List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Savinien Bonheur, Franz Thaler, Michael Pienn, Horst Olschewski, Horst Bischof, Martin Urschler
Abstract
Since their introduction by Sabour et al., capsule networks have been extended to 2D semantic segmentation with the introduction of convolutional capsules. While extended further to 3D semantic segmentation when mixed with Convolutional Neural Networks (CNNs), no capsule-only network (to the best of our knowledge) has been able to reach CNNs’ accuracy on multilabel segmentation tasks. In this work, we propose OnlyCaps-Net, the first competitive capsule-only network for 2D and 3D multi-label semantic segmentation. OnlyCaps-Net improves both capsules’ accuracy and inference speed by replacing Sabour et al. squashing with the introduction of two novel squashing functions, i.e. softsquash or unitsquash, and the iterative routing with a new parameter free single pass routing, i.e. unit routing. Additionally, OnlyCaps-Net introduces a new parameter efficient convolutional capsule type, i.e. depthwise separable convolutional capsule.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16443-9_33
SharedIt: https://rdcu.be/cVRyN
Link to the code repository
https://github.com/savinienb/OnlyCaps-Net
Link to the dataset(s)
https://www.isi.uu.nl/Research/Databases/SCR/
http://www.sdspeople.fudan.edu.cn/zhuangxiahai/0/mmwhs/
Reviews
Review #1
- Please describe the contribution of the paper
Authors propose OnlyCaps-Net for 2D and 3D multi labels semantic segmentation for improving accuracy and inference speed by replacing the squashing function with Softsquash and Unitsquash function with introducing unit routing, paramter free single pass routing mechanism. Authors also propose a new convolutional capsule, depthwise convolutional capsule. Authors evaluate proposed method on public datasets using Dice Similarity score.
Contibutions are as follows: -Implementing two novels squashing functions softsquash and unitsquash. -Introduce unit routing, a parameter-free single pass routing mechanism -Comparing with sota public method on public dataset -New convolutional capsule, depthwise convolutional capsule.
- 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.
-Implementing two novels squashing functions softsquash and unitsquash. -Introduce unit routing, a parameter-free single pass routing mechanism -Comparing with state of art public method on public datasets -New convolutional capsule, depthwise convolutional capsule.
- 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.
-Only using one metric for comparison. Authors could use Hausdorff, accuracy, sensitivity and specificity metrics beside DICE metric because using just one metric such as DICE is not very reliable for showing segmentation results quantitatively. -Authors can update one reference for capsule paper with a newer one because it fits more to medical field. “Capsules for biomedical image segmentation” https://www.sciencedirect.com/science/article/abs/pii/S136184152030253X
- 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
Work can be reproducible.
- 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/2022/en/REVIEWER-GUIDELINES.html
Authors could use more metrics for more fair comparison such as Hausdorff, accuracy, sensitivity and specificity.
- 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Authors propose a new way for optimizing current capsule network on public dataset with novel squashing functions such as softsquash and unitsquash, and compare their work with another public work from literature.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
6
- [Post rebuttal] Please justify your decision
the authors addressed all my concerns.
Review #2
- Please describe the contribution of the paper
The contribution of this paper is 3-fold: • the authors combine separable depthwise convolution to reduce the complexity of capsule networks, • they provide two new squashing functions, • they introduce a parameter free single pass routing mechanism which does “normalization”.
- 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 three contributions seem interesting liste above, in particular the gain in matter of speed and memory consumption (while preserving the accuracy).
- 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.
I do not see real weaknesses in this paper, except that the given formulas are not easy to read for someone which is not expert in capsule networks and should be completed with parentheses for operators priorities (see formula (1)).
- 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
This paper seems reproducible since the formulas are given (even in not always very clear to me).
- 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/2022/en/REVIEWER-GUIDELINES.html
I do not have any particular remark to do on this paper: references are given, formulas are clear (except (1)) the notion of capsule network is given with its comparison versus standards CNN’s. Perhaps that a comparison with separable convolutions (used in some CNN’s) should be done since the concepts seem to be similar.
- 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?
The major factor which decides me is the gain of speed and the reduced memory consumption compared to usual capsule networks.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Somewhat Confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
The proposed method proposes an optimized capsule network architecture from state of the art for 2D and 3D segmentation. The authors implemented some efficiency strategies, such as Separable Convolutions and approximation of the Squash-Softmax functions for Dynamic Routing, to speed up the training process and limit the model memory footprint and could be of interest for the community. The method was tested on different datasets for 2D and 3D image segmentation, and the results were validated with 3-fold cross-validation.
- 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 rationale of the method is clear
- The performance of the method are validated using a cross-fold validation
- Ablation study
- The improvements introduced in this work can be adapted to 2D and 3D applications
- 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.
- Introduction of optimisation strategies known in the literature to a state-of-the-art architecture.
- The contribution in terms of performance compared to state of the art is limited.
- Unclear savings of memory footprint of the current method compared to state of the art.
- No visual examples for 2D segmentation.
- Lack of statistical tests to strengthen the discussion of results
- 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
The authot won’t release any code or dataset
- 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/2022/en/REVIEWER-GUIDELINES.html
The work does not address any clinical problem directly. However, it is interesting from an implementation point of view. However, it is not sufficiently justified, especially when looking at the results, which are numerically comparable to existing state-of-the-art methods. In the absence of significantly better results than state of the art, the contribution would have been complete if the author had provided an in-depth analysis of the impact in terms of computational cost and memory of this type of model. An analysis of variability through graphs (e.g. boxplots) and more comparison images for 2D and 3D segmentation results would have provided more insights into the discussion of the results obtained. One of the aims of the work is to reduce the memory footprint of a MatwoCapsNet-based capsule network (cost 2z^2) to z, where is the achievement of this aim demonstrated?
- 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
3
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although it does not solve any critical problems, it is, in my opinion, a topic of interest to the community to try to optimise the resource cost of complex models to make them more accessible. Unfortunately, these observations have not been satisfactorily discussed by the author.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
2
- 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
6
- [Post rebuttal] Please justify your decision
The authors addressed most of my concerns. I think a more in-depth analysis is worth to be done, but the limited amount of space makes this problematic.
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.
Reviewers think that the paper has substaintial merits. In particular, authors introduced a novel separable depthwise convolution to reduce the capsnets’ computation. In addition, they developed a non-parametric single pass routing mechanism. However, reviewers also pointed out significant weaklesses related missing memory saving analysis, lack of visual segmetation examples, and some concerns over using of only one metric for comparison. After considering all reviewers’ comments, the area chair would like to invite authors to submit a rebuttal addressing the reviewers’ concerns.
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
18
Author Feedback
First of all, we would like to thank the reviewers for the time and effort they put into reviewing our paper as well as for the feedback they provided.
We would like to thank Reviewer 1 for suggesting updating our citations to better fit the model application. To address the concern of the Area Chair and Reviewer 1, we re-performed our comparison with 3D-UCaps. We therefore complemented our Dice metrics with precision, sensitivity, standard distance, and Hausdorff distance. Moreover, we added the statistical significance of those metrics as per Reviewer 3 request. Following the suggestion of Reviewers 1 and 2, we updated the method section to allow for an easier understanding of the paper’s equations.
We would like to address a possible overlooking of Reviewer 2 and assure it that a comparison of our Depthwise Separable Capsule Convolution and CNN’s depthwise convolution is provided in Table 1 under the name ”U-Net Depth”.
Despite our mistake in the MICCAI submission formular, we would like to assure Reviewer 3 that our code will be publicly available. We would like to correct a mistake from Reviewer 3 in the routing description as we do not approximate the squash-softmax function to improve the dynamic routing algorithm. Indeed, we replace all-together the dynamic routing algorithm (iterative in nature) by a novel single pass routing algorithm. As well, we would like to indicate that we introduced two novel squashing functions rather than approximate the Sabour et al. squashing function. Unfortunately, we were unable to add more 2D and 3D segmentation results or boxplots analysis as per Reviewer 3 wishes, due to the conference 8 pages limitation. However, we did, as pointed out by Reviewer 3, add a memory and speed comparison of a 3D-UCaps modified with our unit squashing, unit routing, and depthwise separable capsule convolution. This comparison shows that not only does our method conserve the performances of the original 3D-UCaps while reducing parameters counts by ≈17%, but, as well, reduce memory by ≈35% and increases the training speed by ≈29%. Reviewer 3 also pointed to our lack of direct comparison of memory usage between our network and Matwo-CapsNet. We did not include such comparison on the JRST dataset as it can be calculated that, for a single x,y location, and at any given layer, our OnlyCaps-Net needs to store 7x6=42 digits (number of capsules x quantity of digits per capsule). In comparison, for a single x,y location and with their minimal amount of capsules per layer, a Matwo-CapsNet’s capsule layer requires at least 5x(4x4+5x5)=205 digits (a ≈ ×4.9 greater memory consumption for Matwo-CapsNet). Lastly, Reviewer 3 commented ”Introduction of optimization strategies known in the literature to a state-of-the-art architecture.”. We believe it to be referencing our novel depthwise separable convolutional capsule. Although we do draw parallels with CNN’s depthwise separable convolution, we believe our approach to be the first to adapt this method to Capsule’s Convolution.
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.
Reviewers think that the paper has substaintial merits. In particular, authors introduced a novel separable depthwise convolution to reduce the capsnets’ computation. In addition, they developed a non-parametric single pass routing mechanism. The rebuttal has sufficiently addressed reviewers’ concerns. The AC recommends to accept this paper.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
NR
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.
Authors proposed novel components for capsule, the major concern before rebuttal is the justification, and authors successfully address reviewers’ questions.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
4
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.
This paper has received positive reviews from the reviewers who acknowledge the novelty of the work and contributions, and succesful addressment of the weaknesses during the rebuttal. I consider this work is well suited for MICCAI.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
3