List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Zhihong Lin, Danli Shi, Donghao Zhang, Xianwen Shang, Mingguang He, Zongyuan Ge
Abstract
Recent studies have validated the association between cardiovascular disease (CVD) risk and retinal fundus images. Combining deep learning (DL) and portable fundus cameras will enable CVD risk estimation in various scenarios and improve healthcare democratization. However, there are still significant issues to be solved. One of the top priority issues is the different camera differences between the databases for research material and the samples in the production environment. Most high-quality retinography databases ready for research are collected from high-end fundus cameras, and there is a significant domain discrepancy between different cameras. To fully explore the domain discrepancy issue, we first collect a Fundus Camera Paired (FCP) dataset containing pair-wise fundus images captured by the high-end Topcon retinal camera and the low-end Mediwork portable fundus camera of the same patients. Then, we propose a cross-laterality feature alignment pre-training scheme and a self-attention camera adaptor module to improve the model robustness. The cross-laterality feature alignment training encourages the model to learn common knowledge from the same patient’s left and right fundus images and improve model generalization. Meanwhile, the device adaptation module learns feature transformation from the target domain to the source domain. We conduct comprehensive experiments on both the UK Biobank database and our FCP data. The experimental results show that the CVD risk regression accuracy and the result consistency over two cameras are improved with our proposed method. The code is available here: https://github.com/linzhlalala/CVD-risk-based-on-retinal-fundus-images
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16434-7_57
SharedIt: https://rdcu.be/cVRsp
Link to the code repository
https://github.com/linzhlalala/CVD-risk-based-on-retinal-fundus-images
Link to the dataset(s)
N/A
Reviews
Review #2
- Please describe the contribution of the paper
The manuscript proposes a novel method to adapt fundus images captured by two different fundus cameras to explore the domain discrepancy issue. First, the authors collect a dataset (FCP) containing pair-wise fundus images captured by two cameras with different image quality of the same patients. Second, the authors propose a cross-laterality feature alignment pre-training scheme and a self-attention camera adaptor module. I think the work is of importance for device adaption for fundus image analysis.
- 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 constuction of the dataset is important for research on the camera adaption. The proposed scheme consider the camera adapation problem from a novel view.
- 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 of selection of CVD risk estimation after the camera adaption is not clear.
- The organization and the writing needs improvement.
- 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
The description of the proposed proposed method is clear.
- 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 manuscript proposes a novel method to adapt fundus images captured by two different fundus cameras to explore the domain discrepancy issue. First, the authors collect a dataset (FCP) containing pair-wise fundus images captured by two cameras with different image quality of the same patients. Second, the authors propose a cross-laterality feature alignment pre-training scheme and a self-attention camera adaptor module. Overall, this work is of importance for device adaption for fundus image analysis.
Suggestions and questions are as follows:
- The motivation of selection of CVD risk estimation after the camera adaption is not clear.
- The manuscript doesn’t compare the proposed method to Cycle-Gan, which is famous for image translation.
- The organization and the writing needs improvement.
- 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 main idea of this work and the construction of the dataset are of great importance. The propose scheme considers the camera adaption problem in a new view.
- Number of papers in your stack
4
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
- This paper makes a fundus pair (left and right eyes) dataset with both high-precision and low-precision equipments.
- The paper proposes a cross-laterality feature alignment method for model generalization in the task of cardiovascular disease risk estimation.
- The paper design a self-attention camera adaptor module for domain adaptation, bridging the domain gap for data from different OCT cameras.
- 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 proposed cross-laterality feature alignment pre-training scheme is realized by minimizing the CVD risk difference between left and right fundus photos, hoping to extract common features in left and right fundus photos. And the self-attention camera adaptor modul is realized by minimizing the predction value between fundus photos from two equipments. This research show potential application value in disease prevention.
- 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 description of the method is not clear and accurate, for instance, “ω^r, ω^r, λ” in chapter 3.2 paragraph 1 is a spelling mistake and the meaning of the superscript of variable “y” in chapter 2.1 paragraph 2 is not mentioned. Most references of this paper are review articles, lacking of references that related to ophthalmic images or similar research. The experiment data is not sufficient and the results are not convincing.
- 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
The paper is believed to 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
- Is the assumption that “the visual clues of CVD risk have invariant representation over the two eyes” supportted by any literature? Or this can be verified by your experiment result.
- There are some loose expressions in this paper, which are easy for readers to misunderstand, for example expression “S={x_i^l, x_i^r, y_i^r, y_i^c}”, in chapter 2.1 paragraph 2, the symbol “r” has two definitions (right fundus photo and regression).
- The collapsing performance with SimSiam is confusing. Have you analyzed the reason? Your approach is not limited to a specific network architecture. Thus, more typical network architures need to be evaluated. Besides, the proposed methods do not compare with recent domain adaptation methods.
- The data provided is insufficient, such as the results of multi-task network. What’s more, the accuracy metrics of CVD risk estimation mentioned in the title is not demonstrated in the article. I think the accuracy is one of the most importanct factor to determine the application value of this work.
- 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?
In my opinion, the experiment setting in this paper is generally reasonable. However, the investigation of related resarch is not enough. The experiment results is not convincing and do not compare with state of the art methods.
- Number of papers in your stack
4
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #5
- Please describe the contribution of the paper
This paper proposed a cross-laterality feature learning training method and a camera adaptor module to improve a fundus image-base CVD risk predicting algorithm. In addition, they collected a Fundus Camera Paired (FCP) dataset containing pair-wise fundus images captured by the high-end Topcon retinal camera and the low-end Mediwork portable fundus camera of the same patients. This dataset is of great significance to the study of data in different domains for CVD risk prediction.
- 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.
In this paper, domain adaptive technology is used to increase the accuracy of CVD risk prediction from images collected by low-quality portable fundus camera, which is of great significance in clinical application. Moreover, experiments in this paper are abundant.
- 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 loss function in Fig. 2 is not described in detail.
- 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
I think this article is 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
- The quantification results are suggested in the abstract。
- Please provide the detailed description or reference of the stop-gradient operation in Section 2.1.
- In Fig.2, z^s and z^t were input into the loss function, but there is no description of this operation. Please specify the relationship between the loss function and z^s and z^t.
- 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?
I think this paper is of great significance for clinical application, and the method is feasible and the experimental results are promising.
- Number of papers in your stack
2
- 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
6
- [Post rebuttal] Please justify your decision
As I commented before, I recognize the motivation and innovation of this article. The author also answered my questions. However, perhaps due to the word limit, I have not seen the added sentences and corresponding references for question 2. But it doesn’t affect my comment of the article.
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.
TBA
- 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).
7
Author Feedback
We appreciate the suggestive comments provided by the reviewers and the area chair. Based on the comments and concerns, we provided our responses sorted by the reviewer number and the importance.
Q1: What is the motivation for the selection of CVD risk estimation? The accuracy metrics of CVD risk estimation mentioned in the title are not demonstrated. (Reviewer #2, Reviewer #3) A: The motivation lies in developing a clinical tool to predict CVD risk without conducting standardized measurements and blood tests. Thus, we explore the camera adaption for CVD risk estimation. Regarding the accuracy of the WHO-CVD score, [10] provides the CVD outcome of the WHO-CVD score. At the initial stage of this study, we analyzed the CVD outcome of different scores based on the UK Biobank data. We have added this information to our supplemental material. Q2: More method comparisons, such as Cycle-GAN and recent domain adaptation methods. (Reviewer #2, Reviewer #3） A: As our unique problem involves paired images, we considered methods utilizing the pairing information and did not adopt other methods. Recently, we have done some experiments on unsupervised domain adaptation methods such as Cycle-GAN, Domain-Adversarial Neural Networks, and Margin Disparity Discrepancy. Additional experimental results are not allowed here, but we can add them to the camera-ready version. Q3: The collapsing performance with SimSiam is confusing. (Reviewer #3) A: Our proposed method (CLFA) is inspired by the SimSiam framework, so we include it for benchmarking purposes. As mentioned in the [17], self-supervised learning can help domain generalization. We tried to combine SimSiam with supervised learning by joining the loss function. We experimented heavily but didn’t achieve the desired result. We assume the structure did learn the image contrast, which may be useless to our tasks. In our CLFA framework, we remove the MLP projector and change the feature alignment to asymmetric. Q4: Writing improvements such as “ω^r, ω^r, λ” in chapter 3.2 and S={x_i^l, x_i^r, y_i^r, y_i^c}” in chapter 2.1) (Reviewer #3) A: We have fixed the mentioned problem and will conduct more proofreading by the professional native writer before submitting the camera-ready version. Q5: Lacking references that related to ophthalmic images or similar research (Reviewer #3) A: From our knowledge, we have tried to cite the most relevant works for the task addressed in this paper, e.g. [11] and [18] are based on domain adaption for ophthalmic images, although for segmentation task and Diabetic Retinopathy task, respectively. We have additionally included Ju et al. (10.1109/TMI.2021.3056395) and Lei et al.(10.1109/JBHI.2021.3085770) for our reference. Q6: Is the assumption that “the visual clues of CVD risk have invariant representation over the two eyes” supported by any literature? (Reviewer #3) A: There has been an association between the retinal fundus and its CVD risk factors [3, 13]. The left and right eyes are in the same circulation system and have the same vessels. Therefore, they reflect the person’s health status in the same way. Q7: In Fig.2, z^s and z^t were input into the loss function, but this operation is not described. Please specify the relationship between the loss function and z^s and z^t. (Reviewer #5) A: We prepared both image features (z^s, z^t) and the prediction result (p^s, p^t) for the loss function, as shown in Fig.2. We compared three losses: MSE(z^s,z^t), MSE(p^s,p^t), and MK-MMD loss between (z^s,z^t) in the experiment (Table 2). And we found that MSE(p^s,p^t) has the best performance. Hence, in Eq.5, we only keep MSE(p^s, p^t) for our proposed method. Q8. Please provide a detailed description or reference of the stop-gradient operation in Section 2.1. (Reviewer #5) A: We have added some sentences that the motivation for using stop-gradient operation. We have also assigned [1] as its reference because [1] provides a detailed description and relative experiment.
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.
This paper proposed a cross-laterality feature learning training method and a camera adaptor module to improve a fundus image-based CVD risk predicting algorithm. In addition, they collected a Fundus Camera Paired (FCP) dataset containing pair-wise fundus images captured by the high-end Topcon retinal camera and the low-end Mediwork portable fundus camera of the same patients. This dataset is of great significance to studying data in different domains for CVD risk prediction. In this paper, domain adaptation technology is used to increase the accuracy of CVD risk prediction from images collected by a low-quality portable fundus camera, which is of great significance in clinical application. Moreover, experiments in this paper are abundant. In addition, they answered well to the reviewers’ questions about motivation and more experimental comparisons in the rebuttal. I think this article is of great significance for clinical application. Therefore, I recommend accepting this submission.
- 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).
5
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.
Prediction from low-cost cameras for possible screening for cardiovascular disease would have high clinical impact. I find the paper original and interesting for the conference, with similar approaches finding its use for other modalities. In the rebuttal, the authors have addressed the main remark of R3 on the accuracy metric being WHO-CVD, which I find appropriate.
- 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.
#R3 does not change the score. But the explanations in the rebuttal look reasonable and correct to me.
- 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).
10