Authors

Thanh Nguyen-Duc, Trung Le, Roland Bammer, He Zhao, Jianfei Cai, Dinh Phung

Abstract

Medical semi-supervised segmentation is a technique where a model is trained to segment objects of interest in medical images with limited annotated data. Existing semi-supervised segmentation methods are usually based on the smoothness assumption. This assumption implies that the model output distributions of two similar data samples are encouraged to be invariant. In other words, the smoothness assumption states that similar samples (e.g., adding small perturbations to an image) should have similar outputs. In this paper, we introduce a novel cross-adversarial local distribution (Cross-ALD) regularization to further enhance the smoothness assumption for semi-supervised medical image segmentation task. We conducted comprehensive experiments that the Cross-ALD archives state-of-the-art performance against many recent methods on the public LA and ACDC datasets.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43907-0_18

SharedIt: https://rdcu.be/dnwcg

Link to the code repository

https://github.com/PotatoThanh/Cross-adversarial-local-distribution-regularization

Link to the dataset(s)

https://www.creatis.insa-lyon.fr/Challenge/acdc/databases.html

http://atriaseg2018.cardiacatlas.org

Reviews

Review #1

Please describe the contribution of the paper

The paper introduces a novel regularization to semi-supervised medical image segmentation and achieves superior performance on the benchmark 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.
- Novel regularization (Cross-ALD).
- Extension of SVGD to produce diverse adversarial samples .
- Improved performance in semi-supervised medical image segmentation.
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 organization and the presentation of the paper is difficult to follow. Some of the claims made in the work are not backed properly either with explanations or with any experimental work.
- Objectively, the proposed method is the replacement of l_vat in SS-NET with the proposed l_crossALD. As such, the most important ablation study would be to understand how this replacement has impacted the overall performance which is missing in this work.
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

Could be improved.
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
Some questions/comments to authors:
- How ALD with Dice loss benefits medical problem “specifically”?
- The proposed framework is argued to extend and overcome the drawback of Mixup techniques. What drawback of mixup is solved by the proposed approach?
- Authors are encouraged to carefully proofread their paper. a. Consider using either all capital or small “phi” in Equation 6. b. Grammatical issues: “being be” “distribution can helps to” “intractable to find” “and etc..”
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 there are some outstanding questions about how the paper is written and organized, the problem this paper is trying to solve is quite exciting. The provided experiment demonstrates improvement over existing methods.
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 #4

Please describe the contribution of the paper

This paper introduces a novel technique called Cross-Adversarial Local Distribution (Cross-ALD) regularization for semi-supervised medical image segmentation. The technique aims to enhance the smoothness assumption, which is commonly used in existing semi-supervised segmentation methods. The smoothness assumption encourages the model to produce similar output distributions for similar data samples, which helps to improve the accuracy of the segmentation.
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 ability to achieve high performance in medical image segmentation tasks with very limited annotated data, only 3% of the total data.
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.
- While the proposed technique shows excellent performance with very few annotations, it appears to have limited performance improvements with a higher number of annotations (10%). Therefore, it may be better for the authors to submit their work to the “few-shot” segmentation track instead.
- The content of the organ to segment in the two datasets may not be diverse enough, as it appears to have a narrow focus, making evlauated segmentation task too easy to solve.
- The authors should compare their method with some recent works, including a large number of the latest CVPR-22 and MICCAI-22 semi-supervised medical segmentation methods, to validate the SOTA performance. Overall, addressing these concerns would improve the quality and impact of the paper.
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

I can not sure readers can reproduce this 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

A dataset with more diverse lesions to segment, such as BraTS2019, should be used to validate the algorithm’s performance.
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 setting of the compared mathods, as well as datasets, can not effectively validate the SOTA performance agianst a large number of methods.
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 #2

Please describe the contribution of the paper

This work aims to enhance the smoothness assumption in semi-supervised learning and. It firstly formulates an adversarial local distribution, which can cover all possible adversarial examples within a ball constraint. Moreover, the authors propose a new cross-adversarial local distribution regularization between two adversarial local distributions. The overall framework achieves superior performance on multiple semi-suupervised segmentation benchmarks, verifying the effectiveness of the proposed 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.

Main strengths:

(i) The motivation is clear and make sense: the authors aim to apply the mixup between informative samples for boosting the VAT to explore sufficient perturbations space and improve the smoothness of the neural function.

(ii) The proposed methods are technical sound and the theoretical analysis is reasonable.

(iii) The entensive experimental results over two benchmarks verify the effectiveness of the proposed framework.

(iv) The overall writing quality is good, and the paper is easy to follow.
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.

Main weaknesses:

(i) The statistical results and significantance test are beneficial to further verify the performance improvments of the proposed framework.

(ii) More latest works, for example, [1] and [2], on semi-supervised segmentation should be compared to prove the superiority of the proposed method.

[1] Yang, et al., “ST++: Make Self-training Work Better for Semi-supervised Semantic Segmentation”, in CVPR, 2022. [2] Wang, et al., “Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels”, in CVPR, 2022.
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 proposed method is clear and the implementation details are exhaustive, therefore, I believe the paper can be 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/2023/en/REVIEWER-GUIDELINES.html

As listed in the above part of 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

6
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The proposed methods are technical sound and the experiments are solid. The merits of this work outweight the drawbacks, therefore, I tend to accept this paper currently.
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

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 authors proposed a novel regularization for semi-supervised image segmentation, i.e., cross-adversarial local distribution (Cross-ALD). The technique aims to enhance the smoothness assumption used in existing semi-supervised segmentation methods. The smoothness assumption encourages the model to produce similar output distributions for similar data samples (adding small perturbations), which helps to improve the rebustness and accuracy of segmentation. In terms of evaluation, the proposed method achieved superior performance on two benchmark cardiac datasets, ACDC and LA seg dataset. The organization and presentation of the paper may be difficult to follow, and the comparison experiments need to be improved. The authors could refine their manuscript accordingly.

Author Feedback

We thank all the reviewers for their constructive comments and valuable suggestions.

Here, we proceed to respond to the reviewers’ comments as follows.

The idea behind mixup is to create new training examples by linearly interpolating between pairs of natural examples. However, not all of these natural examples are beneficial. We show that mixing between more informative samples (e.g., adversarial examples near decision boundaries) can lead to a better performance enhancement compared to mixing natural sample (see Section Ablation study).

Dice loss directly optimize the Dice coefficient which is the most commonly used segmentation evaluation metric. Therefore, we use Dice loss instead of KL from VAT.

back to top

Cross-adversarial local distribution regularization for semi-supervised medical image segmentation