List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Paul Friedrich, Julia Wolleb, Florentin Bieder, Florian M. Thieringer, Philippe C. Cattin
Abstract
Advances in 3D printing of biocompatible materials make patient-specific implants increasingly popular. The design of these implants is, however, still a tedious and largely manual process. Existing approaches to automate implant generation are mainly based on 3D U-Net architectures on downsampled or patch-wise data, which can result in a loss of detail or contextual information. Following the recent success of Diffusion Probabilistic Models, we propose a novel approach for implant generation based on a combination of 3D point cloud diffusion models and voxelization networks. Due to the stochastic sampling process in our diffusion model, we can propose an ensemble of different implants per defect, from which the physicians can choose the most suitable one. We evaluate our method on the SkullBreak and SkullFix datasets, generating high-quality implants and achieving competitive evaluation scores. The project page can be found at https://pfriedri.github.io/pcdiff-implant-io.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43996-4_11
SharedIt: https://rdcu.be/dnwOL
Link to the code repository
https://github.com/pfriedri/pcdiff-implant
Link to the dataset(s)
https://www.fit.vutbr.cz/~ikodym/skullbreak_training.zip
https://files.icg.tugraz.at/f/2c5f458e781a42c6a916/?dl=1
https://www.sciencedirect.com/science/article/pii/S2352340921001864
Reviews
Review #1
- Please describe the contribution of the paper
The paper analyzes the usage of diffusion probabilistic models on point clouds for the generate of cranial implants for cranioplasty. The work builds on top of two recent challenges AutoImplant 2020 and 2021 and relies therefore on a solid reference base.
- Please list the main strengths of the paper; you should write about 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 rely on sparse point cloud instead of voxel grids as previous works to reduce the memory footprint.
- It is the first method to evaluate diffusion models on point cloud for cranioplasty
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- Although a large number of methods were suggested and evaluated during the AutoImplant challenges, the authors only compare against 2D/3D U-Net.
- It is not clear why the authors rely on voxelization during training instead of relying on oriented point clouds.
- The suggested method does not substancially improve the state of the art performance according to the reported evaluations.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The authors provide the reviewers with an anonymous link with the implementation and rely on public data for training.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- There appears to be a journal article summarizing one of the challenges [1]. The authors could refer the reader to this article in the background for further information.
- The authors should comment on why they rely on unoriented point clouds and not on oriented point clouds. For oriented point clouds the reconstruction of a triangulated surface can be easily done with approaches like Poisson Reconstruction.
[1] Li, J., Pimentel, P., Szengel, A., Ehlke, M., Lamecker, H., Zachow, S., Estacio, L., Doenitz, C., Ramm, H., Shi, H. and Chen, X., 2021. AutoImplant 2020-first MICCAI challenge on automatic cranial implant design. IEEE transactions on medical imaging, 40(9), pp.2329-2342.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although a large number of methods were suggested and evaluated during the AutoImplant challenges, the authors only compare against 2D/3D U-Net. Also, the suggested method does not substancially improve the state of the art performance according to the reported evaluations.
However, the suggested approach is interesting and might be of interest for the whole community.
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
This paper employs 3D point cloud diffusion models for an automatic patient-specific implant generation task.
The authors evaluate our method on the SkullBreak and SkullFix datasets, generating high-quality implants and achieving competitive evaluation scores.
- Please list the main strengths of the paper; you should write about 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.
Overall, the paper is well-organized and well-written. The topic is interesting and has practical significance
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
Firstly, the proposed method lacks technical innovation as it only applies the diffusion model to the implant generation. The paper should provide a clearer explanation of how this approach differs from other existing methods.
Secondly, the results presented in Table 1 do not demonstrate the superiority of the proposed method. The paper should provide a more detailed comparison with other state-of-the-art methods and discuss the strengths and weaknesses of the proposed method.
Thirdly, it would be useful if the authors could provide information about the time required for the entire generation process. This information would help readers understand the practicality of the proposed method.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
Very Good.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
Please see the comments on the main weaknesses of the paper.
Also, the authors should do more expriments to evaluate the effeciveness of the proposed method.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Good paper writting. New application.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [Post rebuttal] Please justify your decision
I keep my decision
Review #3
- Please describe the contribution of the paper
This paper presents a novel approach for the contribution of the paper based on a combination of a point cloud diffusion model and a voxelization network.
- Please list the main strengths of the paper; you should write about 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.
- This paper uses a new method (point cloud diffusion model) to achieve automatic implant generation.
- The paper organization is simple, clear and easy to understand.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- There is no methodological innovation in this paper.
- The experimental results are not as good as 3D-UNet, and the experimental design is not convincing.
- Without further explaining the advantages of the diffusion model, why is the diffusion generation model suitable for this task?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
yes
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
please refer to weaknesses.
It is suggested that the author add more comparative experiments to verify the claims.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I am working on diffusion model about three years.
- Reviewer confidence
Somewhat confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
4
- [Post rebuttal] Please justify your decision
Thanks for the author’s response, but I didn’t see substantial improvement, so I’m sticking with the original score.
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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 manuscript received mixed scores based on quite similar concerns.
The overall strengths are appreciated as a well written paper with a method is the first to apply point diffusion to this specific application, and that effectively reduces the memory footprint.
There are, however, several concerns. From my perspective by far the most important concern (echoed in slightly different ways across all reviewers) is the limitation in both, innovation and experimental validation, including the overall performance of this method compared to the limited baselines. Reviewers suggest additional experiments, which is discouraged based on this year’s MICCAI policy. Overall, I echo these concerns and urge the authors to provide as much clarification as possible.
There are additional minor concerns and other clarifications that should be addressed as possible in the rebuttal.
Author Feedback
We would first like to thank all Reviewers (R1-R3) and the Area Chair (AC) for their valuable feedback and insightful comments. Below, we will address the main concerns and clarify some issues. (R2, R3, AC) Concerns about the method’s novelty: We present a novel non-deterministic approach for automatic implant generation that can generate multiple anatomically reasonable implant designs for a single defect. This has not yet been possible with previous state-of-the-art methods and allows the physician to choose from different implant designs. We combine a sparse point cloud representation of the skull, significantly reducing the memory footprint while preserving full context awareness, with a simple yet effective method for voxelization. Our method is straightforward to train and produces high-quality implants for various defects. To the best of our knowledge, we are the first to successfully employ a point cloud diffusion model to a medical application. (R3) Concerns about the experimental design & (R1, AC) the comparing methods: To ensure high comparability, our experimental setup and evaluation metrics are deliberately aligned with both AutoImplant challenges. We retrained 3 approaches on the publicly available datasets used in those challenges. We chose to compare to the current benchmark [30], which won the AutoImplant 2021 challenge by a significant margin. Since we mention the memory efficiency of our approach, we added a method that takes a different memory reduction approach by using sparse convolutions [10]. For comparing 2D and 3D methods, we added the currently best-performing 2D approach [32]. We have not added more comparisons with approaches that have clearly been outperformed in the mentioned challenges. (R2, R3, AC) Concerns about the reached evaluation scores & (R1) no substantial improvement on the state-of-the-art performance: The evaluation scores presented in the paper are generally at a high level. Our method produces very competitive results, i.e. it mostly achieves equal or marginally lower, but also better results compared to [30]. By visually comparing the implants, we find that our method produces the most realistic surfaces that are closest to the ground truth. This property is not reflected in the chosen metrics, which is why we provide visual comparison (Fig. 4). Regarding R1’s concern, it remains questionable whether a significant improvement on the state-of-the-art performance can even be achieved given the anatomical variation of the underlying population. (R1) Why we chose voxelization rather than oriented point clouds: While using oriented point clouds is a very interesting idea, we combined the well-performing diffusion models on unoriented point clouds with the voxelization network. The voxelization network offers the advantage of learning to set the normal vectors of possible outliers to zero, effectively ignoring those points for the Poisson surface reconstruction. This makes our method robust to potential noisy output from the diffusion model. Furthermore, it is not clear or straightforward how the noising and denoising process on oriented point clouds would need to be defined. This would need to be explored in future work. (R1) Additional reference: We agree that the journal article mentioned by R1 is of value to readers of our paper and will add a reference to our final version. (R2) Generation time: The entire generation process takes about 20 min, which is negligible compared to the time required to 3D printing or milling an implant. We will add runtime information to the final version of the paper. (R3) Why a diffusion model suits the task: There are 2 main reasons for choosing a diffusion model: Firstly, diffusion models have shown good performance in point cloud completion [18,36]. Secondly, their stochastic sampling process allows for the generation of multiple reasonable implant designs for a single defect, one of the main advantages of our method compared to other approaches.
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The most
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
Overall this is a very borderline paper as indicated by the reviewers. However, overall reviewers thought the strengths of the paper outweighed its limitations.
The main critiques were focused on not improving on the SOTA based on quantitative measures, which as the authors highlighted are not necessarily the only way to evaluation methods. Although being able to produce a few qualitative figures showing more plausible shapes in a handful of examples is not conclusive, in the future I would recommend trying to obtain expert ratings of the implantations plausibility or some other form of assessment to address this more directly.
The concerns over novelty were not fully addressed, just stating a method is novel is not enough you should explain in detail how your method differs from similar works.
I did appreciate the clarification on the comparisons, and I think aligning comparisons to an existing challenge (AutoImplant 2021) both strengthens the manuscript and alleviates concerns on this point.
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The innovation presented in the paper is the application of the modified point cloud diffusion model to a specific medical field: the generation of cranial implants for cranioplasty. The topic is interesting and has practical significance. The authors’ response in the rebuttal has addressed some of the concerns.