMA) that projects features into the graph Laplacian eigenbasis and uses cross-attention conditioned on cortical attributes (sulcal depth, thickness, curvature) to modulate geometric representations. Evaluated on dHCP and BCP, CortexAdapt3D outperforms both fully supervised baselines and existing fine-tuning methods.

  • 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.

    • S1.Important and well-motivated research direction. Bridging the gap between large-scale 3D point cloud pre-training and cortical surface analysis is a timely problem. The paper correctly identifies three key domain mismatches — topology (Euclidean vs. geodesic proximity), positional encoding (absolute vs. relative), and feature richness (coordinates-only vs. derived neuroanatomical attributes) — and proposes a targeted component for each. Figure 1 effectively illustrates the motivation for geodesic patching and attribute incorporation. This systematic decomposition of the domain gap is a clear methodological contribution.

    • S2.Comprehensive multi-task, multi-dataset evaluation with strong results. The paper evaluates on three distinct tasks (preterm classification, cortical parcellation, PMA age regression) across two independent datasets (dHCP, BCP), comparing against three categories of baselines: fully supervised from-scratch models (SphericalUNet, AGConv), full fine-tuning (PointMAE, PointDif), and state-of-the-art PEFT methods (IDPT, DAPT, PointGST). Tables 1-2 show consistent improvements across tasks and backbone choices. Notably, the improvement over full fine-tuning is substantial (Table 2: CortexAdapt3D Dice 89.90/89.96 vs. PointMAE full fine-tuning 75.01), which demonstrates that the domain-specific PEFT design provides genuine regularization benefit, not merely computational savings.

    • S3.Principled domain-specific adapter design via cortical attribute conditioning. The ASMA module’s key insight — conditioning spectral modulation on neuroanatomically meaningful attributes (sulcal depth, cortical thickness, curvature) via cross-attention — is a well-motivated design choice that goes beyond generic PEFT methods. These attributes encode local brain morphology that cannot be inferred from coordinates alone, and their explicit integration provides the model with domain-specific inductive bias. The ablation in Table 3 (comparing with and without attribute conditioning: ACC 91.03 vs. 87.18 for PointDif) supports the claim that attribute awareness contributes meaningfully to performance.

    • S4.Practical relevance for clinical deployment. The framework keeps the pre-trained backbone frozen, training only lightweight modules (RPB, ASMA, task head). This is appealing for clinical settings where labeled cortical data is scarce and computational resources are limited. The ability to leverage large-scale 3D pre-training from non-medical domains for specialized neuroimaging tasks addresses a real practical need.

  • 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.

    • W1.ASMA’s spectral modulation mechanism builds closely on PointGST [4] — the novelty boundary should be clarified. The ASMA module (Sec. 2.4) follows the same spectral projection pipeline as PointGST (Liang et al., IEEE TPAMI 2025): project token features into a graph Laplacian eigenbasis, apply modulation, then project back. The principal difference is the cross-attention conditioning on cortical attributes. While this is a meaningful domain-specific adaptation (supported by the ablation in Table 3), the paper does not clearly delineate what is inherited from PointGST versus what is novel. A side-by-side comparison or explicit discussion would help the reader assess the incremental contribution. The paper’s framing presents ASMA as an entirely new component, which overstates the novelty.

    • W2.No statistical significance testing — some performance gains are marginal. Tables 1-2 report single-point estimates without standard deviations, confidence intervals, or p-values. Some improvements are small: e.g., MAE 0.74 vs. 0.76 (Table 2, PointDif) is a 0.02-week difference. While other gains are more substantial (ACC 91.03 vs. 88.28, Dice 89.96 vs. 87.60), reporting variance over multiple random seeds would strengthen all claims and is expected practice for datasets of this size (hundreds of subjects).

    • W3.Eigenbasis computation U is underspecified. Section 2.4 states “we project features into the spectral domain using the eigen basis of the graph Laplacian,” but does not clarify whether U is computed per-sample (preserving individual geometry but expensive and preventing batching) or on a template mesh (enabling batching but losing subject-specific information). This is a fundamental design choice affecting theoretical soundness. Clarifying this is essential for reproducibility and for understanding the method’s behavior.

    • W4.Generalizability claims exceed experimental evidence. The abstract and conclusion claim CortexAdapt3D establishes “a unified, scalable paradigm.” However, both datasets contain only infant brain data. No experiments on adult cortical data, neurodegeneration, or pathological conditions are presented. The claims should be qualified to match the experimental scope.

  • Please rate the clarity and organization of this paper

    Good

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

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

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

    N/A

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

    1.Clarify the relationship to PointGST [4]: Explicitly state which ASMA components are inherited from PointGST and provide a concise comparison. The ablation row “w/o Attr” in Table 3 appears to approximate PointGST — stating this explicitly would be transparent and informative.

    2.Add variance estimates: Report mean and standard deviation over at least 3 random seeds. For marginal improvements (Table 2, MAE differences of 0.01-0.02), this is especially important.

    3.Extend ablation to parcellation: The BCP parcellation task would reveal each component’s contribution to fine-grained dense prediction more clearly than binary classification.

    4.Specify eigenbasis computation: State whether U is per-sample or template-based, report the number of eigenvectors used (full N or truncated to K), and discuss the trade-off.

    5.Qualify generalizability claims: Restrict claims to infant cortical analysis on the tested datasets.

  • 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?

    CortexAdapt3D addresses a timely problem and proposes a well-structured framework (S1). However, the methodological novelty is limited: ASMA closely follows PointGST [4] with cross-attention conditioning as the primary addition (W1), and the paper does not transparently delineate inherited vs. novel components. The experimental evaluation has notable gaps — no variance estimates despite marginal improvements on some metrics (W2), and an underspecified eigenbasis computation that affects reproducibility and theoretical interpretation (W3). The generalizability claims also exceed the evidence from infant-only datasets (W4). While the consistent multi-task improvements (S2) and the attribute conditioning ablation (S3) are positive, the cumulative weight of novelty and evaluation concerns places this paper marginally below 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.

    I have carefully read the reviews, and the author rebuttal, and respectfully find that my recommendation remains a Reject. I thank the authors for their honest and diligent response, and offer the following reasoning behind this decision.

    First, the technical novelty is incremental rather than clearing the bar typically expected of a MICCAI methodological contribution. As all reviewers and Meta-Reviewer #2 have raised in different forms, the paper’s relationship to prior work is the central concern. ASMA’s spectral projection-modulation-back-projection pipeline is inherited from PointGST (Liang et al., TPAMI 2025), with cortical-attribute cross-attention as the sole novel architectural delta; the geodesic patching also overlaps in spirit with Surface Vision Transformer (Dahan et al., MIDL 2022) and STF (Cheng et al., NeuroImage 2025). The rebuttal helpfully delineates this inheritance, but the underlying contribution remains a narrow adaptation of a single recent prior method.

    Second, the generalizability is, as the authors themselves kindly acknowledge, restricted to infant cortical tasks and has not been validated on adult cohorts. Given that adult cortical analysis is the more common clinical demand, and that few infant-specific methodological innovation has been offered in place of the broader claim, the work’s overall completeness still falls short.

    I am sincerely grateful for the authors’ care, and hope these comments support a strengthened future version of the work.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a parameter-efficient fine-tuning approach for cortical surface analysis using 3D foundation models. Since the cortex has anatomical properties not present in standard 3D pre-training data, such as its characteristic folding patterns, the authors argue that targeted adaptation is needed. To this end, they introduce patch sampling on an icosphere, a relative positional bias, and an Attribute-Aware Spectral Modulation Adapter (ASMA). The method is evaluated on parcellation and age regression tasks, outperforming two fully supervised baselines, full fine-tuning of 3D foundation models, and other parameter-efficient fine-tuning 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.
    • Finding effective ways to adapt large 3D foundation models to cortical surface analysis is an important problem, and the proposed components are well-motivated.
  • 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’s main weakness is a lack of clarity which makes it difficult to fully evaluate the contributions. The following points need to be addressed:

    • ASMA design choice: What are the cortical-specific attributes used as input to ASMA (Page 4, first paragraph)? Additionally, how does ASMA compare to a simpler alternative, such as concatenating the cortical attributes directly to the input coordinates?
    • Spherical U-Net for age regression: It is unclear how a U-Net architecture, which is typically used for segmentation or map prediction tasks, is adapted for age regression, which requires a single scalar output. Are additional layers added after the bottleneck or decoder? Could the model be overfitting due to having too many parameters for this task?
    • The fully fine-tuned models perform surprisingly poorly. It would be helpful to understand why. For instance, how does fully supervised training of PointMAE or PointDif compare?
    • The ablation study compares against spectral domain adapter fine-tuning without cortical attributes, which is an informative baseline. It would be valuable to include this condition in the main Table 2 experiments (parcellation and age regression) as well, unless this is equivalent to PointGST, in which case please clarify
  • 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
    • Patch sampling and input representation (Page 4): The patch sampling procedure is not clearly described. Is the cortical surface registered to a sphere of a given resolution and then remeshed accordingly? It is also unclear whether each triangular face on the icosphere corresponds to one patch, and whether the input consists solely of the three vertex coordinates of that face. If more coordinates are used, does the network receive any connectivity information? Please clarify the full input representation.
    • Relative positional bias (Page 4): It is unclear what B_i,j looks like concretely. Is it a scalar, vector, or matrix? The authors describe it as learnable, yet in Figure 2 the RelPos-Attn block appears to be frozen. Please clarify how and when this component is learned.
    • W_d and W_u in Figure 2 and d_a on Page 5 are not defined. Please introduce all notation before use.
    • It is not stated what data Point-MAE and PointDif were pre-trained on. This context is important for understanding the transfer learning setting.
    • Baselines in Table 2: Please add references for all baseline models and include a measure of variance (e.g., standard deviation) for all reported results.
  • 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 proposed method is interesting and the results are promising, but the paper is currently difficult to evaluate due to insufficient methodological clarity. I would be willing to revise my score upward if the authors can address the points above in their rebuttal.

  • 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 sufficiently addressed my concerns in the rebuttal.



Author Feedback

We thank the reviewers for the constructive comments. We address the main concerns below. [R1,R2,R3] Prior work and relation to PointGST. Our contribution is not geodesic patching as standalone novelty, but adapting generic 3D pre-trained priors to cortical surfaces. For ICO-based cortical surfaces, topology-preserving patches are a natural domain adaptation, since Euclidean KNN can connect points across folds (Fig. 1). The ablation supports this: even with frozen patch embedding, changing only the patch construction improves generic-prior transfer. We will better discuss the cortical transformer works mentioned by R1.ASMA inherits the spectral projection–modulation–back-projection paradigm from PointGST, but differs in two cortical-specific aspects: we omit PointGST’s Euclidean local-basis branch, which does not respect intrinsic cortical topology, and introduce cortical-attribute-conditioned cross-attention. Thus, the ablation baseline is not exactly PointGST, but an empirically close PointGST-like spectral-adapter baseline without attribute conditioning. We will explicitly clarify inherited vs. new components. [R2,R3] Details of ASMA, patch sampling, eigenbasis, and RPB. ASMA uses sulcal depth, curvature, and cortical thickness as cortical attributes. Directly concatenating attributes with coordinates is possible in principle, but would alter the input distribution of the frozen pre-trained encoder and may weaken preserved generic priors; therefore, we inject attributes through a lightweight spectral adapter. All cortical surfaces are remeshed to ICO5 topology with 10,242 vertices and 20,480 faces. We construct 320 geodesic triangular patches; each patch contains 64 ICO5 faces and 45 vertices, not only three triangle corners. The input includes coordinates and cortical attributes. Connectivity is not explicitly fed into the backbone, but is used for topology-preserving patch assignment. The Laplacian eigenbasis U is computed per sample from the patch centers. Since N=320, using the full eigenbasis without truncation is computationally manageable. For RPB, B_ij is a scalar attention-logit bias for token pair (i,j), forming a bias matrix aligned with QK⊤. The pre-trained attention weights are frozen; only the added RPB MLP, computed from patch-center coordinates, is learnable during PEFT. We will define all symbols before use, including W_d/W_u as adapter down-/up-projection matrices and d_a as the adapter embedding dimension. [R3] Comparison methods and transfer setting. For age regression, we used the spherical-CNN encoder with global pooling and an MLP regression head, not the full U-Net decoder; we will rename it as spherical-CNN to avoid ambiguity. Point-MAE and PointDif are mainly pre-trained on ShapeNet, a synthetic object-level dataset with about 50K–57K shapes from 55 categories, rather than cortical surfaces. This domain gap and limited labeled infant data may make full fine-tuning prone to overfitting/catastrophic forgetting, while PEFT better preserves generic 3D priors and learns cortical-specific adaptation. [R2,R3] Statistical robustness, ablation scope, and claims. We initially reported point estimates due to space limits, as many key improvements are large. Nevertheless, we agree that statistical robustness is important for marginal differences. We will therefore add missing references and report mean±std and p-values from existing results. Our ablation was intentionally conducted on dHCP classification as a compact and sensitive setting to isolate component effects under the same transfer protocol. For dense prediction, the main BCP parcellation results already show a substantial gain over PointGST, with ACC improving from 71.14 to 91.91.We will restrict broader claims to the tested infant cortical tasks, leave adult/pathological cortical analysis as future work. [R1, R2, R3]-Reproducibility For sure, We will release the code after acceptance.



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 authors should focus on addressing the lack of clarity and detail regarding the methodological contributions, particularly the Attribute-Aware Spectral Modulation Adapter (ASMA) and its relationship to prior work like PointGST. They must explicitly clarify the eigenbasis computation and whether it’s per-sample or template-based. This explanation impacts reproducibility and theoretical soundness. The other main concern is the generalizability claims which is limited to infant brain data. Reporting standard deviations for results will boost the performance claims especially for marginal gains.

  • 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 clarified the reviewers, specified the patch sampling and eigenbasis computations, and restricted their claims to infant data. Two reviewers upgraded their scores to accept after these clarifications, demonstrating that the primary concerns regarding reproducibility and methodological details were resolved. The paper is recommended for 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.

    This paper proposes CortexAdapt3D, a parameter-efficient fine-tuning framework that adapts generic 3D point-cloud foundation models to cortical surface analysis through geodesic-aware patch construction, learnable relative positional bias, and an attribute-aware spectral modulation adapter. The reviewers agree that adapting large 3D pre-trained priors to cortical surfaces is a timely and well-motivated problem, and that the multi-task, multi-dataset evaluation shows consistent improvements over supervised, full fine-tuning, and prior PEFT baselines. The rebuttal addressed the principal concerns in concrete terms, clarifying the per-sample eigenbasis computation, the patch construction and input representation, the relative positional bias and previously undefined notation, and the boundary between inherited and newly introduced adapter components. The main remaining reservation concerns novelty, as the contribution is best understood as a domain-specific adaptation of an existing spectral adapter paradigm; the authors are encouraged to retain the transparent inherited-versus-novel framing and to qualify the generalizability claims to the infant cortical setting actually evaluated.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    This submission proposes a cortical surface analysis using fine-tuning of pre-trained models. R1 and R3 appreciated the multi-task performance. R2 noted the incremental nature of the method and heavy reliance on the existing PointGST. The rebuttal attempted clarifying the novelty of the components but the core concern on novelty remains. As is, the consensus appears to converge towards a lack of substantial technical innovation. For all these reasons, and with respect to the other submissions, the recommendtion is towards Rejection,



back to top