Abstract

Couinaud segmentation is a fundamental step in surgical planning for anatomical liver resection. Anatomically, Couinaud liver segments are defined by the vascular system. However, in medical imaging, the critical issue is that the defining landmarks, i.e., vessels, are extremely sparse within the vast liver parenchyma. Current methods employ handcrafted dense sampling strategies to exploit more vascular data points, leading to high energy consumption and a heavy dependency on manual data rebalancing. We propose Spiking Serialized Point Transformer (SSPT), a Spiking Neural Network (SNN) architecture, to tackle these challenges efficiently and effectively. In particular, SSPT devises Leaky Integrate-and-Fire (LIF) neurons to enable a 3D transformer framework, serializing unordered point cloud data into structured temporal spike trains. This spike-based design enables temporally consistent aggregation of critical anatomy (e.g., vessels) context from sparse medical imaging data, while avoiding manually designed sampling strategies and their associated biases. Experimental results on MSD8 and LiTS datasets demonstrate that SSPT delivers high-fidelity segmentation (up to 92.14% Dice) and robust boundary precision (down to 2.24 mm ASD) with a minimal energy footprint of 10.97mJ. Moreover, our experimental results demonstrate that spiking dynamics can inherently capture sparse anatomical topologies, presenting a robust and energy-efficient solution for offline preoperative liver surgery planning. Code is available at https://github.com/Yan0918/SSPT.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2435_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Yan0918/SSPT

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiYan_SSPT_MICCAI2026,
        author = { Li, Yan AND Li, Ruiyang AND Qiu, Yue AND Liu, Qixuan AND Pei, Jialun AND Qiu, Shi AND Heng, Pheng-Ann},
        title = { { SSPT: Spiking Serialized Point Transformer for Couinaud Segmentation in 3D Medical Data } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper presents SSPT, a spiking neural network (SNN) point transformer for Couinaud liver segmentation. The key idea is to take the 3D liver point cloud and lay it out as a 1D sequence using a Z-order curve that keeps nearby points close together, then feed chunks of this sequence to the network over a few time steps like a rhythmic signal (with TSD: Temporal Scanning Dispatcher). Attention is then restricted to local windows, and the expensive multiplications in the attention computation are replaced with simple additions driven by binary spikes (SSA). Evaluated on LiTS and MSD8, SSPT achieves 92.14% Dice at 10.97 mJ, which beats the SOTA (APVN) and allows a 98% energy reduction versus the PTv3 counterpart at equal parameter count (12M).

  • 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.
    • Novel and well-scoped contribution: it seems to be the first transformer-based SNN for 3D Couinaud point-cloud segmentation. The energy-efficiency narrative is clean and quantitative (45nm CMOS model, per-layer firing rates reported).
    • The PTv3 comparison at equal parameter count is the right experiment. It isolates spiking dynamics from capacity effects.
    • TSD is a well-motivated solution to the spatio-temporal mismatch problem in SNN point processing. The ablation Table 2 Cat. III is convincing.
    • Serialization ablation (Z-order vs. random vs. k-NN) and time-step analysis (T=1/2/3) are informative and support the design choice made.
  • 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.
    • Dataset labels are undocumented. Neither LiTS nor MSD8 ships with Couinaud annotations. The paper never states where they come from.
    • Preprocessing details for voxel baselines are not given.
    • Table 2 Cat. III contradicts Table 1.The Scanning Wave row reports LiTS ASD=2.24 / MSD8 ASD=2.44, but Table 1 and the abstract report LiTS=2.44 / MSD8=2.24.The columns appear transposed.
    • No statistical testing performed, the tables lack standard deviations.
    • The claim “the binary gating mechanism of LIF neurons serves as a dynamic threshold, prioritizing high-entropy structural regions such as vessel bifurcations” is not verified. No spike-activity visualization is provided, only segmentation outputs are shown.
    • The “Real-time intraoperative” claim in the conclusion is unsupported. No wall-clock latency or neuromorphic hardware deployment is reported. The energy estimate is theoretical.
    • The theoretical energy model (versus actual neuromorphic hardware measurement) and the single-dataset / single-split evaluation for both MSD8 and LiTS are limitations that should be stated more explicitly.
  • 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

    I would appreciate if the authors publish the code if the paper gets accepted.

  • 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 core idea is to show that an SNN point transformer can match ANN accuracy on a clinically relevant 3D medical task at a fraction of the energy: this is interesting and novel. The TSD mechanism and the point-specific PTv3 comparison are well-executed. However, the undocumented Couinaud labels, the internal table inconsistency, the likely under-tuned nnU-Net and the unsupported intraoperative claim prevent me from giving a strong accept. With fixes to these issues, this would be a solid contribution.

  • 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 tried to answer the main concerns raised by the reviewers. Also code will be released which makes their approach at least reproduceable for the scientific community.



Review #2

  • Please describe the contribution of the paper

    The paper proposes SSPT, a spiking neural network for Couinaud liver segmentation from point cloud representations of 3D medical images. Its main design includes a Temporal Scanning Dispatcher (TSD), which organizes 3D points along a space-filling curve and feeds them over time steps to Leaky Integrate-and-Fire neurons, and a Spiking Serialized Attention (SSA) module, which uses binary spike-gated queries within local windows to reduce arithmetic complexity. Experimental results on LiTS and MSD8 suggest that the method is competitive with existing voxel-based, hybrid, and point-based approaches in segmentation accuracy, and may offer advantages in theoretical energy efficiency.

  • 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.Introduces a principled spatio-temporal serialization (TSD) of unordered point clouds to create coherent spike trains, addressing the common mismatch between static 3D geometry and SNN temporal dynamics. 2.Designs a spiking self-attention variant (SSA) that uses binary query spikes with localized windows to approach linear complexity while replacing much of QK multiplications with additions, a compelling hardware-aware rethinking of attention for point clouds. 3.Incorporates a spiking conditional positional encoding (SCPE) with parallel sparse 3D convolutions to compensate spatial information loss due to serialization; the end-to-end hierarchical U-Net-like spiking transformer is 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.

    Major concerns: 1.The energy analysis is theoretical (45 nm CMOS model) and not hardware-measured; it does not include memory movement costs or GPU/CPU runtime, which can dominate in practice for point-cloud pipelines. 2.SSA still performs floating-point softmax/value aggregation; the overall fraction of operations actually reduced to additions versus multiplications is not fully quantified per layer, making the magnitude of energy savings somewhat assumption-sensitive. 3.The gating granularity and stability of spike generation (per-channel/per-head) and its impact on attention quality are under-specified; no sensitivity analysis for spike threshold or membrane dynamics is provided. 4.The most serious issue is the unclear source of the Couinaud ground truth. LiTS and MSD8 do not provide Couinaud annotations by default, yet the manuscript does not explain how these labels were obtained, generated, or validated. This is not a minor omission but a fundamental problem that undermines the rigor and credibility of the entire study. 5.The evaluation is incomplete and insufficiently convincing. The paper reports only overall metrics, without segment-wise Dice results for each Couinaud region, especially for anatomically challenging segments such as Segment I (caudate lobe). This makes it difficult to judge the actual clinical usefulness of the method. 6.The computation of ASD is unclear and lacks reproducibility. Since the predictions are point-wise, the paper should clearly describe how point labels are converted into surfaces or volumes before boundary distances are measured. Without these details, the reported ASD values are difficult to interpret. 7.The fairness of the comparisons is questionable. Voxel-based methods are evaluated on dense full-resolution volumes, while the proposed method uses sampled point clouds. The energy comparison is therefore not fully fair, and the paper does not provide runtime or peak memory results to support its practical efficiency claims. 8.Important baselines are missing. The manuscript does not compare against recent Couinaud-specific, vessel-guided, or clinically relevant methods. In particular, Couinaud segmentation in CT has already been addressed in mature frameworks such as TotalSegmentator v2, yet no comparison is provided. This makes the practical contribution unclear. 9.The qualitative results do not convincingly support the claimed advantages. Even in the visual examples, the proposed method still shows noticeable boundary offsets and inaccurate delineations. This weakens the claim that point-cloud-based modeling is particularly advantageous for this task. 10.Several table entries appear duplicated or inconsistent (e.g., identical Dice/ASD in ablation for different serialization settings; some text/table typos), which undermines confidence in reported numbers. 11.Data preprocessing and feature composition are under-specified (e.g., what per-point features are used beyond coordinates/intensity? any normalization, augmentation?). 12.Training/validation/test splits, case counts per split, and cross-validation protocols are not reported; reproducibility is limited without code or detailed settings (e.g., window size w, neighborhood/window overlap, spike thresholds). 13.Does not discuss or compare to recent Couinaud-specific methods that incorporate vascular priors or registration-based augmentations. Given the paper’s focus on topology-aware segmentation and boundary precision, omitting these is a notable gap.

  • Please rate the clarity and organization of this paper

    Poor

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

    The submission does not provide sufficient information for reproducibility.

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

    N/A

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

    N/A

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

    (2) Reject — should be rejected, independent of rebuttal

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

    The paper introduces a creative and potentially impactful SNN-based framework for 3D point-cloud liver segmentation, with principled serialization and a hardware-conscious attention design. The methodological novelty is credible and the energy-efficiency narrative is compelling. However, serious concerns about experimental validity and reproducibility remain: most critically, the provenance of Couinaud labels on LiTS and MSD8 is not explained, undermining confidence in the central claims; the boundary metric computation pipeline is unspecified; and energy/runtimes are presented largely as theoretical estimates without corroborating system-level measurements. The omission of direct comparisons to recent Couinaud- or vessel-guided methods further weakens contextualization. With detailed clarification of labels, reconstruction steps, and practical runtime/memory measurements—and with stronger contextual baselines—this work could be a valuable contribution. In its current form, I recommend rejection, primarily due to dataset/labeling ambiguity and incomplete experimental reporting, despite the promising technical ideas.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    still lack novelty



Review #3

  • Please describe the contribution of the paper

    This paper presents a spiking neural network (SNN) architecture called Pulse Sequence-Encoded Point Cloud Transformer (SSPT) for Couinaud liver segmentation of 3D medical point clouds. By introducing a time scanning scheduler (TSD), the 3D spatial topology is serialized into one-dimensional time pulses, and the pulse-encoding attention (SSA) is utilized to convert floating-point multiply-add operations into low-power accumulation operations. This model demonstrates excellent performance in terms of theoretical energy consumption. Introducing the low-power characteristic of SNN into 3D medical point cloud processing is an inspiring and highly promising research direction. However, the current manuscript has significant flaws in data preprocessing as a black box, the fairness of cross-modal baseline comparison, the reproducibility of experiments, and the consistency of clinical motivation. These issues directly affect the validity and credibility of the experimental results. Therefore, I currently give a preliminary evaluation of “Weak Reject”. But I am very much looking forward to seeing the authors’ strong responses to the following core issues in the Rebuttal stage.

  • 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.

    Major Strength 1 The paper presents a novel and well-motivated architecture adaptation by elegantly combining the Z-order space-filling curve with the sliding window mechanism (TSD). This design effectively resolves the critical spatiotemporal mismatch between static 3D geometric point clouds and the continuous spike dynamics of spiking neural networks (SNNs). Such a cross-paradigm alignment is non-trivial and represents an original way to bridge discrete geometric representations with event-based, continuous-time computation, making the work technically innovative and conceptually interesting. Major Strength 2 A key strength lies in the extremely high theoretical energy efficiency enabled by the proposed SSA module. The module maintains linear complexity while drastically reducing floating-point operations. Compared with traditional artificial neural networks, the work reports a remarkable 98.1% reduction in theoretical energy consumption, which is outstanding and highly relevant to resource-constrained medical imaging and clinical deployment scenarios. This strong efficiency advantage significantly enhances the practical value of the method.

  • 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.

    Major Weakness 1 The paper is primarily motivated by real-time intraoperative surgical planning and ultra-low-power constraints to justify the use of SNNs. However, Couinaud segmentation is clinically a preoperative offline planning task, where energy consumption and millisecond-level latency are generally less critical than absolute segmentation accuracy. This makes the core motivation less convincing. Major Weakness 2 The paper lacks critical details regarding 3D point cloud extraction from raw CT/MRI volumes. The authors state that data are converted to 3D point clouds via windowing/cropping and then sampled to 60,000 points using FPS. However, how the initial point cloud is generated from the 3D volume is completely unclear, whether by intensity thresholding, uniform grid sampling, or another strategy. This missing step is essential for reproducibility, because segmentation performance strongly depends on the initial point distribution, especially when extracting sparse vascular structures. Major Weakness 3 The claimed energy efficiency is based on theoretical calculations under a 45nm CMOS process, which is acceptable for conceptual validation but incomplete for SNN research. Standard evaluation in this field also requires practical runtime metrics independent of custom ASIC hardware. The authors should supplement their results with actual inference speed (ms/volume or FPS) and peak GPU memory usage (MB) on a commodity GPU such as RTX 3090 to complement the theoretical energy claims and enable fairer comparison with existing ANN methods. Major Weakness 4 To ensure full reproducibility of the proposed method, the authors are strongly encouraged to release the source code (including data preprocessing, model implementation, and evaluation scripts). Without accessible code, it is difficult for the community to verify, reproduce, or build upon the presented approach, which limits the impact and transparency of the work.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The submission does not provide sufficient information for reproducibility.

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

    N/A

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

    N/A

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

    (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?

    My recommendation of a ‘Weak Reject’ reflects a careful balance between the paper’s highly original methodological contributions and its shortcomings in clinical motivation, reproducibility, and practical evaluation. Major factors favoring acceptance: The paper presents a highly innovative cross-paradigm alignment by adapting Spiking Neural Networks (SNNs) for 3D unstructured medical point clouds. The elegant combination of the Z-order space-filling curve with the sliding window mechanism (TSD) successfully resolves the spatiotemporal mismatch between static geometry and continuous spike dynamics. Furthermore, the theoretical 98.1% reduction in energy consumption achieved by the SSA module is an outstanding technical achievement that highlights the potential of SNNs in resource-constrained scenarios.

    Major factors driving the ‘Weak Reject’: Despite the strong technical core, the current manuscript falls marginally below the acceptance threshold due to four critical issues: Misaligned Clinical Motivation: The paper justifies the use of SNNs by citing “real-time intraoperative” constraints. However, Couinaud segmentation is standardly a preoperative offline planning task where absolute accuracy far outweighs energy efficiency or millisecond latency. Critical Reproducibility Gap (Data Preprocessing): The exact pipeline for extracting the initial 3D point cloud from dense raw CT/MRI volumes (before applying FPS to reach 60k points) is opaque. Because vascular structures are sparse, blind sampling on raw voxels is physically problematic. This missing step makes the results impossible to replicate.

    Lack of Practical Evaluation Metrics: Relying solely on theoretical ASIC energy calculations (45nm CMOS) is insufficient. To fairly compare with ANN baselines, practical runtime metrics, specifically inference speed (ms/volume) and peak memory usage (MB) on standard commodity GPUs (like the RTX 3090 used)—must be reported.

    Closed Source Code: The absence of source code further exacerbates the reproducibility issues, limiting the community’s ability to verify the data pipeline and the SNN implementation. Condition for Acceptance: I “would not mind if accepted” and am very open to raising my score to an ‘Accept’ if the authors can use the rebuttal to: 1) transparently detail the raw-volume-to-point-cloud preprocessing pipeline, 2) provide practical GPU-based inference latency and memory metrics, and 3) explicitly commit to releasing their source code.

  • 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



Author Feedback

We sincerely thank the reviewers for recognizing the design of our SSPT model. We will release the full source code, including the model implementation, preprocessing, augmentation, evaluation, and energy-estimation scripts, upon publication.

1.[R1,R2,R3] Preprocessing pipeline from CT to point clouds We follow APVN (TMI 2024) for raw-volume preprocessing: from paired CT/label volumes, we crop and retain the liver regions covered by Couinaud annotations, apply CT windowing and normalization, and convert the retained voxels into point samples with coordinates, intensity, and labels; training uses controlled point sampling and testing uses full-point chunked inference, while voxel baselines directly use the same ROI-constrained normalized volumes. All methods are consistent in terms of data source, anatomical scope, intensity normalization, and spatial preprocessing. We will detail the pipeline in the revision.

2.[R1,R2,R3] Dataset labels For LiTS, we use the Couinaud annotations released by APVN (TMI 2024), for MSD8, we use the Couinaud annotations released by GLC-UNet (MLMI 2019). We will add the reference in the revised manuscript.

3.[R1,R2,R3] Practical efficiency (peak GPU memory and inference speed) We thank the reviewers for noting this issue. Regarding the efficiency comparison, we report measured efficiency on a single RTX 3090 (24 GB). On the MSD8 test set, SSPT achieves 3.79 s/volume and 6.47 GB peak GPU memory, compared with 6.933 s/volume and 5.91 GB for PTv3, and 4.438 s/volume and 6.08 GB for APVN. Peak memory is reported using maximum memory allocated. Thus, SSPTachieves an inference speed 45.3% faster than PTv3 and 14.6% than APVN, while consuming moderately more GPU memory. We will add the measured results to Table 1. 4.[R1,R2] Energy calculation We thank the reviewers for raising the fairness concern. In our experiment, to fairly evaluate the computational cost of point-based and voxel-based representations, we designed point and voxel inputs with equivalent input sizes (point 65,000, voxel (16* 64 *64)). As stated in the manuscript, we estimate the theoretical compute energy of each forward pass. Following widely adopted protocols of SPT (AAAI 2025) and SDT (NeurIPS 2023), we estimate theoretical compute energy under the 45 nm CMOS model, where ANN energy is computed from total MACs, while SNN energy is computed from first-layer MACs plus firing-rate-weighted ACs in subsequent spiking layers.

5.[R1,R3] Clarifying the clinical scope: offline Couinaud segmentation. We agree that Couinaud segmentation should be positioned as an offline preoperative planning task. We will revise any wording that could be interpreted as claiming demonstrated real-time or intraoperative deployment. Thank you for reminding.

6.[R2] ASD computation and reproducibility. Following the evaluation protocol of the APVN (TMI 2024), We map the point-wise predictions back to the NIfTI label volume using the stored voxel-index coordinates, evaluate them only within the annotated foreground mask, and then compute ASD for labels 1–8 with the surface-distance library using the physical spacing from the reference NIfTI header. We will add the details in the revised manuscript.

7.[R2] Evaluation against SOTA baselines and qualitative analysis In the submitted experiments, we selected APVN (TMI 2024), which is currently the state-of-the-art Couinaud segmentation model with publicly available source code and data.

8.[R1,R2] Table typo and inconsistency correction We thank the reviewers for carefully checking the reported results. We will correct all identified table/text inconsistencies in the revised manuscript.

9.[R1,R2] Statistical robustness and reporting completeness In the revised manuscript, we will report results as mean ± standard deviation, and provide segment-wise Dice. We will also extend spike-activity visualization to support the discussion of binary gating behavior.




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.

    This paper has some technical merit on 3D medical point cloud segmentation. The reviewers has concern on the evaluation and datasets, lack of metrics of claims. Unequal comparison, etc.

  • 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.

    For this paper, i found some reviewers’ comments and opinions are less informative or purely AI-generated; therefore, a complete re-review is conducted by AC. Based on AC’s addition review, the major concern of this paper is the evaluation and datasets; this paper seems to have only evaluated several old datasets, which may already be saturated and cannot fully reflect the real effectiveness of this proposed method.



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.

    Given the author’s rebuttal which addresses the main criticisms of the paper, my recommendation would be to accept the paper given the points for rejection raised by reviewers could be be addressed fairly easily in the revised version.



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 rebuttal substantially clarifies the dataset labels, CT-to-point-cloud preprocessing, ASD computation, practical GPU efficiency, and clinical scope, and although some concerns about baseline breadth, statistical reporting, and reproducibility remain, the technical novelty and addressed core ambiguities support acceptance with these details incorporated in the camera-ready version.



back to top