Authors

Shishuai Hu, Zehui Liao, Yong Xia

Abstract

Deep learning-based medical image segmentation models suffer from performance degradation when deployed to a new healthcare center. To address this issue, unsupervised domain adaptation and multi-source domain generalization methods have been proposed, which, however, are less favorable for clinical practice due to the cost of acquiring target-domain data and the privacy concerns associated with redistributing the data from multiple source domains. In this paper, we propose a \textbf{C}hannel-level \textbf{C}ontrastive \textbf{S}ingle \textbf{D}omain \textbf{G}eneralization (\textbf{C$^2$SDG}) model for medical image segmentation. In C$^2$SDG, the shallower features of each image and its style-augmented counterpart are extracted and used for contrastive training, resulting in the disentangled style representations and structure representations. The segmentation is performed based solely on the structure representations. Our method is novel in the contrastive perspective that enables channel-wise feature disentanglement using a single source domain. We evaluated C$^2$SDG against six SDG methods on a multi-domain joint optic cup and optic disc segmentation benchmark. Our results suggest the effectiveness of each module in C$^2$SDG and also indicate that C$^2$SDG outperforms the baseline and all competing methods with a large margin. The code is available at \url{https://github.com/ShishuaiHu/CCSDG}.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43901-8_2

SharedIt: https://rdcu.be/dnwCC

Link to the code repository

https://github.com/ShishuaiHu/CCSDG

Link to the dataset(s)

https://zenodo.org/record/6325549

Reviews

Review #3

Please describe the contribution of the paper

This paper proposed a Channel-level Contrastive Single Domain Generalization (C2SDG) method to alleviate the problem of inconsistent distributions due to various medical equipment and protocols. To be specific, a style augmentation module is used to produce another style image, and a feature disentanglement module is used to extract style representations and structure representations, finally, structure representations are used to segment and performance improvement is achieved. Experimental results on RIGA+ dataset is shown.
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 organized and written, and the logic is clear. The proposed method is relatively useful and practical. Experimental results show its effectiveness.
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 effectiveness might need to be further verified instead of testing on only RIGA+ dataset. More complex dataset should be considered. Besides, one source domain generalization might not be some practical in real-world applications. Multi-source domain generalization is possible. The method is not very novel.
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

This work is relatively more reproducible because the code and the model will be available.
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

Please check the weakness. More experiments should be helpful. Besides, the figures, especially figure 2 may be simplified and reorganized.
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?

Proposed method is relatively practical and the experimental results prove the statements. The paper is well organized.
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 #1

Please describe the contribution of the paper

This paper proposes a new method, called C2SDG, to address the issue of performance degradation of deep learning-based medical image segmentation models when deployed in new healthcare centers. Existing methods for addressing this issue are not practical for clinical practice due to high costs and privacy concerns. C2SDG disentangles style and structure representations using contrastive training, allowing feature disentanglement with a single source domain.
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.
1. The paper addresses a very important and clinically relevant problem
2. The paper is well-written and well-organized
3. The results seem to be good
4. The authors performed necessary ablation study
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.
1. The authors only did one experiment with a single dataset.
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 authors intend to release the code, otherwise the details of training is written in 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/2023/en/REVIEWER-GUIDELINES.html

Please refer to the strength and weakness 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

7
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The proposed method is sound and the results show a definite win of the method over other methods. The paper is also well-organized and clearly written hence this a clear accept for me.
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

This work proposes a method for contrastive optimization based on the feature channels. The proposed method outperforms the previous SOTA methods.
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.
1. The experimental results are good on some datasets.
2. The feature distanglement method sounds somewhat novel.
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.
1. How to ensure the quality of feature distanglement as the channel mask prompt is randomly initialized?
2. The length of mask prompt is not studied.
3. There is no attention visualization of style and structure features.
4. The proposed method is not compared with the following: Liu, Quande, Cheng Chen, Qi Dou, and Pheng-Ann Heng. “Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary.” In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 2, pp. 1756-1764. 2022.
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 major experimental details are given.
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

Please refer to the strength and weakness.
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 experimental analysis is not sufficient.
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.

The paper proposes a Channel-level Contrastive Single Domain Generalization (C2SDG) method to mitigate issues related to inconsistent distributions from different medical equipment and protocols. The approach uses a style augmentation module and a feature disentanglement module, leading to improvements in segmentation performance. The work is praised for its clear logic and organization, as well as its practical relevance, with experimental results on the RIGA+ dataset demonstrating its effectiveness. The paper is critiqued for only testing on one dataset, a lack of comparative analysis with similar research such as the AAAI 2022 paper, and not visualizing style and structure features. Concerns are also raised regarding the method’s novelty and its practicality. Overall, the strengths of the paper significantly overweight the relatively minor weaknesses. Thus the paper is recommended for provisional accept.

Author Feedback

We sincerely thank all reviewers and ACs for their invaluable comments. The code of this work is in supplementary files and will be available on GitHub.

#R1Q1&R3Q1: Only testing on one dataset Due to the page limitation, we have to make a compromise between sufficient analysis of each module in the proposed CCSDG and testing on more datasets. But we are willing to test the proposed CCSDG on more datasets in our future work.

#R2Q1: Quality of feature disentanglement Although the channel mask prompt is randomly initialized, it is updated to minimize L_{str}+L_{sty} during training. Once L_{str}+L_{sty} is minimized, the channel mask prompt can well disentangle f into f_{str} and f_{sty}, as shown in Fig. 2 in the supplementary. Also, the experimental results in Table 2 in the supplementary demonstrate either 1-0 initialization or random initialization can improve the performance, and random initialization is better than 1-0 initialization.

#R2Q2: Length of mask prompt The length of channel mask prompt can be adjusted according to the specific network design but is always equal to the channel number of f, i.e., 64 in our experiments.

#R2Q3: Visualization of style and structure features The visualization of style and structure features was provided in in Fig. 2 (c) and (d) in the supplementary. It reveals that the style representations (c) can be largely affected by the style discrepancy, whereas the structure representations (d) are relatively robust to the style variation.

#R2Q4: Comparison with AAAI 2022 method TASD As described in TASD, this method adopts test-time adaptation strategy with dual-consistency regularization to adapt the model to target domain data. Therefore, this method is more like a test-time adaptation method but not a method focusing on single domain generalization, which is beyond the scope of this paper. Thus, we compared more recently proposed single domain generalization methods, such as SLAug (AAAI 2023) and Dual-Norm (CVPR 2022), instead of TASD.

#R3Q2: Single domain generalization might not be some practical We acknowledge that multi-source domain generalization is a promising research direction, but we disagree that single domain generalization is less practical. Privacy concerns can be raised during multi-source domain data collection for training multi-source domain generalization models, but single domain generalization models only require data from a single domain, and thereby avoid data redistributing, which is more friendly to privacy protection. Therefore, we proposed CCSDG for single domain generalization in this paper. Its novelty mainly lies in the channel-level contrastive design and its superiority has been demonstrated by ablation studies and comparison experiments.

back to top

Devil is in Channels: Contrastive Single Domain Generalization for Medical Image Segmentation