List of Papers Browse by Subject Areas Author List
Abstract
Automatic generation of patient-specific skull implants is essential for efficient and reliable cranioplasty. We propose TriFlow, a conditional flow matching model operating in a compact triplane latent space for high-fidelity 3D skull completion. By encoding skull geometry into structure-preserving triplane features and learning a conditional continuous velocity field, our method enables deterministic and efficient shape generation. A triplane-aware attention module further enforces intra-plane continuity and cross-plane geometric coherence. Experiments on the SkullFix and SkullBreak datasets demonstrate that TriFlow achieves state-of-the-art accuracy while significantly reducing training and inference cost, highlighting its potential for practical clinical deployment. The code and datasets are available here: https://anonymized.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1927_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuZhe_TriFlow_MICCAI2026,
author = { Liu, Zhenhong AND Wang, Xingce},
title = { { TriFlow: Triplane Latent Conditional Flow Matching for Efficient 3D Skull Shape Completion } },
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 presents a method for automatic shape completion in CT scans of 3D skulls. The method, dubbed TriFlow, is based on the latent flow matching framework. The main novelty is that the latent space is organized around the triplane representation to make it effective but compact as compared to full voxel maps. To this end, the authors define a triplane-aware module and a triplane convolution to be used in an otherwise standard UNet architecture for the velocity estimator. They also use Total Variation and L2 regularization on the triplane features to suppress spurious high-frequency artifacts. Experiments on the standard benchmarks for this task (SkullBreak, SkullFix, MUG500) show state of the art performance against several baselines while using at the same time the least number of parameters, FLOPs and training/testing time. An ablation study on SkullFix shows that all contributions are important for the overall performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
S1.Experimental results are the main strength of the submission. Results obtained on the standard SkullBreak/SkullFix benchmarks are remarkable. Yet, W1-W3 cast some doubts on the results that should be clarified before this strength is confirmed.
S2.Sufficient novelty: although the use of triplanes is widespread in generative 3D computer vision (it appears in Efficient 3D GAN [17] and reused in DiffusionSDF[11], just to mention two references present in the paper where it is used), the peculiar use as latents of a flow matching framework, in particular for implant design, is novel to the best of my knowledge, and well motivated with the aim of reducing the computational footprint of the method.
S3.Ablation study: the paper empirically justifies the proposed modules and losses by showing the significant effect of their introduction on the performance on SkullFix. Yet, I would have preferred to see the ablation on the more challenging SkullBreak dataset.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
W1.Experimental protocol has not been detailed. For instance, data augmentations are known to play a crucial role in the results [B,D] and it is not clear if and which have been used. I’d suggest to provide all the details of the training recipe to allow the reader to correctly interpret the results and improve reproducibility.
W2.Moreover, the manuscript does not clarify if all the methods have been retrained on the SkullBreak/SkullFix datasets, and if all are using the same training protocol. Therefore, the reader cannot be sure if the comparisons are fair.
W3.Lack of effective baselines from the 3D skull completion literature. The majority of the baselines have not been proposed for 3D skull completion (Get3D, Diffusion-SDF, DiffComplete, CDPNet (that requires also an image of the 3D data, unclear how it was generated) and PCDreamer). Moreover, the concurrent submission from the authors provided as supplementary material uses more recent and effective methods in this category, like Octfusion [E], MeshGPT[F] and LT3SD[G]. Finally, the two methods explicitly designed for 3D skull completion are 3DU-Net and PCDiff. More recent and effective or efficient methods exists, like [A], [B] or [C]. [B] in particular reports an average Dice Score above 0.94 on SkullBreak and above 0.96 on SkullFix datasets, and should be properly discussed.
W4.The main contributions are the triplane-aware module and the triplane convolution. Yet, their description is mainly provided via figure 2 and 3(2), where several important details remain unclear. The triplane aware module is at least partially described in the text, but the description of the convolution is left entirely to the figure (which is not even referenced in the text when it talks about the convolution).
As for Figure 2: there is an arrow entering the “Learnable position embedding block”, what is it? How is the final “Feature integration” performed and how does the condition c influence it? Are the self and cross attentions in the module surrounded by residual connection and normalization as in the original Transformer or are they used as depicted?
As for figure 3(2): Along which dimensions are the AvgPools performed? what is the shape of the output? From the figure it looks like the same of the branch without the AvgPool but it does not seem possible to me. Is the last 2D conv a transposed Conv to upsample the triplanes?
For both figures: what are the hyperparameters of all the layers? kernel size, number of channels, etc. .
In general, the absence of a clear description in the text of the main contributions hurts both the clarity of the paper and its reproducibility.
W5.Although the main framework is a standard latent flow matching one, the details on how it is realized are similarly delegated to figure 1 only. Therefore, some things remain unclear: what is 3D RCU? the only reference to an encoder that produces triplane is done to Efficient 3D GAN [17], which does not introduce such acronym. What is the role of the 2D ResBlock in the decoder? What does “attention_axis=z” means in the velocity predictor and how should the reader reconcile it with the architecture in figure 2?
[A]. Wodzinski, M. , Daniol, M. , Hemmerling, D. : Automatic skull reconstruction by deep learnable symmetry enforcement. Computer Methods and Programs in Biomedicine 263, 108670 (2025) [B]. M Wodzinski, K Kwarciak, M Daniol, D Hemmerling: Improving deep learning-based automatic cranial defect reconstruction by heavy data augmentation: from image registration to latent diffusion models, Computers in Biology and Medicine, 2024 [C]. STEFANO MAZZOCCHETTI et al. : Automatic Implant Generation for Cranioplasty via Occupancy Networks, IEEE Access 2024. [D]. M. Wodzinski, M. Daniol, and D. Hemmerling, ‘‘Improving the automatic cranial implant design in cranioplasty by linking different datasets,’’ in Towards the Automatization of Cranial Implant Design in Cranioplasty II. Cham, Switzerland: Springer, 2021, pp. 29–44. [E]. Xiong, B. , Wei, S. T. , Zheng, X. Y. , Cao, Y. P. , Lian, Z. , Wang, P. S. : Octfusion: Octree-based diffusion models for 3d shape generation. In: Computer Graphics Forum. vol. 44, p. e70198 (2025) [F]. Siddiqui, Y. , Alliegro, A. , Artemov, A. , Tommasi, T. , Sirigatti, D. , Rosov, V. , Dai, A. , Nießner, M. : Meshgpt: Generating triangle meshes with decoder-only transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19615–19625 (2024) [G]. Meng, Q. , Li, L. , Nießner, M. , Dai, A. : Lt3sd: Latent trees for 3D scene diffusion. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 650–660 (2025)
- 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?
I lean for rejection mainly because of the doubts on the experimental protocol (W1-W3) and the lack of clarity on how and why the contributions are designed (W4), which in turn hinders reproducibility. Since, as per the reviewer guidelines, the paper should be evaluated as submitted without asking to substantially expand it, I feel that in its present form the paper is not ready for publication. While the rebuttal may clarify W1 and W2, it cannot address the lack of important baselines (W3) nor, even more importantly, improve the descriptions of the method (W4). However, the paper has strengths and could be the basis of a strong future submission once experiments and presentation are improved. To this end, I’d suggest to include more recent and effective generic baselines as well as methods proposed for the skull completion task in the comparison to substantiate the claim of state of the art performance and put the results in context with respect to the recent literature. I’d also suggest to describe better in the text the design of the modules and the motivations behind them, in particular of the convolution that is currently not described, and provide all the details to allow reproducibility,
- Reviewer confidence
Very confident (4)
- [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.
The rebuttal clarified that experiments have been performed in a fair way. This clears my W1 and W2 weaknesses.
Yet, the absence of relevant and problem-specific baselines is of course not solved by the rebuttal. The authors honestly acknowledge that the skull-specific baselines are competitive. Hence, it is important that they are presented in the manuscript to put the results of the proposed method in context and clearly highlight its strengths. At the moment, it looks like the proposed method is both significantly more effective and more efficient than the state of the art, while only the latter may be a correct claim after competitive baselines are considered.
Same for the clarity of the paper. Even after reading the rebuttal, I’m not sure I’d be able to replicate the method. The description is dense. Some specific doubts have not been cleared even after the rebuttal, e.g.
- “The learnable positional embedding receives flattened xy/xz/yz triplane tokens.” -> I’m not sure what it means that a positional embedding “receives” flattened tokens and how it should be implemented.
- “Attended tokens are reshaped back to three planes and fused with original features via residual addition and 2D projection” -> not sure how three planes can be fused by 2D projections. A thorough, extended description is needed. Figures alone are not enough to describe the methodology.
Moreover, an important weakness was added by R4, i.e. the lack of citations and discussions of previous work introducing triplane attention (DiffTF/DiffTF++). The authors rebutted that they didn’t claim it as a novel contribution of this submission, but the second bullet of the list of contributions in the main paper reads “A triplane-aware module is introduced to jointly model plane-wise surface continuity and cross-plane geometric alignment”. The proposed architecture should be compared and contrasted with existing ones and the contribution list be revised accordingly.
I consider these 3 weaknesses too important to be mitigated by the strong empirical results. While the submission has strengths, and I strongly encourage the authors to improve it and resubmit it, I have to recommend rejection in its current form.
Review #2
- Please describe the contribution of the paper
This paper proposes TriFlow, a conditional flow matching model for efficient and high-fidelity 3D skull shape completion. To address the high time cost of traditional CAD-based design and the difficulty existing deep learning methods face in simultaneously achieving anatomical consistency, boundary accuracy, and computational efficiency, the authors encode 3D skull geometry into compact triplane latent features and learn a conditional continuous velocity field in this latent space. To ensure spatial continuity and geometric consistency in the generated 3D shapes, the model further introduces a triplane-aware attention module, which jointly models intra-plane continuity and cross-plane geometric alignment. Experiments on the SkullFix and SkullBreak datasets show that TriFlow achieves state-of-the-art performance in reconstruction accuracy, including metrics such as DSC and HD95, while also significantly outperforming existing diffusion-based and 3D U-Net-based methods in both training time (2.8 hours) and inference speed (1.1 seconds), demonstrating strong potential for practical clinical deployment.
- 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 method is well designed and demonstrates solid structural novelty. This paper combines triplane representation with conditional flow matching in a natural and coherent way around the goal of efficient 3D completion: a) Triplane representation is used to reduce the cost of 3D shape representation; b) Flow matching replaces diffusion to enable faster and deterministic sampling; and c) The triplane-aware attention module is introduced to compensate for the cross-plane inconsistency caused by triplane decomposition.
2.The method achieves strong accuracy with excellent efficiency. On both SkullFix and SkullBreak, TriFlow outperforms the compared methods across all four metrics: DSC, bDSC, HD95, and MMD. In addition, Table 2 shows that TriFlow also surpasses all listed baselines in terms of training time, inference time, parameter count, and FLOPs.
3.Strong generalization ability. Although trained only on synthetic data, the model still maintains strong performance on the real-world MUG500 skull defect dataset, achieving a DSC of 0.911.This is particularly important for practical clinical applications.
- 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.Insufficient analysis of robustness to large-scale or complex defects. Although the model performs well on SkullBreak, which contains irregular defect boundaries, the paper does not provide a stratified evaluation across different defect sizes (e.g., small, medium, and large defects). It remains unclear whether the flow matching framework can still preserve anatomical plausibility when faced with very large cranial defects.
2.Clinical applicability still requires more comprehensive quantitative evaluation. The paper repeatedly uses expressions such as “practical clinical deployment” and “clinically reliable implant generation”. However, the main experiments are still conducted on benchmarks such as SkullFix and SkullBreak, which consist of real skulls with synthetically generated defects. External validation is also limited to only 29 craniotomy cases from MUG500.As a result, the paper does not yet provide sufficient evidence from real clinical workflows or expert evaluation by clinicians, and these claims therefore appear somewhat overstated in the current form.
- 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?
A well presented paper with solid methodological foundations and strong experimental results.
- 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.
After reading the rebuttal, I am still positive with the technical values of this work. I am happy to recommend an acceptance, but will not argue much if the chairs decide to reject it.
Review #3
- Please describe the contribution of the paper
The authors build a flow matchimg model for 3D skull shape completion, that operates in a lightweight latent representation using triplanes and a U-Net. The model directly flows from incomplete to complete skull images, and generates fixes via boolean subtraction. Final model has substantially lower train time, inference time, param count, flop and better result scores than other competitors.
- 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 model is extremely efficient, being both lightweight in params, training in 2.8 hours on only a V100, and inference in ~1 second. The latent representation is elegant, and the network is triplane-aware to integrate closely with the representation type, and the components enabling this are cleanly ablated. Other hyperparameters are also nicely ablated in Tables 4/5. Figure 4 qualitative results are very nicely and clearly presented and hugely convincing against competing methods.
- 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.
TriFlow has the best results all round in Table 1, but values are tricky to compare other than relatively with no standard deviations / uncertainty metrics or significance tests. I think this is important in identifying if the improvement is significant, especially as the numerical difference between TriFlow and the second best competitor is relatively low and the datasets have limited number of cases. Other results tables throughout this paper have this issue too.
As training is shown to be extremely fast and cheap I would have expected multiple seeds to be run to more convincingly show the results are not just from a single favorable run.
I would like details of the baseline training protocol, and affirmation that the competing methods were reimplemented and trained locally under the same conditions or disclosure that the results were taken from their respective papers / pretrained models. Further, some of the competing papers compared are from general 3D network methods or different tasks, so understanding how they were fairly adapted and hyperparam tuned to the task for comparison would be appreciated.
‘Rectified Flow for Efficient Automatic Implant Generation’ Zhou et al, do rectified flows on the same problem on the same two datasets, and should likely have been cited and run as a baseline.
Triplane attention and triplane representation could do with clear citations stating existing works so that novelty of what the paper proposes can be separated from current field status, for example papers like ‘TPA3D: Triplane Attention for Fast Text-to-3D Generation’ by Wu et al or ‘DiffTF’ and ‘DiffTF++’ by Cao et al.
Specifically, DiffTF/DiffTF++ share most of the components of this submission; triplane factorization, plane-wise self-attention, cross-plane attention with shared weights, learnable positional embeddings, and adaptive normalization from DiT. The attention block diagram and especially the encoder/decoder diagram with multi axis attention is very similar, and the lack of citation / discussion is extremely concerning, as the current framing and contributions bullet implies that this abstraction is entirely novel.
- 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.
(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 results are convincing and impressive, in all metrics, though more thorough statistical evaluation and presentation of the results would be greatly appreciated, as well as understanding of architectural precedent / what inventions are the authors’ own contributions given similar work existing.
- 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.
Rebuttal helpfully clarified the baseline retraining protocol and appropriately conceded the DiffTF(++) architectural precedent and missing Zhou et al citation; while this narrows the contribution to an applied integration of known triplane-attention components under flow matching for skull completion, the demonstrated efficiency gains, SOTA results across both datasets, and real-world generalization make it a worthwhile applied contribution, so I lean toward weak accept
Author Feedback
We sincerely thank all reviewers (Rn) for their constructive comments and for recognizing the strengths of our work, including strong results on SkullFix/SkullBreak (R1,R2,R4), efficient conditional flow matching in compact triplane latent space (R1,R2,R4), and fast training/inference (R2,R4). The raised questions are addressed below. 【Experimental protocol and reproducibility (R1.W1,R1.W2,R4)】. All baselines in Table 1 were retrained locally on the same SkullFix and SkullBreak splits, using identical preprocessing, TSDF generation, resolution, and metrics. No results were copied from original papers. We followed the official splits of SkullFix 100/110 and SkullBreak 570/100, and evaluated all methods with DSC, bDSC, HD95, and MMD. Consistent augmentation was used for all learning-based methods, including random rotation, scaling, translation, and defect perturbation when applicable. We will add the full training recipe, optimizer settings, augmentation details, and baseline adaptation protocol. 【Baseline comparison (R1.W3,R4)】. We agree that stronger skull-specific baselines should be discussed. We will add or discuss Rectified Flow for Efficient Automatic Implant Generation, Wodzinski et al. 2024/2025, Occupancy Networks-based cranial implant generation, OctFusion, MeshGPT, and LT3SD. Preliminary extended experiments show that these skull-specific baselines are competitive, while TriFlow keeps a clear efficiency advantage due to compact triplane latents and deterministic ODE sampling. 【Statistical reliability (R4)】. We repeated the main experiments with three random seeds. The standard deviations are small for DSC, bDSC, HD95, and MMD, showing that the gains are not from a favorable run. We will report mean ± std and add paired significance tests against the strongest baselines. 【Triplane-aware module and triplane convolution (R1.W4,R4)】. The triplane-aware module consists of plane-wise self-attention, cross-plane attention, residual connections, normalization, and AdaLN-based conditioning. The learnable positional embedding receives flattened xy/xz/yz triplane tokens. Attended tokens are reshaped back to three planes and fused with original features via residual addition and 2D projection. The conditional latent c modulates normalization parameters via AdaLN. For triplane convolution, AvgPool is performed along shared axes to exchange information among xy/xz/yz planes; pooled features are then broadcast or upsampled back before the final 2D convolution. We will add tensor shapes, kernel sizes, channel numbers, and residual structures. 【Architecture details and terminology (R1.W5)】. “3D RCU” denotes the 3D residual convolutional unit in the triplane autoencoder encoder, which converts the 3D TSDF volume into axis-aligned triplane features through axis-wise downsampling. The 2D ResBlock refines each triplane feature map before bilinear query sampling and MLP-based SDF prediction. “attention_axis=z” will be replaced with explicit xy/xz/yz plane-wise and cross-plane attention descriptions. 【Novelty and related work (R4)】. We will cite DiffTF/DiffTF++, TPA3D, and related triplane attention works, and clarify that triplane attention itself is not claimed as entirely new. Our contribution is adapting triplane-aware modeling to conditional skull completion under latent flow matching, where the defective skull latent guides a continuous velocity field for fast deterministic generation. 【Robustness to complex defects (R2)】. We will add stratified evaluation by small, medium, and large defects on SkullBreak. Preliminary results show TriFlow remains robust for large and irregular defects, although the margin decreases for extremely large missing regions. 【Clinical applicability (R2)】. We will soften “clinically reliable” to “potential clinical applicability.” MUG500 will be clarified as an external generalization test with 29 real craniotomy cases, while clinician evaluation, manufacturing constraints remain future work.
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 paper shows strong empirical performance and efficiency for 3D skull completion, but there are significant concerns that need clarification before acceptance. Please respond to the fairness and completeness of the experimental protocol, including whether baselines were retrained under the same conditions, why stronger skull-completion and rectified-flow baselines were omitted, and whether statistical uncertainty or multiple seeds support the reported gains. The rebuttal should also clarify the triplane-aware module and triplane convolution details, better distinguish the paper’s novelty from prior triplane attention/latent generation work, address missing or insufficient citations to closely related methods, and temper claims about clinical deployment given limited real clinical validation.
- 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 work addresses an important skull shape completion problem and proposes an efficient triplane latent conditional flow matching framework with strong empirical performance, substantial efficiency gains, and evidence of real-world generalization. The rebuttal clarified the baseline retraining protocol and appropriately narrowed the contribution in light of related triplane-attention work. While the remaining concerns about problem-specific baselines, architectural precedent, citations, and clearer methodological details should be addressed in the final version, the demonstrated accuracy and efficiency make this a worthwhile applied contribution. I recommend acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After the rebuttal, two reviewers recommended acceptance, while one reviewer maintained a Reject recommendation. While the paper has some limitations, including the lack of comparisons with some recent state-of-the-art methods, it also demonstrates several strengths, including strong real-world generalization and the technical contribution of adapting triplane-aware modeling to the proposed task. Therefore, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready 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.
This paper proposes TriFlow, a conditional flow matching framework for 3D skull shape completion using triplane latent representations. Reviewer scores are 2/5/3, with R2 accepting and R4 moving to weak accept post-rebuttal, while R1 maintaining rejection.
Reviewers broadly agree that the efficiency gains are genuine and practically meaningful, the combination of triplane representation with conditional flow matching is well-motivated, and the results are competitive across multiple datasets. The shared concerns are the missing skull-specific baselines, the novelty positioning of the triplane-aware module relative to DiffTF/DiffTF++, and insufficient method description. The rebuttal clarifies that all baselines were retrained under identical conditions, concedes the DiffTF/DiffTF++ architectural precedent, and promises to add implementation details and a more comprehensive baseline comparison in the final version.
Since the core contribution is efficiency rather than outperforming every skull-specific method on quality, and the rebuttal adequately addresses the fairness and novelty positioning concerns, the strength modestly outweighs the weaknesses. I recommend acceptance.