List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yanda Meng, Xu Chen, Hongrun Zhang, Yitian Zhao, Dongxu Gao, Barbra Hamill, Godhuli Patri, Tunde Peto, Savita Madhusudhan, Yalin Zheng
Abstract
Glaucoma is a chronic eye disease that permanently impairs vision. Vertical cup to disc ratio (vCDR) is essential for glaucoma screening. Thus, accurately segmenting the optic disc (OD) and optic cup (OC) from colour fundus images is essential. Previous fully-supervised methods achieved accurate segmentation results; then, they calculated the vCDR with offline post-processing step. However, a large set of labeled segmentation images are required for the training, which is costly and time-consuming. To solve this, we propose a weakly/semi-supervised framework with the benefits of geometric associations and specific domain knowledge between pixel-wise segmentation probability map (PM), geometry-aware modified signed distance function representations (mSDF), and local boundary region of interest characteristics (B-ROI). Firstly, we propose a dual consistency regularisation-based semi-supervised paradigm, where the regional and marginal consistency benefits the proposed model from the objects’ inherent region and boundary coherence of a large amount of unlabeled data. Secondly, for the first time, we exploit the domain-specific knowledge between the boundary and region in terms of the perimeter and area of an oval shape of OD & OC, where a differentiable vCDR estimating module is proposed for the end-to-end training. Thus, our model does not need any offline post-process to generate vCDR. Furthermore, without requiring any additional laborious annotations, the supervision on vCDR can serve as a weakly-supervision for OD & OC region and boundary segmentation. Experiments on six large-scale datasets demonstrate that our method outperforms state-of-the-art semi-supervised approaches for segmentation of the optic disc and optic cup, and estimation of vCDR for glaucoma assessment in colour fundus images, respectively. The implementation code is made available.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16440-8_50
SharedIt: https://rdcu.be/cVRwB
Link to the code repository
https://github.com/smallmax00/Shape_aware_Weakly-Semi_ODOC_seg
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper discusses the segmentation of glaucoma with an hybrid supervision setting, mixing weak labels and no labels.
The idea is to regularize the training with regularization that taps into anatomical priors about the problem, such as the shape and diameter relationship of the two circles to be segmented.
- 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 is overall well written and well motivated, and the topic is of interest.
The dual prediction and regularization is interesting.
- 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.
Adding a clear example of the weak labels would help the reader to understand exactly the supervision setting.
Though the handling of the multi-classes is not clear and hamper the comprehension of the paper
- 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 code is already shared and the data downloadable. The documentation seems enough to be able to re-run their experiments.
- 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
How would the method handle multi-class setting? I have troubles to understand the motivation of Eq (2), and I absolutely do not see how it could be generalized to multi-class. Which is a problem here since we have three classes to segment, so I am a bit lost.
For equation (6), would it be possible to use “default” or “safe” values for unlabeled samples?
- 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?
Despite the shortcomings (which are fixable in the camera ready), I believe that the paper (both method and its application) has value and is of interest to the community.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- 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 #2
- Please describe the contribution of the paper
The paper proposes a semi-supervised framework for optic disc and cup segmentation via a dual-task level of geometric consistency between pixel-wise segmentation mask with distance map.
- 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 proposes a dual task which injects geometric consistency between the pixel-wise segmentation and distance map;
- It proposes a differentiable estimation layer to predict the vCDR directly without offline post-processing.
- 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 motivation and the contribution of the proposed mSDF is not clear. In the previous work, e.g., [15], SDF is adopt to the represent the target mask. In this paper, it proposes a modified mSDF. What’s the difference between the proposed mSDF with the previous work? And what’s the advantage of the proposed mSDF compared to the previous work?
- The experimental setting is not clear. Did it train a single model using both the labeled SEG dataset and unlabeled UKBB dataset or train separate models?
- The way to utilize the labeled images and unlabeled images is not clear. It mentioned that in one batch, it has both labeled images and unlabeled images. Does it have the same training loss for both the labeled and unlabeled images?
- 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
Meet the requirement.
- 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
See the weaknesses.
- 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?
Even though the paper brings some new ideas in semi-supervised segmentation and vCDR prediction, the technique contribution is limited.
- Number of papers in your stack
7
- What is the ranking of this paper in your review stack?
3
- 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 #3
- Please describe the contribution of the paper
The authors proposed a weakly/semisupervised framework with the benefits of geometric associations and specific domain knowledge between pixel-wise segmentation probability map (PM ), geometry-aware modified signed distance function representations (mSDF), and local boundary region of interest characteristics (B-ROI ). Experiments on six large-scale datasets demonstrated that their method outperform state-of-the-art semi-supervised approaches for segmentation of the optic disc and optic cup, and estimation of vCDR for glaucoma assessment in color fundus images, respectively.
- 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 research is of practical significance and can solve the problem of less labeling in medical image segmentation. In addition, the manuscript is well organized and fluent in language.
- 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 don’t think the manuscript has any major weaknesses.
- 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
I think the method can be replicated, and the author has provided the code.
- 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
- It can be added in the abstract that how much the performance of the proposed method is better than that of SOTA.
- Why is the data size set to approximately 1:2 for the weak supervised set and the test set allocated in the UKBB data set?
- In each batch of the training data, must the proportion of labeled data and unlabeled data be 1:1?
- In the experiment, what is lambda in Eq.(9) set to? Why?
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The research has a positive effect on computer assisted diagnosis, and the method is novel, the results are promising, and the writing is fluent. Vertical cup to disc ratio (vCDR) is an important observation factor in the diagnosis of glaucoma, the usual method in the computer-aided diagnosis is to segment the optic cup and disc region first and then calculate the vCDR. The optic disc and cup segmentation via deep learning can obtain positive results, however, these deep learning methods need a large size of training data. This paper designed a framework which can train a segmentation model by using small size data of optic disc and cup labels and large size data of vCDR labels. This is an effective weakly/semi-supervised approach and is reasonably interpretable. In my opinion, this study is more enlightening to the field of optic cup optic disc segmentation.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- 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
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.
All the reviewers have found value in this paper for the community as well as from the clinical use case. It shows a good exploitation of anatomical priors as well as providing a differentiable loss that exploits the vCDR supervision signal, which has not been done before. In addition, the paper was found to be well written and the availability of the source code was helpful. Nevertheless, further clarification would be helpful, primarily regarding the multi-class setup as indicated by R1, as well as the motivation and the contribution of the proposed mSDF loss as indicated by R2
- 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).
1
Author Feedback
N/A