Abstract

Analyzing hippocampal structure is important for characterizing localized geometry that is poorly captured by volumetric summaries. Yet existing surface reconstruction pipelines often suppress the hippocampal sulcus—a thin internal indentation—because it is under-resolved by partial-volume effects or introduce incorrect topology. Deformation based methods have a risk of self-intersections. We present a sulcus-aware, topology-consistent surface reconstruction framework that anchors subjects to a sulcus-preserving template and learns a smooth template-to-subject deformation to transfer the surface with vertex-wise correspondence. Sulcus masks are used only in training to encourage sulcal alignment in a training-time auxiliary manner. Inference hence uses only standard structural volume and hippocampus masks without requiring sulcus masks. Experimental results show improved sulcus fidelity while maintaining competitive overall boundary agreement and stable meshes with minimal geometric failures. The resulting topology-consistent and sulcus-preserved surfaces enable surface-based sulcus morphometry; in a pilot downstream task, we show the proposed method captures sulcus-related surface statistics in cognitively normal (CN) and Alzheimer’s disease (AD) cohorts. The software is available at https://github.com/Shape-Lab/Sulcus-Aware-Hippo-Surface.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1315_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Shape-Lab/Sulcus-Aware-Hippo-Surface

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KimYuy_SulcusAware_MICCAI2026,
        author = { Kim, Yuyeon AND Kim, Banseok AND Son, Jiwon AND Joo, Mingyu AND Kim, Kangmin AND Ha, Minji AND Kim, Minah AND Lyu, Ilwoo},
        title = { { Sulcus-Aware Hippocampal Surface Modeling with Training-Time Sulcus-Guided Learning } },
        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

    The authors propose a method for hippocampal surface modeling that takes into account the hippocampal sulcus in the image and shape model. They construct a sulcus-aware template by iterative registration by incorporating a signed distance field term. The sulcal indentation is computed from the difference of the sulcal mask from the hippocampus shape mask. In order to learn the deformation field, they use two complementary methods, i) compute a low-resolution sulcus probability map for voxel-wise soft supervision, and ii) extract a boundary point cloud from the sulcus masks used only to compute surface-distance penalty (using the Chamfer distance) against the deformed template surface. The template-to-subject deformation is represetned by a diffeormophic transformation by integrating stationary velocity fields. The template based deformation approach guarantees topology consistency across registrations. In the experimental results, the authors show a pilot study of AD vs CN group differences.

  • 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.

    Few papers have explicitly tackled sulcal-based modelling in hippocampal shape analysis. Thus this paper highlights a unique and commonly overlooked problem.

    In formulating this problem, the authors have generated a dataset of 0.33 mm isotropic sulcus masks, intially automated and then refined manually by hippocampal experts. This is a valuable dataset by itself, used for training by the authors but may also be generally useful to the community.

    The authors present a thorough evaluation of their idea across multiple methods.

    The use of signed distance fields instead of binary masks also ensures topological correctness.

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

    While the paper primarily tackles sulcus indentation aware modeling, it does not explicitly visualize the indentation. This is a major weakness. Figure 2(a) shows small lateral-view thumbnails with yellow arrows pointing to surface curvature variations, but the hippocampal sulcus is absent from these views. Thus it was harder to visualize what exactly was modeled and represented here.

    The sulcal mask did not explicitly directly feature in the mathematical formulation. The sulcus masks are defined in terms of soft occupancy maps, which in turn are not defined explicitly. The authors refer to the paper by Hwang et al. but omit the definition of the sulcus map from this paper. Instead they just denote these maps using the symbol P_S and P_T. Similarly the boundary point cloud B_S is not formally defined. The authors state that these maps are obtained using the block-averaging operator, but this operator is not defined either. This made the paper difficult to follow.

    The sulcus shape is not explicitly modeled. The main geometric features of a sulcus are it’s depth, curvature, concave/convex nature etc. are not directly represented in the formulation. One would expect a sulcus-aware method to be represented by explicit geometric features such as sulcal depth fields, curvature-based losses, or medial axis constraints that represent the shape information instead of global appearance based measures.

    Since the formulation of the occupancy maps was not specified, it was not clear if the signed distance fields were only used for the purpse of extraction of point clouds, or was it also passed through the alignment network?

    There were several weaknesses in the the experimental results.

    in Figure 3(b), the SPHARM-PDM representation computes alignment using a spherical harmonic parameterization while the proposed method uses template deformation. It is not clear what is being compared when calculating group differences. Is it the magnitude of the deformation field or is it vertex wise displacement to the template?

    Why is the ODC+smooth that yields a better hippocampal distance (Table 1) absent from the results in Figure 3(b)?

    An ablation study that includes the proposed method without the sulcus-aware modeling is missing. This is because for experimental purposes, a sulcus-aware template is already constructed along with the sulcal indentation information. Thus it is not clear how much gain is obtained from the sulcus-aware template itself and whether incorporating sulcal constraints later during hte semi-supervised alignment stage does make a significant difference. Thus it was not clear whether the sulcus mask supervision during training actually did anything. Similarly voxel occupancy maps or the boundary extraction step could also undergo ablation to determine the performance of the proposed method.

    Important papers such as HippUnfold (DeKraker, J. , Haast, R. A. , Yousif, M. D. , Karat, B. , Lau, J. C. , Köhler, S. , & Khan, A. R. (2022). Automated hippocampal unfolding for morphometry and subfield segmentation with HippUnfold. elife, 11, e77945.) are not cited. While subfield segmentation is out of scope of the proposed work, the above paper does explicitly model the SRLM (stratum radiatum lacunosum moleculare) landmarks for the hippocampal sulcus. Thus it will be useful to compare the proposed method with their work.

  • 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

    N/A

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

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    The paper tackles an important problem in the field. There has been a general lack of extensive studies that have undertaken hippocampal modeling with sulcal-aware methods. Thus the paper tries to address an important gap. However, aspects of the sulcal representation are not adequately described as that’s one of the main contributions of the paper. Further the reliance on masks that capture the global features instead of local features such as sulcal depth, curvature, geometry etc, weakens the explicit shape constraints that are important for hippocampal modeling. The experimental results have limitations in terms of lack of ablation studies, while the statistical results were unconvincing.

  • 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 author’s response clarified my doubts and explained certain methodological details.



Review #2

  • Please describe the contribution of the paper
    • A sulcus-aware hippocampal surface reconstruction framework that preserves the thin hippocampal sulcus — typically washed out by mask-based extraction and smoothing — while enforcing a fixed topology.
    • A template-based diffeomorphic deformation that jointly warps a sulcus-preserving template volume and mesh to each subject, yielding topology-consistent surfaces with vertex-wise correspondence and propagating the template sulcus to obtain subject-specific sulci.
    • Experiement results showing improved sulcus fidelity and stable meshes with competitive boundary agreement, plus a pilot downstream application to surface-based sulcus morphometry in CN and AD cohorts.
  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper tackles a well-motivated and technically challenging problem: preserving the hippocampal sulcus in surface reconstruction. The sulcus is narrow, highly susceptible to partial-volume effects, and routinely washed out by conventional mask-based pipelines, yet it carries morphometric signal relevant to aging and Alzheimer’s disease. Addressing this gap is of clear interest to dementia researchers.
    • The template-based diffeomorphic deformation design is well suited to the problem, simultaneously delivering fixed topology, vertex-wise inter-subject correspondence, and stable meshes with minimal self-intersections — directly addressing a known limitation of deformation-based surface methods.
    • The evaluation is comprehensive, combining surface-distance-based measures, mesh quality metrics, and a pilot downstream morphometry analysis in CN and AD cohorts, and demonstrates improved sulcus fidelity over state-of-the-art baselines such as SPHARM-PDM.
  • 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.
    • There is a concern that the method may attenuate clinically meaningful disease effects to fit the sulcus. In Figure 3, disease effects in the hippocampal head appear diminished relative to SPHARM-PDM and prior literature, where AD-related atrophy is consistently reported across the entire hippocampus. This raises the possibility that the template-to-subject deformation establishes noisy correspondences in the hippocampal head as a by-product of prioritizing sulcal alignment.
    • The “semi-supervised” framing is unclear. Semi-supervised learning typically means training with a small amount of labeled data alongside a large volume of unlabeled data. If only a subset of subjects have sulcus annotations, the number of labeled versus unlabeled subjects and how they are combined during training should be explicitly reported.
    • Important neural network details are missing, including the architecture and network hyperparameters, which limits reproducibility.
  • 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
    • Figure 2(b): the meaning of the color map should be clarified. If it represents the Jacobian determinant of the deformation field, the striped pattern observed in the hippocampal head is concerning, as it may indicate that the proposed method establishes noisy or locally unstable correspondences in this region — potentially as a side effect of prioritizing sulcal alignment.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    The paper addresses a well-motivated problem and the template-based diffeomorphic design delivers topology-consistent surfaces with improved sulcus fidelity. However, the possibility that prioritizing sulcal alignment introduces noisy correspondences in the hippocampal head and attenuates well-established AD effects is a significant concern, compounded by the unclear semi-supervised framing and missing network details. A convincing rebuttal on these points could move the paper above the acceptance threshold.

  • 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 successfully addressed all my concerns in the rebuttal.



Review #3

  • Please describe the contribution of the paper

    The authors present a hippocampus surface reconstruction pipeline that is sulcus-aware, improving the geometric accuracy of hippocampal surface analyses. First, the authors construct a sulcus-aware template consisting of an intensity image, hippocampus soft mask, and sulcus soft mask defined on a 64×64×64 voxel grid. A spatial transform is then learned (VoxelMorph) using supervision from the hippocampal SDF and MRI, and applied to the template surface generated via marching cubes. During training, sulcal labels are used to supervise the spatial transformation, upsampled to increase their gradient weighting, along with the transformed boundary point cloud. The method is evaluated on 1mm isotropic T1-weighted MRI and compared against SPHARM-PDM and PTSD as baseline methods, as well as generic reconstruction approaches, using 120 cases for training and 15 for evaluation (across both hemispheres). The authors also conduct a clinical study demonstrating that their method produces more sulcus-localized statistical clusters than both baselines when comparing AD and CN groups.

  • 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.

    Clinical evaluation: The authors perform a preliminary clinical study, comparing vertex-wise analyses comparing AD-CN patient groups on surfaces reconstructed using different methods. While these results need further validation, they illustrate the importance of handing surface geometry on downstream analyses.

  • 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 reconstructed mesh is dependent on the quality of the sulcus-aware template
    • While the proposed method minimizes topological errors, the registration does not guarantee no self-intersections.
    • Fig 2c, the plots are missing axes labels on the x-axis.
  • 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

    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?

    The authors address a clinically meaningful problem of reconstructing hippocampal sulcal geometry. The proposed framework is well-motivated and validated against established baselines, demonstrating improved sulcal fidelity without sacrificing overall geometric accuracy. The downstream morphometry result further confirms that these improvements translate into meaningful clinical findings.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Author Feedback

We appreciate the reviewers for providing insightful comments that help us improve the quality of our paper. Please see our responses below.

Sulcus representation and equations (R1). Geometric descriptors such as sulcal depth, curvature, and medial axis may benefit sulcus-aware modeling. However, reliable extraction of these features typically requires an explicit surface mesh generated by a specific surface reconstruction procedure. As a result, our model may be biased toward reproducing the characteristics of that reconstruction protocol. We instead use expert-validated high-resolution sulcus labels and convert them into two explicit supervision signals: sulcus occupancy maps and sulcal boundary points. The sulcus occupancy maps (P_S) and (P_T) are obtained by applying standard 3X3X3 average pooling (block averaging in Sec. 2.4) to the binary sulcus masks (L_S) and (L_T). The boundary points (B_S) are extracted from the SDF of the sulcus-aware mask (Y~), whereas the alignment network receives the SDF of the whole hippocampal mask (Y). Thus, Y~ is used only for boundary point extraction and is not fed into the network. We will explicitly state their math notations in Sec. 2.4. Necessity of the sulcus-aware template and sulcus-guided components (R1, R3). Since subject meshes deform from the template, its quality is central (Sec. 2.2). During method development, we found that a simple template often failed to preserve sulcal folding as smoothness and mesh-regularity constraints tended to suppress the indentation. This motivates our unique sulcus-aware template construction. Also, as discussed in Sec. 1, the sulcus is often under-resolved at typical clinical resolution (~1 mm^3). Thus, we use a “soft” sulcus occupancy map for learning since the sulcus tends to be occluded in the hippocampal hard label. The boundary points provide explicit supervision of surface geometry. In Eq. (7), the Chamfer loss computed in Euclidean domain for finer level of supervision for deformation. Resulting in over 32% reduced sulcus distance than existing methods (Table 1).

Shape correspondence and clinical analysis (R1, R2). We appreciate the concern regarding possible correspondence issues. Since ground-truth landmarks are unavailable in our dataset, directly evaluating regional correspondence errors was infeasible. In Fig. 2(c), we instead evaluated correspondence using PCA-based generalization and specificity following [31]; our method achieved the lowest reconstruction errors, suggesting stable shape correspondence. We note that the improved correspondence quality influences the observed statistical differences in Fig. 3.We apologize for the confusion. The colormap in Fig. 2(b) represents Koenderink’s curvedness but not the Jacobian determinant; the striped patterns do not indicate deformation instability. The per-vertex surface-area analysis requires inter-subject shape correspondence. Thus, ODC-based methods were excluded in area analysis as shape correspondence is not provided during their surface reconstruction.

Surface visualization (R1): The sulcal indentation is visible in the surface-volume overlay in Fig. 2(a). However, it is less apparent in the surface rendering due to the selected viewpoint and limited panel size. We will update the figure to better highlight the existing sulcal structure using enlarged slice-based overlays and transparent or cut-away renderings that expose the indentation.

Terminology (R2): We agree that the term “semi-supervised” is ambiguous. We used this to indicate that sulcus labels are used only as auxiliary training supervision and are not required at inference. To avoid confusion, we will rename this component as training-time sulcus supervision.

Minor fixes (R1, R2, R3): We will correct typos and add the missing axis labels as well as missing citations. To further support reproducibility, we will add implementation details about the U-Net architecture and release the code as stated in the abstract.




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 reviewers have provided valuable feedback on the paper, which proposes a novel sulcus-aware hippocampal surface modeling framework. Given the mixed reviews (two Weak Reject, one Accept), the authors must address the lack of clarity in formulation explicitly defining the sulcal mask’s soft occupancy maps and the role of signed distance fields. The most crucial experiment is to show the gain only to the sulcus-aware components.

  • 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 addressed the main concerns of the reviewers and thus warrants an 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.

    Reject

  • Please justify your recommendation.

    Reviewers summarized strengths of your work such as potential clinical relevance and novel aspects related to methodological changes in view of the relatively new domain of postmortem MRI. However, they also detailed significant weaknesses and shortcomings of concept, methodology and descriptions. There is no doubt that the application area is an emerging field that requires adaptation and of known segmentation and modeling concepts but also new approaches – also facing challenges regarding validation and comparison. However, balancing pros and cons and MICCAI’s high bar, your paper may be considered not to be mature enough and thus not fully ready to be accepted at MICCAI.



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.

    This submisson proposes a sulcus-aware hippocampal surface modeling using template-based diffeomorphic deformation. R3 liked clinical relevance. R1 and R2 noted mathematical inconsistencies, minor. The rebuttal has clarified all concerns. The consensus is acceptance and novelty in method. For all these reaons, and with respect to other submissions, recommendation is towards acceptnace.



back to top