List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Jadie Adams, Shireen Y. Elhabian
Abstract
Statistical Shape Modeling (SSM) is a valuable tool for investigating and quantifying anatomical variations within populations of anatomies. However, traditional correspondence-based SSM generation methods have a prohibitive inference process and require complete geometric proxies (e.g., high-resolution binary volumes or surface meshes) as input shapes to construct the SSM. Unordered 3D point cloud representations of shapes are more easily acquired from various medical imaging practices (e.g., thresholded images and surface scanning). Point cloud deep networks have recently achieved remarkable success in learning permutation-invariant features for different point cloud tasks (e.g., completion, semantic segmentation, classification). However, their application to learning SSM from point clouds is to-date unexplored. In this work, we demonstrate that existing point cloud encoder-decoder-based completion networks can provide an untapped potential for SSM, capturing population-level statistical representations of shapes while reducing the inference burden and relaxing the input requirement. We discuss the limitations of these techniques to the SSM application and suggest future improvements. Our work paves the way for further exploration of point cloud deep learning for SSM, a promising avenue for advancing shape analysis literature and broadening SSM to diverse use cases.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43907-0_47
SharedIt: https://rdcu.be/dnwdt
Link to the code repository
https://github.com/jadie1/PointCompletionSSM
Link to the dataset(s)
https://github.com/jadie1/PointCompletionSSM
Reviews
Review #1
- Please describe the contribution of the paper
Statistical Shape Models (SSMs) require establishment of point correspondences in a separate step prior to learning the SSM. Traditional point correspondence establishment methods need knowledge about the topology of the shape in form of a binary volume or mesh. This work proposes to use neural networks for point completion with unordered point clouds of a shape as input only.
- 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.
- using point completion networks for SSM generation is an interesting idea and is promising to solve some problems of existing correspondence establishment methods like the need to re-optimise correspondence when new shapes are added to a dataset.
- evaluation shows that the concept works decently for a variety of different shapes
- the method can cope with missing information in the input shape to some degree
- 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.
- the test point clouds are all generated from smooth surfaces without outliers which is not realistic for some of the mentioned target application areas like ultrasound imaging and 3D scanning. It would have been interesting to see the impact here
- one of the highlighted benefits of the method is to work with limited data. However, the method struggles with rather simple shapes like the spleen when small datasets are used. In such cases, traditional methods are likely the better choice
- 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 paper is using existing methods which allows for easy reproduction.
- 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
The paper is well written and the concept is clear and to my knowledge novel. The evaluation could be improved by comparing the SSM quality to traditional PDMs which would give a better indication of the performance. Furthermore, more realistic point clouds for the highlighted application areas including outliers could be used. A main benefit of the proposed method is that it works with unordered point clouds and that no meshing is required. However, the used test point clouds are all from smooth surfaces which are quite easy to tesselate. In terms of performance, the results for the pancreas are quite good which indicates that the method also works on complex shapes.
- 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?
I think the work would be interesting to the community as the use of unordered point clouds for SSM generation has good potential. The proposed method leaves some questions in terms of generalisability, but I think the results are promising enough for further work in that area.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
In this work, they demonstrate that existing point cloud encoder-decoder-based completion networks can provide an untapped potential for SSM, capturing population-level statistical representations of shapes while reducing the inference burden and relaxing the input requirement. Extensive experiments demonstrate that point cloud networks can learn accurate SSM of anatomy when provided with a sufficiently large and representative training dataset.
- 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.
- Extensive experiemnts are conducted on left atrium, spleen and Femur datasets, outperforming SOTA methods.
- This paper is well-written and well-organized.
- More discussion to make it applicable is explored.
- 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.
- Too trivial to discuss in the final chapter.
- Table 1 can be decomposed to make it more clear.
- 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 reproducibility of the paper is fair. Release of code is recommended.
- 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
The paper is well written. The attempt to broaden SSM to diverse use cases is encouraged.
- Table 1 is recommended to be more clear and comparable.
- Discussion part can be conclusive in the final chapter.
- 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?
Clear idea and novel contributions: the methods accurately represent shapes via uniformly distributed points constrained to the surface and provide good correspondences that capture population-level statistics. Rich literature review and exploration to make it applicable.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #4
- Please describe the contribution of the paper
This paper investigates variants of PCNs (Point Completion Networks) in generating PDMs (Point Distribution Models). For the selected PCNs, the authors applied the methods to both a simulated dataset consists of ellipsoids and multiple real anatomic datasets. The results show that it is possible to generate PDMs via the general PCNs.
- 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 paper shows the effectiveness of general point cloud completion models in the field of generating SSM for anatomic shape analysis. The paper focuses on important issues in anatomic shape analysis; i.e., good correspondence across a population is critical in analysis. To show the correspondence resulting from PCN, this paper shows the PDMs of simulated ellipsoids. Also, the metrics including compactness, generalization and specificity are helpful in understanding the quality of the PDMs’ correspondence. • The paper notes that the bottleneck features from an encoder capture population-level properties that are consistent across a population. The coarse output is taken as the PDM. • The paper is well written. The experiments provide qualitive and quantitively visualization of various models with various datasets. Moreover, these datasets help to reveal the properties of the chosen models from various perspectives. E.g., the Pancreas dataset shows the properties of the chosen models taking in different amount of boundary points.
- 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.
• The paper shows experiments with existing methods, lacking novel contributions to the field of constructing PDMs for anatomic shape analysis. In fact, the paper discuses promising future works, some of which can be done and significantly boost the novelty of this work. • The paper lacks deep interpretation of the various methods chosen to construct PDMs. From table 1, different methods have different advantages across the chosen datasets. It is of interest to understand how to choose from various methods to generate PDMs given a dataset. Additionally, the discussion section compares PCN (MLP-based) and SFN (transformer-based), saying that “Interestingly, the simplest model, PCN[31], achieved similar, and sometimes better, SSM accuracy than more current state-of-the-art methods, despite its inferior performance in point completion benchmarks”. It would be good to connect the observation with the architectures of the model, looking into a better explanation of how these models perform in generating SSM. Finally, I am curious about the significant performance difference between SFN and PointAttN (attention-based), expecting more discussion.
- Please rate the clarity and organization of this paper
Excellent
- 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 conduct experiments on several datasets that are not public. The supplementary material gives more details about the datasets. The implementation should be accessible in each method chosen in this paper.
- 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
- The paper can enhance the novelty by improving objectives, as suggested in the discussion section.
- The paper can deep dive into a couple of methods (e.g., PCN and SFN) to show the depth of this research.
- The research can show the effectiveness of SSM produced by a general point completion method (e.g., PCN) in anatomic shape analysis (e.g., hypothesis testing).
- It would be good to connect the general coarse-to-fine architecture (figure 1) with the specific methods. The manuscript can provide more insights based on Figure 1, including the variants of encoder/decoder, how the refinement network works.
- In the simulation study, the authors can wrap/twist the ellipsoids to add more modes of variation. As the morphological variation of anatomic shapes is often subtle and complicate, the simulation study can be more convincing to include more complicate shape variation than stretching.
- 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?
This paper is very well written and aims toward an important direction in anatomic shape analysis. However, the paper lacks (1) novel contributions in methodology and (2) deep understanding of the chosen methods in producing SSM. The paper can have a better rank assuming the improvement of either of the two aspects.
- Reviewer confidence
Very confident
- [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
The merits of this paper slightly weigh over its weakness. Merits:
- This paper studies applications of various networks in building SSM from point clouds. The evaluation of resulting SSMs covers both individual point accuracy and statistical properties (i.e., compactness, generalizability and specificity), see table 1.
- This paper is well-written. The figures show rich information that is of interest. E.g., in Figure 4, the top-left figure shows best, median and worst case together. Also, the bottom row in figure 4 shows that the training size of target cohort can be large (i.e., 1000).
- In the rebuttal, the authors more clearly compared the studied methods in the paper with traditional methods. This comparison highlights the significance of this work. Weakness:
- No new method is proposed in this paper. In fact, the discussion section mentioned a few possible directions to improve the novelty (e.g., modify the training objective). The paper can have a better rank by including some of those improvements.
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.
While two reviewers viewed the idea behind general point cloud completion favorably. One reviewer pointed out that the paper lacks novel contributions. Instead applications using solutions of point completion, SSMS generation etc. are “promised” as future works.
After going through the paper in detail, it seems that the contribution of the paper is in evaluation different networks that predict point clouds, that perform point cloud completion, and those that perform variations on the above. Thus the paper does not propose new methods. While evaluation and ideas-based papers should indeed be promoted, provided they make significant contributions to the field or propose some novel ideas, it is not clear that this paper achieves that. One of the reviewer also points out that “the method struggles with rather simple shapes like the spleen when small datasets are used. In such cases, traditional methods are likely the better choice”.
The paper also does not provide appropriate context in evaluation and the choice of the various methods selected to construct point cloud models. This point is also brought up by a reviewer who was also intrigued by “difference in performance between SFN and PointAttN (attention-based)”.
The paper does not adequately differentiate between the point cloud construction and the statistical shape model that follows the construction. The PDMs/SSMs are a step beyond the initial prediction/construction. Thus table 1 compares existing point cloud models. However, since the authors make a case about statistical shape models, they should also evaluate the resulting PDMs after the initial step. This is currently not done systematically.
The actual proposed problem “learning statistical shape models from point cloud anatomies” is an interesting and fundamental one. This has implications for sequential or online learning, as well as incremental model updating, however, none of these ideas are discussed or at least evaluated in the paper, limiting its impact. Thus based on the evaluation results from this paper, one could simply used this approach for point cloud completion/prediction, but stil use classical models to compute SSMs, which will be significantly faster. This should be discussed.
Author Feedback
We appreciate the reviewers’ feedback and the opportunity to respond. Point completion entails predicting an unordered complete point cloud from a partial one, while correspondence-based SSM requires generating ordered correspondence points. Previous studies have not assessed point completion networks on SSM. To clarify for the meta-reviewer, this approach is end-to-end, directly predicting the PDM without a second step. The suggestion to first use our approach for point completion and then utilize classical models to compute SSMs would not be faster or feasible since classical methods cannot operate on point clouds (complete or partial).
Regarding the novelty of our contributions, our paper is an “application studies” track, not focused on methodological advancements. Our primary contribution lies in applying and benchmarking existing point completion methods on a new task: correspondence-based SSM of anatomy. We reveal a significant, unnoticed outcome of coarse-to-fine point completion networks: ordered coarse output that provides correspondence. We apologize if we did not make it clear what a significant contribution this is. Predicting correspondence-based SSM from point clouds has significant benefits over traditional methods:
- It only requires point cloud input, significantly expanding the potential use cases of SSM. Traditional methods require complete, noise-free shape representations, which are time-consuming and costly to acquire. Point clouds are readily accessible in biomedical applications via lightweight acquisition techniques such as thresholding CT scans and anatomical surface scanning. Although we evaluated only smooth point cloud inputs, similar architectures have shown success in point cloud denoising tasks, suggesting that our approach may handle noise as well.
- Our solution is scalable with fast inference. As mentioned in the discussion and highlighted by Reviewer #1, traditional methods might outperform our approach given a very small training cohort. However, given a large cohort such as the left atrium, this approach is accurate and much faster than optimization-based methods.
- Our approach enables simultaneous SSM prediction and completion. Traditional methods cannot be applied when dealing with partial observations.
- As highlighted by the meta-reviewer, our approach enables sequential or online learning, as well as incremental model updating. This feature makes it more applicable to real clinical scenarios where shapes are collected over time, allowing for iterative updates and fine-tuning of the model.
- This data-driven approach avoids biases introduced by metrics and parametric representations used in classical methods. If accepted, we will make these benefits clear in the revision. This is the first work to provide a solution for SSM from point clouds. Although the networks are not novel, we believe this is an important contribution that paves the way for further exploration and will be of great interest to the community.
We’d also like to address the feedback regarding the lack of deep interpretation of various method’s performance. As there is no universal backbone for point cloud networks, we selected a breadth of SOTA models to compare in evaluation. SFN utilizes k-Nearest Neighbors (kNN) to capture geometric relations in the point cloud, while PointAttn does not. PointAttn showed this was advantageous for point completion of man-made objects such as chairs and lamps, where kNN information could be misleading. However, in the task of anatomical SSM, it is likely that the kNN information assisted SFN by providing accurate spatial shape information, given the more convex shape of organs and bones. We aim to provide insights into models strengths and weaknesses to guide future research and will extend this discussion appropriately.
We appreciate the reviewer’s thoughtful comments and will improve the clarity of the contributions and the model performance discussion in revision.
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 rebuttal addressed the reviewer comments. The evaluation strategy of the networks used for building the SSM from point clouds was deemed as a strength. While the paper did not propose any new methods, the results, comparisons, and discussion of the other methods were all perceived as strengths.
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.
I am not sure whether the rebuttal regarding the “application studies” track convinces me, as “application studies” should demonstrate the impact or clinical value of the techniques with evaluations on large datasets or in-human feasibility studies.
However, the rebutal has satisfactorily addressed my all other concerns, in particular the motivation of this work. Overall, I think this is an interesting evaluation work on anatomical shape modelling that is worth discussing in the MICCAI conference.
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 authors response addresses issues about deep interpretation of various method’s performance as well as the novelty. I suggest we ask the authors to revise the paper according to their response letter and comments by the reviewers and the original AC.