Abstract

Clinically significant geometrical measurements of the aortic valve, such as the intercommissural distances and angles, are enabled by aortic valve morphometry. While morphometry typically relies on detecting key anatomical landmarks of the aortic valve (nadirs and commissures), the accuracy of morphometric measurements derived from these landmarks remains under-investigated. In this study, we devise a graph-based representation, where morphometric measurements of the aortic valve are represented as graph nodes, and their anatomical relationships are represented as graph edges that encode the dependencies of the underlying landmarks. Such representations, which therefore encapsulate the geometric variability of landmark detection, are optimized through a graph neural network (GNN) that integrates geometric features and structural context to compute anatomically consistent morphometric measurements. The proposed GNN approach was evaluated on 280 cardiac computed tomography images consisting of three cohorts from two different institutions with both healthy and pathological aortic valve cases. By computing 29 different morphometric measurements (18 linear and 11 angular) with the proposed GNN approach and comparing them to those obtained from direct landmark detection, we observed a reduction in the mean absolute error by around 15-25% on average. The results demonstrate an improved anatomical consistency of the proposed GNN approach in extracting morphometric measurements, and its robustness to landmark detection uncertainty.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5465_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/kam-ibragimov/morpho-gcn

Link to the Dataset(s)

N/A

BibTex

@InProceedings{IbrKam_Graph_MICCAI2026,
        author = { Ibragimov, Kamil AND Jelenc, Matija AND Vrtovec, Tomaz},
        title = { { Graph Neural Networks for Morphometric Characterization of the Aortic Valve } },
        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 paper suggests a graph-based approach for the geometric representation and assessment of the aortic valve via a connected landmarks.

  • 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 authors address a relevant clinical problem. The idea of the consideration of the spatial relations between relevant landmarks via graph representations is well motivated. The approach is well described.

  • 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 seems to be work-in-progress. There are several mistakes (e. g. surgerical, landmarks of the aortic,. . . ) in the text and references, which do not support the statements in the text ([7]). There is no assessment of the of inter- or intraobserver variability of the reference annotations or an ablation study.

  • 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?

    The paper presents a well motivated approach for improving the extraction of relevant geometrical parameters to TAVI planning CTs. The method strongly depends on a preprocessing step, which is implemented with a standard method. In the evaluation the assessment of the reproducibility of the reference method is missing. The authors do not present an ablation study.

  • 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 GNN-based method for optimizing aortic valve morphometric measurements derived from pre-segmented landmarks. Landmarks are first created from nnU-Net, producing N=10 candidate points per landmark. The method then constructs a graph-based representation where each node corresponds a measurement, and edges are weighted by the similarity of shared landmark subsets between measurements. Each node holds K=1000 randomly sampled candidate landmark-point combinations, described by geometric descriptors (distances, angles, anisotropy). A GNN with graph convolutional layers shares context across interdependent measurements, then produces softmax weights over candidates. The final measurement is a weighted sum over the candidate measurement values — the GNN selects among candidates rather than regressing measurements directly.

  • 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 problem is clinically relevant. Accurate aortic valve measurements are important for surgical planning, and automating robust measurement from detected landmarks addresses a real clinical need.

    2.The graph-based representation is well-motivated and clearly formalized. Defining edges via similarity of landmark subsets is an intuitive and principled way to encode anatomical dependencies between measurements.

    3.The GNN architecture — from row-wise MLP embedding, through mean-pooled context, GCN message passing, context re-injection via residual MLP, to softmax-weighted candidate selection — is described in detail with clear notation and equations. The design choice to select among anatomically plausible candidates rather than directly regressing measurements is a sensible inductive bias.

    4.The evaluation covers a comprehensive set of 29 measurements (18 linear, 11 angular), and results show consistent improvement over the baseline.

  • 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.Reference [6] is cited in the context of deriving measurements from landmarks, but it is a general anatomy review that describes aortic root structure — it does not define a measurement protocol. The 29 evaluated measurements remain undefined: which landmark subsets are used and what geometric operations are applied for each measurement are not specified in the paper or in the cited reference.

    2.The geometric descriptor for larger set of landmarks is described with “e.g.” and “etc.”, leaving the exact feature set and dimensionality D unspecified. Since the descriptor is the foundational input to the entire pipeline, the complete feature list should be explicitly stated.

    3.The baseline comparison is insufficient to isolate the GNN’s contribution. The baseline uses a single Center-of-Mass point per landmark, while the proposed method uses a GNN-weighted average over K=1000 candidate combinations. The observed improvement could arise from three distinct sources: (a) the inherent robustness of ensembling over multiple candidates, (b) the GNN’s learned weighting, or (c) diversity among candidate points. Without an ablation comparing uniform (equal-weight) averaging over the same K=1000 combinations against the GNN-weighted result, the reader cannot determine how much of the improvement is attributable to the GNN architecture versus the trivial benefit of multi-candidate averaging.

    4.No training loss function is specified anywhere in the paper. The architecture is described in full detail, but the optimization objective — what the GNN is trained to minimize — is absent. This raises several unanswered questions: whether MSE or MAE is used, how losses across linear (mm-scale) and angular (degree-scale) measurements are balanced, whether any regularization is applied to the softmax weights, and what serves as the ground truth reference for supervision.

    5.The ground truth was provided by “an experienced cardiovascular surgeon” but no inter-observer variability analysis is reported. For clinical measurements, a single annotator’s measurements are insufficient to establish a reliable reference standard. Without inter-observer analysis, it is unclear whether the GNN’s improvement is within the noise floor of human measurement variability.

    6.The authors acknowledge dependence on landmark detection accuracy as a limitation but provide no quantification. No landmark localization error is reported, and no sensitivity analysis examines how upstream detection noise affects downstream measurements. Without this, it is impossible to assess whether the GNN is learning genuinely optimal geometric weighting or compensating for detection noise.

  • 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

    Questions for the authors:

    1.Could you provide the complete landmark-to-measurement assignments for all 29 measurements, including the geometric operation used to compute each measurement from its landmarks? Are these based on an established clinical protocol?

    2.What is the training loss function? How are linear and angular measurement losses scaled or balanced?

    3.What is the performance of uniform (equal-weight) averaging over K=1000 candidate combinations compared to CoM and compared to GNN-weighted averaging? This ablation is essential to separate the ensembling benefit from the GNN’s learned contribution.

    4.What is the landmark detection accuracy (e.g., mean localization error in mm)? How does downstream measurement accuracy vary across cases with good vs. poor landmark detection?

    5.K=1000 random combinations are sampled, but the total combinatorial space is N^L per measurement where L is the number of landmarks. For measurements with many landmarks, is K=1000 sufficient to cover the space? Does this vary by measurement, and was any analysis performed on the sensitivity to K?

    Minor Issues:

    • There’s a typo in Page 6, results section: “resuls” should be “results”
    • There’s a type in Page 3, method section, Landmarks and Landmark Candidate Points: “Ket” should be “Let”
  • 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 clinically relevant problem and proposes an interesting graph-based formulation connecting interdependent aortic valve measurements through shared landmarks. The GNN architecture is clearly described at the equation level.

    However, the paper has several significant methodological gaps that collectively prevent reproducibility and undermine the evaluation. The landmark-to-measurement assignments — which define the graph structure, the input features, and the measurement computations — are never specified or cited. The geometric descriptor features are incompletely described. No training loss function is provided. These are not minor implementation details but core components of the method: the reader cannot determine what is being measured, what features describe it, or what the network optimizes.

    The experimental evaluation is weakened by an insufficient baseline. Comparing a single-point estimate (CoM) against a learned ensemble of 1000 candidates conflates ensembling benefits with the GNN’s learned contribution. Key ablations — uniform averaging baseline, graph vs. no-graph comparison — are missing. The ground truth relies on a single expert without inter-observer analysis, and the acknowledged limitation regarding landmark detection accuracy is stated without any quantification.

    While the core idea has merit, the current manuscript does not provide sufficient methodological detail for reproducibility or sufficient experimental evidence to support its claims. Addressing the missing specification (landmark assignments, full feature set, loss function) and adding the ablation studies described above would substantially strengthen the paper.

  • 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 thank the authors for the thorough rebuttal, which substantively addresses my main concerns. The clarified loss function (unweighted MSE), geometric descriptor set, inter-observer variability analysis, and uniform averaging ablation are all reassuring and demonstrate that the GNN’s contribution is genuine and clinically meaningful. I am updating my recommendation to Accept.

    I strongly encourage the authors to incorporate the following in the camera-ready version:

    • The landmark-to-measurement assignment table as promised, ideally accompanied by a figure illustrating landmark positions and measurement geometry, with citations to the measurement protocol followed.

    • The training loss function — even a single sentence suffices.

    • The complete geometric descriptor list and dimensionality.

    • The inter-observer variability and landmark localization accuracy results, which were available but unreported and are essential for interpreting the method’s clinical significance.

    • The uniform averaging ablation result, which isolates the GNN’s contribution from ensembling.

    These details would transform the paper from a promising method with insufficient specification to a strong, reproducible contribution.



Review #3

  • Please describe the contribution of the paper

    The primary objective of this article is to obtain precise three-dimensional landmarks of the aortic valve in order to derive more accurate morphometric measurements. The study focuses on the development of a refinement mechanism based on graph neural networks (GNNs). This mechanism takes as input the anatomical landmarks detected using “conventional” methods (nnU-Net) and produces morphometric measurements of the subjects under study as output. The challenge arises from the fact that the nnU-Net framework produces a point cloud for each detected anatomical landmark rather than a single representative point. Consequently, an additional processing step is required to determine a unique and reliable landmark position from this point cloud. In this study, two strategies are investigated: (i) the point cloud is approximated by a sphere, and its center of mass (CoM) is taken as the landmark location from which morphometric measurements are derived (conventional approach); and (ii) the point cloud is directly introduce to a graph neural network (GNN), which infers the morphometric measurements without explicitly estimating a single landmark point (proposed approach).

  • 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.

    Although the proposed methodology has been applied in other domains for extracting anatomical measurements, to the best of our knowledge, it has not previously been used to derive measurements related to the aortic valve. Additionally, the method has been trained and evaluated using three different cohorts from two independent institutions, which supports its robustness and generalizability.

  • 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 main limitation of the proposed approach, as acknowledged by the authors, is that it functions primarily as a refinement tool. Consequently, if the initial landmark detection results are highly inaccurate, the method cannot effectively correct them. Nevertheless, in practical terms, current landmark detection techniques are generally robust, and in cases of failure, it is reasonable to assume that inconsistent results would be readily identified by a specialist. Furthermore, the non-public nature of the datasets used in this study limits the ability to fully assess the generalizability and external validity of the proposed method.

  • 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.

    (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?

    Overall, this is a comprehensive and original study that presents promising results. However, from a reproducibility standpoint, neither the datasets nor the source code are made publicly available through platforms such as GitHub, which restricts independent validation and broader adoption of the method.

  • 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.

    It is clear that this article exhibits certain limitations in terms of technical detail and the reporting of results, as noted by several reviewers. Perhaps the most significant limitation is the absence of an ablation study that would explicitly demonstrate the contribution of GNNs to model performance. Nevertheless, the manuscript presents a novel and well-motivated approach, whose primary strengths lie in its originality and potential clinical applicability. The idea of refining landmark-derived measurements using GNNs is conceptually appealing and potentially impactful, and the reported results provide some support for this premise. In summary, while a more comprehensive ablation study would strengthen the work, this limitation does not, in my view, substantially detract from its overall quality.



Review #4

  • Please describe the contribution of the paper

    This work proposes a graph neural network (GNN)-based approach for automated morphometric analysis of the aortic valve. Landmark detection outputs are used to construct a graph where nodes represent measurements (with uncertainty modeled via multiple candidate points per landmark) and edges encode interdependencies between measurements. The method is evaluated on CT data from two institutions (including healthy and pathological cases), showing improved accuracy (15–25%) over measurements derived directly from automatically detected landmarks, using manual measurements as reference.

  • 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.

    Novelty: The work explores an original hypothesis by shifting the focus from landmark detection to the refinement of landmark-derived morphometric measurements using a GNN. This is relevant, as many pipelines directly compute measurements from detected landmarks without modeling anatomical constraints or interdependencies. The proposed approach explicitly incorporates these relationships, which is conceptually appealing and, according to the results, impactful. Given the prevalence of landmark-based pipelines in medical imaging, this idea may generalize to other clinical applications and be of broader interest to the community.

    Mathematical formalism and illustration: The methodology is clearly described, well thought out, and mathematically well formulated. The authors are also commended for Fig. 1, which effectively illustrates the proposed pipeline and aids understanding.

    Adequate experimental design: The evaluation is well designed, using CT data from two institutions with both healthy and pathological cases. Comparison against landmark-based measurements, using manual measurements as reference, is appropriate, and implementation details are sufficient to support reproducibility. A limitation is the absence of a dedicated generalization assessment, as data from both institutions appear to be used in both training and testing.

  • 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.

    Limited interpretability: The approach lacks interpretability in the final measurements. Since the GNN aggregates multiple candidate landmark configurations, each measurement is derived from a weighted combination rather than a single consistent landmark set. Even in the theoretical case where a single configuration dominates, different measurements may rely on different candidates of the same anatomical landmark. This prevents identifying or visualizing the exact landmark positions underlying the results, limiting verification and reuse in downstream tasks. In practice, one would need to revert to the original landmark outputs (e.g., center-of-mass), despite their lower accuracy.

    Lack of ablation study: The manuscript lacks analysis of key design choices in the GNN. The impact of the number of candidate points per landmark (Nᵢ), the number of candidate geometric descriptors per measurement (K), the contribution of specific feature groups in these descriptor vectors (e.g., is compactness relevant?), the embedding dimension (E), and other architectural components (e.g., edge weighting mechanism) is not analyzed. Such ablations would be important to assess the method’s robustness.

  • 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?

    The authors make use of private datasets (pages 5-6), but no explicit statement is provided regarding ethical approval or IRB compliance.

  • 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) In the related work section, please clarify the distinction between the strategies described in the first and third groups. Both appear to rely on segmentation-based pipelines. Is the difference primarily related to the downstream task (morphometry vs. other applications), or is there a methodological distinction? 2) It is not explicit from the Introduction, (and to a minor extent the abstract) that the proposed method explicitly leverages detected landmarks and their associated uncertainty, rather than solely optimizing measurements based on anatomical relationships. This is a key aspect of the methodology (as it anchors the solution to the input image) and should be more clearly reflected in the contribution description. Its importance only becomes evident in the Methods section. 3) Contribution 3 appears to follow directly from Contribution 2 (i.e., the GNN formulation modeling landmark uncertainty and interdependencies), rather than constituting an independent contribution. Please clarify or reconsider this distinction. 4) Are there measurements that depend on more than three landmarks? An example descriptor is provided (anisotropy), but the wording (“can compute”) introduces ambiguity. Please clarify whether such cases are effectively used in the method. 5) Consider discussing the interpretability limitation (see above) in the discussion or limitations sections. 6) Please clarify the rationale for including the compactness feature. If C_f is a candidate set of landmarks for a given measurement f, the centroid is a physical location in the input volume. Why is compactness needed? If used as just another geometric descriptor, why was compactness chosen over alternatives? 7) Please provide further details about the dataset, particularly regarding the types of pathologies included in the cohorts. 8) Was nnU-Net trained using the default configuration (automated hyperparameter selection), with the only particularity being how the GT masks were created? If so, please state this explicitly for transparency and reproducibility. 9) Consider highlighting the best-performing method for each measurement in Table 1 (e.g., use bold font) to improve readability and emphasize the results.

    Minor remarks:

    • Please proofread the manuscript for grammatical errors or unclear phrasing. A non-exhaustive list includes:
    • First sentence of Related Work: (p.2) the aortic root and valve are complex anatomies, not their 3D representation.
    • “key anatomical landmarks of the aortic, where” (p.2): a word appears to be missing.
    • “where architectures both CNN architectures” (p. 2): duplicated wording.
    • “Ket the key” should read “Let the key” (p.3)
    • “probability that are at least for distance d apart from the” (p.3): unclear phrasing.
    • Variable N is not defined in the landmark set notation, nor its relationship to index i.
    • Consider removing “with shared landmarks” from the initial edge definition (p.4). In practice, edges exist between all measurement pairs (as reflected in the square symmetric adjacency matrix), hence the need to define the zeroed connection strength where no actual shared landmarks exist.
  • 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 manuscript presents a novel and well-motivated approach, with strengths in its originality and the adequacy of its experimental design. The idea of refining landmark-derived measurements through a GNN framework is both conceptually interesting and potentially impactful, and the evaluation convincingly supports it. While the lack of an ablation study is a notable limitation, it does not detract from the overall validity of the work.

  • 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 maintain my original accept recommendation. The manuscript presents a valid, interesting, and potentially impactful contribution, particularly given the originality of refining landmark-derived morphometric measurements through a GNN framework and its possible applicability to other landmark-based problems in medical imaging. I further commend the authors for the additional methodological clarifications, particularly regarding the valve measurements, descriptors, and loss formulation. The inclusion of previously unreported landmark detection baseline performance and inter-/intra-observer variability analyses further strengthen and contextualize the results. Finally, although ablation and noise robustness studies cannot be added, the discussion provided increases confidence in the robustness and validity of the proposed approach.



Author Feedback

We thank the reviewers and meta-reviewer for providing valuable feedback. Several comments are pointing to new/additional experimental results, which are however not allowed in the rebuttal. Nevertheless, we indicate on our framework performance that will be considered in our eventual future work. We therefore focused on the following comments:

Measurement details (MR, R#2, R#6): The evaluated morphometric measurements are computed from anatomically defined landmark subsets using deterministic geometric operations. Depending on the measurement, these operations include Euclidean distances, angles between vectors or plane normals, centroid-based angular calculations, and ellipse fitting using SVD and Powell optimization. We will provide a list of landmark-to-measurement assignment for each morphometric measurement in Table 1. Geometric descriptors, loss function and compactness (MR, R#2, R#6): Candidate descriptors consist of: 1) Euclidean distances between landmarks, 2) angular relationships encoded as cosine similarities, 3) PCA descriptors using the two largest eigenvalues, and 4) compactness as the mean distance of participating landmarks to their centroid (describes spatial dispersion and overall scale of landmark configuration, complementing pairwise distances with a rotation-invariant global descriptor). The model is trained using a standard MSE regression objective applied jointly across all predicted measurements (with no weighting of linear and angular measurements, auxiliary losses or softmax regularization terms).

Ablation study (MR, R#1, R#2, R#6): We differentiated the contribution of graph-based weighting from candidate-based averaging, and evaluated a uniform averaging baseline using the same candidate sets and the same number of candidate combinations, but without GNN. Compared to CoM, the reduction in MAE was the greatest for GNN, whereas uniform averaging achieved smaller improvements, indicating that the observed gains are not solely explained by ensemble averaging, but additionally benefit from GNN. We acknowledge this experimental suggestion, and will consider it in our future work.

Observer variability (MR, R#1, R#2): In terms of normalized MAE, the proposed GNN achieved an error of 2.5±0.9% for linear and 9.6±10.2% for angular measurements. For a subset of 25 cases, the inter-observer variability was 3.0±1.6% and 15.4±15.8%, while the intra-observer variability was 1.6±1.0% and 8.4±9.5%, respectively, confirming that the proposed framework operates within the inter-observer variability. Note that no new experiments were performed, as these results were already available but not reported.

Noise (MR, R#5): We evaluated the effect of perturbations applied to candidate landmark coordinates using different Gaussian noise levels, and GNN showed substantially greater robustness compared to the CoM baseline. The relative advantage of GNN over CoM increased with noise, indicating that graph-based aggregation improves robustness by leveraging anatomical contextual information. We acknowledge this experimental suggestion, and will consider it in our future work.

Landmark localization (R#1, R#2, R#6): This was performed by nnU-Net with the default configuration and automatic hyperparameter selection. The only task-specific modification was ground truth mask generation (spherical landmark-centered regions with a radius of 3.6 mm). Based on the predicted heatmaps, the landmark localization error was below 0.95 mm for 99.4% of landmarks (only two exceeded 3.6 mm). Note that no new experiments were performed, as these results were already available but not reported.

Code and Data (All): We are evaluating the possibility of publicly releasing the source code. For data, we will allow sharing upon reasonable request. Ethical approval for data acquisition and usage exist and will be reported.

Text/referencing errors (MR, R#2, R#6): Thank you, we will correct the typos and referencing errors.




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.

    While reviewers agreed that the paper addresses a clinically relevant problem with a well-motivated, detailed graph-based GNN architecture and validation across multiple cohorts, they also raised serious concerns about the paper’s methodological clarity, evaluation rigor, and reproducibility: First, critical methodological details are missing, including undefined aortic valve measurements, unspecified geometric descriptors, and no stated training loss function, hindering reproducibility. Second, the evaluation is insufficient—baseline comparisons fail to isolate the GNN’s contribution, there is no inter-observer variability analysis for ground truth annotations, and no quantification of how landmark detection errors affect downstream measurements. Third, the paper contains text/referencing errors and uses non-public datasets, limiting external validity and trustworthiness.

    Since the reviews are mixed, we would like to invite the authors to provide a rebuttal. In addition to clarify other concerns, please try to address the following points during rebuttal: provide more details about the aortic valve measurements and their landmark subsets, specify the full geometric descriptor set and training loss function; conduct an ablation to isolate the GNN’s contribution (vs. multi-candidate averaging), report inter-observer variability for annotations, and quantify how landmark detection noise impacts measurements; correct text/referencing errors and provide details to enhance external validity (e.g., dataset access or more cohort details).

  • 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.

    Reviewers agreed that the paper addresses a clinically relevant problem with a well-motivated graph-based approach for refining aortic valve measurements, and that the method is comprehensively evaluated across multiple institutions, demonstrating potential robustness and generalizability. However, they also raised serious concerns about the paper’s technical completeness and scientific rigor: (1)​critical methodological details were missing or underspecified, including the measurement protocol linking landmarks to clinical metrics, the complete geometric descriptor set, and the training loss function; (2) the contribution of the GNN was not convincingly isolated, as the improvement could stem from multi-candidate ensembling rather than learned relational modeling, and no ablation with uniform averaging was provided; and (3)​the reliability of the reference standard was unclear, with no inter-observer variability analysis, no quantification of landmark localization accuracy, and limited discussion of clinical interpretability and robustness. The rebuttal has successfully addressed most concerns by supplying the missing specifications, reporting inter-observer variability and landmark detection performance, and clarifying the loss formulation and descriptor design. Since the review scores are all positive after rebuttal, I recommend accepting this manuscript.



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.

    Though with some limited novelty and short in ablation study, in general it is a quite interesting and novel study



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 paper presents a clinically relevant and well motivated method for refining aortic valve morphometric measurements using graph-based modeling of landmark dependencies. The rebuttal addressed several important concerns, including measurement definitions, descriptor details, loss function, observer variability, landmark localization performance, and ethical approval. While the ablation analysis and external generalization remain limited, the contribution is sufficiently novel and the reported results support acceptance.



back to top