List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Rui Xu, Jiaxin Zhao, Xinchen Ye, Pengcheng Wu, Zhihui Wang, Haojie Li, Yen-Wei Chen
Abstract
Retinal vessel segmentation is an essential preprocessing step for computer-aided diagnosis of ophthalmic diseases. Many efforts have been made to improve vessel segmentation by designing complex deep networks. However, due to some features related to detailed structures are not discriminative enough, it is still required to further improve the segmentation performance. Without adding complex network structures, we propose a local-region and cross-dataset contrastive learning method to enhance the feature embedding ability of a U-Net. Our method includes a local-region contrastive learning strategy and a cross-dataset contrastive learning strategy. The former aims to more effectively separate the features of pixels that are easily confused with their neighbors inside local regions. The latter utilizes a memory bank scheme that further enhances the features by fully exploiting the global contextual information of the whole dataset. We conducted extensive experiments on two public datasets (DRIVE and CHASE_DB1). The experimental results verify the effectiveness of the proposed method that has achieved the state-of-the-art performances.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16434-7_55
SharedIt: https://rdcu.be/cVRsn
Link to the code repository
N/A
Link to the dataset(s)
https://drive.grand-challenge.org/
https://blogs.kingston.ac.uk/retinal/chasedb1/
Reviews
Review #1
- Please describe the contribution of the paper
The authors present a novel segmentation network for the retina vessel segmentation task. The main idea of the paper is to apply contrastive learning into the vessel segmentation network. The contrastive learning approaches are studied in the paper. One is the local region contrastive learning by selecting hard samples in a local region manner. The second one is the generalization of local regions to batches inside the whole dataset. Experiments show that incorporating contrastive learning to standard segmentation networks boosts the segmentation results.
- 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 strengths of the paper:
-
The overall paper is well organized and the writing of the paper is great.
-
Although deep learning segmentation network methods are extensively studied for vessel segmentation, the authors present a novel way to further improve the segmentation results by contrastive learning approaches. The presented method consists of local and global contrastive learning means to select hard samples and provide additional supervision into the segmentation network.
-
The authors provide experiments to show the effectiveness of the proposed contrastive learning methods. Results compared to previously reported methods are also analyzed in the paper.
-
- 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 have one question about the proposed method.
The global contrastive learning can be treated as a mini-batch with a large batch size. Therefore the claimed global contrastive learning is also in local contrastive learning manner.
- 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 details of the method are introduced in the paper. It should be easy to reproduce the 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/2022/en/REVIEWER-GUIDELINES.html
The overall paper is well organized and the writing of the paper is great. I have no further comments regarding the presentation of the paper.
- 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 deep learning segmentation network methods are extensively studied for vessel segmentation, the authors present a novel way to further improve the segmentation results by contrastive learning approaches.
- 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 #2
- Please describe the contribution of the paper
The paper introduces a contrastive learning based algorithm for retinal vessel segmentation. The algorithm consists of a local intra-region contrastive learning strategy and a global cross-dataset contrastive learning strategy.
- 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 well written, and easy to read.
The algorithm is well presented.
- 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 novelty of the paper is limited. A very similar method has been introduced in [1], which, however, is not included in the reference list. It seems to me that the paper just adapt the algorithm from natural image segmentation to a medical image segmentation task. The quality-aware anchor sampler is very similar with the segmentation-aware anchor sampling in [1].
[1] Exploring Cross-Image Pixel Contrast for Semantic Segmentation. ICCV 21.
Another major issue is that the performance improvements of the paper are minor on the two datasets.
- 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
It seems to me that the paper can be easily reproduced.
- 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 is essential to provide a detailed comparison with [1] and clarify the contributions of the paper.
[1] Exploring Cross-Image Pixel Contrast for Semantic Segmentation. ICCV 21.
The term “cross-dataset” confuses me a lot. Would it be better to use “cross-image” like [1] because the contrast is conducted over different images rather than different datasets?
I wonder whether it is possible to formulate the two strategies into only one formulation like [1]? Are there any advantages to compute them separately?
- 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?
The algorithm proposed in the paper is mostly adapted from [1], which however is not cited in the paper. It is necessary to clarify the technical contributions of the paper against [1], before it can be accepted.
[1] Exploring Cross-Image Pixel Contrast for Semantic Segmentation. ICCV 21.
- 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
5
- [Post rebuttal] Please justify your decision
The rebuttal provides some clarifications on the difference between the proposed method and the ICCV21 paper. The proposed method indeed does some technical modifications to fit the task, however, the improvements are not significant. My rating is still “borderline”, but I change the score to “weak accept” for easier decision. The technical designs of contrastive loss in this particular tasks, if precisely presented in the article, probably benefit future studies.
Review #3
- Please describe the contribution of the paper
This paper proposes a framework to improve retinal vessel segmentation performance on some challenging pixels via the local-region and cross-dataset contrastive learning. In specific, the authors use a quality-aware anchor sampler to select the challenging pixels of false predictions and then construct contrastive loss in both the local area (with pixels from the same cropped patch) and global region (the other patches stored in the memory bank). The authors have verified the proposed method on DRIVE and CHASEDB1 datasets.
- 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.
-
Authors motivate this work concisely - to extract discriminative features from those challenging vessel pixels by using context in local and global areas. The paper structure is clear and all sections echo the motivation well.
-
The contrastive strategy fits well with this retinal vessel segmentation task. Retinal vasculature is long and tortuous, some vessel pixels in complex vascular structures and low contrast background are elusive, which makes those vessel features are indiscriminative from the background. The idea of the contrastive strategy can enhance the model’s capability of extracting discriminative features.
-
A variety of results are shown, including segmentation visualisation, quantitative segmentation metrics (two connectivity scores), and t-SNE maps.
-
- 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 technical novelty of this paper is moderate, considering these two contrastive strategies have been proposed and exploited early in [4]. The main methodological difference between this paper and [4] is that this paper uses a quality-aware anchor sampler to select a collection of challenging pixels for contrastive learning. However, this selection has also been introduced and verified in the paper [1], which is missed in reference.
[1] Wang, Wenguan, et al. “Exploring cross-image pixel contrast for semantic segmentation.” Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021. [4] Hu, Hanzhe, Jinshi Cui, and Liwei Wang. “Region-Aware Contrastive Learning for Semantic Segmentation.” Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021. (cited by the paper)
-
Correlating to the point above, some descriptions can be more objective. Instead of claiming “we have proposed a local intra-region contrastive learning strategy and a global cross-dataset contrastive learning strategy…”, I would suggest changing it to “we have applied … into retinal vessel segmentation”. I believe a paper with a novel application can also show lots of strengths.
-
The image sizes of DRIVE and CHASEDB1 are smaller than (1000, 1000). Some publicly available datasets with large image sizes, like HRF, can better evaluate the capability of segmenting thin and elusive vessels. Also, the performance is comparable to some miccai works last year [2, 3].
[2] Zhou, Yuqian, Hanchao Yu, and Humphrey Shi. “Study group learning: Improving retinal vessel segmentation trained with noisy labels.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2021. [3] Kamran, Sharif Amit, et al. “RV-GAN: segmenting retinal vascular structure in fundus photographs using a novel multi-scale generative adversarial network.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2021.
- 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 is moderate as parts of training hyperparameters are introduced and network structure is shown in supplementary material.
- 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 suggest revising the paper writing to be more objective and application-motivated. A reasonable and novel application is a kind of novelty.
-
It is recognise that the weights for contrastive loss are small (i.e., 0.001). Have the magnitude of L_ce, L_lc, and L_gc been studied? How do these tiny contrastive losses contribute to the model training?
-
Please keep the naming consistency, like “local-region” and “local intra-region”. Additionally, “easy vessel samples” and “easy region-level vessel samples” only appear once in Figure 1, they are called “high-quality samples” in the main text.
-
Language needs polish - some sentences and words can be more precise. For example, the first sentence of paragraph three in the Introduction says “For addressing the feature discrimination…”. It can be revised as “ To extract discriminative features…” or “To enhance the features’ discriminability…”.
-
- 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 methodology motivates well, but the paper presentation needs to be objective and application-motivated, which requires a certain revision. Some other points also need clarification, summarised in the weakness and comments.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
2
- 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
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.
This paper proposed a contrastive-learning-based method to address the retinal vessel segmentation. The paper was reviewed by three experts who have pointed out several major issues. All the three reviewers given the borderline ratings. However, the major issues arised by the reviewers are consistent. First the technical novelty is limited and very similar exisiting literature ICCV 2021, but it is not included in the reference list. In addition, the experimental results on two datasets did not show its superior on exisiting work such as miccai 2021. Therefore, the authors should address the concerns in rebuttal.
- 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
To All Reviewers:
- Relation of this Paper to [1] The paper [1] proposes a supervised contrastive learning (SCL) based method for semantic segmentation of natural images. Inspired from [1], we present a method to apply SCL for retinal vessel segmentation. We find that the direct application of [1] for vessel segmentation can not earn performance gains. Thus, we modify several techniques to adapt SCL for our task. Besides, we are sorry to miss the citation of [1] unintentionally. We summarize the differences and the results of [1] as follows.
1) Positions where to exploit SCL We exploit SCL at the last layer of decoder(highest spatial resolution), rather than at the deepest layer of encoder(lowest spatial resolution) in [1]. This is important for vessel segmentation. Vessels have a long and thin shape, and its segmentation requires more accurate spatial information, which is contained at the last layer of decoder. Besides, a key point of successfully applying SCL is to accurately detect the pixels that can be selected as anchors and positive/negative(P/N) samples. This detection is more accurate when it is performed at the last layer of decoder.
2) Selection of Anchors and P/N Samples Another difference between our method and [1] is the selection of anchors and P/N samples. We only select pixels that are hard to be segmented as anchors, while [1] selects both of hard pixels and randomly chosen pixels as anchors. Since U-Net performs relatively well on vessel segmentation, it is more possible that randomly chosen pixels have already been segmented correctly. It is not necessary to fine-turn features of these pixels by using SCL. Besides, the fine-turning can even cause worse performance. Thus, we only select hard pixels as anchors. Besides, P/N samples for each anchor are selected inside a SxS local region whose center is the corresponding anchor, while P/N samples are selected from the whole images in [1]. In our task, anchors are hard pixels that are usually capillaries or vessel boundaries. Due to that these anchors are easily confused with their surrounding pixels, the selection of P/N samples directly from their surrounding regions is more straight forward and more beneficial for fine-turning features of anchors in SCL. This selection also causes that P/N samples stored in memory bank are different from [1].
3) Results of direct application of [1] Without the above-mentioned adaptation, SCL can not improve the retinal vessel segmentation, and even the performance is lower than the U-Net. We list the results of direct application of [1] in follows. DRIVE:0.9678(ACC),0.8086(SE),0.9833(SP),0.9820(AUC),0.8140(DICE),0.4178(COR),0.4934(INF); CHASE_DB1:0.9751(ACC),0.8162(SE),0.9859(SP),0.9879(AUC),0.8046(DICE),0.3248(COR),0.6196(INF).
Performance of Our Method and Comparison to MICCAI21 Papers Since mis-segmented pixels by the U-Net are not too many, improving segmentation of these pixels can not distinctly increase traditional segmentation metrics. However, the mis-segmented pixels can cause many breakpoints on segmented vessel-trees, which does harm for further analysis of ophthalmic diseases. Our method can decrease these breakpoints and largely improve the connectivity of vessel-trees, which can be distinctly indicated by the metrics of COR and INF in Tab1. In addition, we find that our method performs at least equally as the MICCAI21 papers mentioned by the reviewers. Even, several segmentation metrics are better. Note that their training strategies and network structures are more complex than ours. Besides, our method can be easily extended to other network and medical segmentation tasks. We will cite these MICCAI21 papers in the revision.
We will cite the paper [1] and indicate the differences between our method and [1] in the revision. Besides, we will accept the comments from the reviewers and describe our method more objectively.
[1] Exploring Cross-Image Pixel Contrast for Semantic Segmentation.ICCV21
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.
In rebuttal, the authors clarify the difference between the paper and ICCV paper. And the reviewer has changed the rating from negative to positive.
- 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 #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.
The authors have clearly clarified the issues raised by the reviewers. 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).
5
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 proposes a retinal vascular segmentation network based on contrast learning. According to the reviewers, the paper has two main problems： 1) the paper is similar to the work of ICCV2021 which needs to be distinguished and cited. 2) The experimental results do not show the advantages of the method. After rebuttal, all reviewers consistently recommended acceptance of the paper and stated that the feedback from the authors clarified the issues raised to some extent.
- 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).
2