List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yanting Xie, Hongen Liao, Daoqiang Zhang, Fang Chen
Abstract
Ultrasound image segmentation plays an essential role in automatic disease diagnosis. However, to achieve precise ultrasound segmentation is still a challenge caused by the ambiguous lesion boundary and imaging artifacts such as speckles and shadowing noise. Considering that the pixels with high uncertainty generally distributing in the boundary regions of prediction maps, are likely to overlap with the confused regions of ultrasound, we proposed an uncertainty-aware cascade network. Our network uses the confidence map to evaluate the uncertainty of each pixel to enhance the segmentation of ambiguous boundary. On the one hand, the confidence map fuses with the ultrasound features and predicted mask using the adaptive fusion module (AFM) which enriches the context features from different modalities. In addition, the uncertainty attention module (UAM) is proposed based on the confidence map. This module focuses on the influential features with cross attention constrained by the uncertainty of pixels which can extract the localized features of confused ultrasound regions. On the other hand, the recurrent edge correction module (RECM) further improves the segmentation of ambiguous boundary. This module increases the weights of confident features neighboring the uncertainty boundaries in order to refine the predictions of edge pixels with low confidence. We evaluated the proposed method on three public ultrasound datasets and the segmentation results show that our method achieved higher Dice scores and lower Hausdorff distance with more precise boundary details compared with state-of-the-art methods.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16440-8_26
SharedIt: https://rdcu.be/cVRvR
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
- The uncertainty of each pixel from ultrasound images is leveraged to improve segmentation performance.
- A new uncertainty-aware network with AFM, UAM, and RECM is proposed.
- The experiment results are shown to be competitive compared to SOTA on three public US 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.
- The motivation and innovation of this work are good.
- A novel recurrent edge correction module (RECM) adjusts the low-confidence pixels based on the neighboring high-confidence pixels to decrease the weights of the indistinguishable features.
- 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 UAM idea is confusing (see detailed comments).
- Some descriptions, figures, and equations need clearing up (see detailed comments).
- The authors lack analysis of design ideas or experimental results.
- 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
since some necessary details are missing, the Reviewer thinks the reproducibility is modest.
- 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 authors apply the UAM structure before each encoder layer of the second U-Net. According to Eq. 2, the UAM requires uncertainty confidence maps, decoder features, and high-scale features. It is puzzling that the size of the uncertain confidence map is 1wh for each layer, which seems to be incorrect. How do you obtain the uncertain confidence map for non-top layers? Also, for the top layer, what are the high-scale features?
-
In Fig. 2, it is best to label the UAM in the U-Net and show more details of the RECM.
-
The operator symbols are difficult to understand. Does the “x” symbol in Eq. 2 reflect element-wise multiplication? Is the “/” symbol in Eq. 3 representing element-wise division? Does element-wise multiply M_conf and X in Eq. 3? And M_ca should be M_conf in Eq. 2.
-
The authors divide the network training into two phases. Can the whole network be trained jointly to achieve better performance?
-
The RECM and boundary loss actually do not contribute much to the improvement.
-
The first sample in Figure 3 on TN-SCUI shows that the segmentation result of the network differs significantly from GT, and even the borders appear curled.
-
For the results in Table 1&2&3, it is better to add the standard deviation.
-
The authors lack analysis of design ideas or experimental results.
-
- 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?
- There are still several errors in the proposed network, and some experimental results cast doubt on the network’s applicability.
- Analysis of design ideas or experimental results is lacking.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
3
- 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
For the task of ultrasound image segmentation, this paper proposed an uncertainty-aware cascade framework. The confidence map-guided uncertainty map is adopt for feature fusion, feature attention and edge correction.
- 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 proposed components seem reasonable and relevant to clinical practice. 2) Extensive experiments were conducted to show the 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.
1) Some technical details are not clear. e.g., what is ‘ultrasound feature’ in Fig.2 and Sec.2.1? 2) There are no qualitative results to show the effectiveness of proposed method in boundary issues, especially in recurrent edge correction.
- 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 reproducibility is acceptable.
- 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
1) In Fig.1, how the uncertainty map obtained? Considering the proposed method rely highly on confidence map, it is better to list confidence map for better motivation introduction. 2) What is ‘ultrasound feature’ in Fig.2 and Sec.2.1? 3) Authors mentioned that the method improves segmentation by “decreasing the weights of features causing the uncertain predictions”. Why “decreasing”? Will “increasing” work? Considering “increasing” will make the model pay more attention to the hard regions. I’d like to hear about the discussion. 4) There are no qualitative results to show the effectiveness of proposed method in boundary issues, especially in recurrent edge correction. (Fig.3 failed in this issue.) 5) Considering AFM and UAM are time consuming components, the time efficiency analysis should be added. 6) Typos. e.g., “duo” in abstract.
- 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 components and corresponding experiments seem to be reasonable. Some minor issues should be revised or discussed as mentioned above.
- Number of papers in your stack
6
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Very confident
- [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
The authors give reasonable responses to reviewers, especially R1. To me, this manuscript is acceptable.
Review #3
- Please describe the contribution of the paper
This paper proposes an uncertainty-aware framework with multiple uncertainty-aware modules to improve the segmentation accuracy on ultrasound images.
- 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 overall structure is clear and informative.
- The experiment is comprehensive and the compared methods are new and related.
- The performance is good with in-depth discussions.
- The method is novel which utilizes multiple forms of uncertainty to enhance feature sharing and embedded in a cascade network.
- 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.
- Some training details are missing, e.g., the number of layers and channels in the network.
- 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
Easy to reproduce.
- 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 writing is clear and well organized with clear motivation
- The experiment is relatively comprehensive (3 datasets and compared with 3 relevant important MICCAI’21 works), and does a good job on ablation study
- The results show superior performance to well support the claims
- The investigation of related work is sufficient
- The code will be published also
- There is no discussion of the failure cases, which may due to the length limit
- There is no standard deviations for the results
- The figure quality is not very good. Suggest to improve the image resolution.
- The novelty of each module may be limited or incremental. However, I think it’s ok to propose a framework with a good 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Good and clear introduction, and comprehensive experiment with good performance.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
7
- [Post rebuttal] Please justify your decision
I am aware of the concerns of other reviewers, and I think the author has well addressed the concerns.
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 an uncertainty-aware cascade framework for ultrasound image segmenation. This paper received two accept and one reject decision. In the rebuttle, the authors should clarify the concerns raised by the first reviewer, including clarify some descriptions, figures, and equations and improve the organization of experimental results.
- 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).
4
Author Feedback
We gratefully thank the reviewers for their remarks and suggestions. We appreciate the encouraging comments like “good motivation, innovation and novel method” of R1, “extensive experiments were conducted to show the effectiveness” of R2, and “the overall structure is clear and informative; the experiment is comprehensive and compared methods are related; the performance is good with in-depth discussions” of R3. For the questions and concerns, our responses are given as below (1)Q:“Can the whole network be trained jointly to achieve better performance” for R1 A:Actually, we compared the results using these two training schemes. For TNSCUI, when we trained the whole network jointly, the network presented a worse performance, with a decrease of 3.06% for DSC↑and an increase of 2.44mm for HD↓ (2)Q:”RECM and boundary loss donot contribute much to the improvement” for R1 A:The improvements of RECM and boundary loss were statistically significant. With the results of the t-test for HD, the values of p were 2.3e-7 and 3.6e-7 for RECM and boundary loss. Additionally, based on the complete ablation results in appendixC, when we added RECM and boundary loss to the baseline directly, HD can be decreased about 5.5mm. When adding RECM, DSC increased from 85.87% to 87.62% and HD decreased from 20.03mm to 16.76mm. Meanwhile, when we added boundary loss, DSC increased to 88.34% and HD decreased to 14.66mm. Thus, with these modules, HD can be decreased significantly (3)Q:“operator symbols distinction” for R1 “more details for each module” for R1,2,3 “standard deviations for the results” for R1,3 A:In Eq2,3, the “x” and “/” represent element-wise operation. The “Mconf” and “X” in Eq3 will do element-wise multiplication. We have cleared up the matric and element-wise operation under each equal For Fig2, we have supplemented descriptions about module details like the number of layers and channels. The standard deviation of result has been added to the Tab1&2. These supplements will not increase the length of the paper (4)Q:“The first sample in Fig3 shows that boundary differs from GT, even the borders appear curled” for R1 “qualitative results to show the effectiveness in boundary issues in Fig3” for R2 A:Borders appear curled: The results visualized didnot have any post-process like filling the holes or getting largest connected area which will improve the results. The boundary of the hole will be drawn which makes the boundary of mask seems curled Qualitative analyses of the first sample: The first sample in Fig3 shows that our method performed well for confused cases in shadow artifacts and boundary-missing situation. Although our result differs from GT in some regions, it is closer to the GT in missing boundary regions compared with all other methods. For the quantity result of this case, DSC and HD of ours are 88% and 22mm, which outperforms UNet++ with DSC of 69% and HD of 75mm (the best results of all other methods). Ours improved an increase of 19% for DSC and a decrease about 53mm for HD In addition, we analyze the reason of improvements. Shadow artifacts lead to the missing boundary in local regions which complicates the segmentation. The information inside the shadow region is lost and less reliable than that in shadow-free regions. However, most existing methods tend to treat all pixels equally no matter where they locate, which fails to deal with the situation. In contrast, ours treat shadow pixels with different attention using UAM and approach the missing boundary with RECM. Thus, ours can preserve the shadow region better than all other methods (5)Q:” Explaining about the confusion of UAM” for R1 A:The key operation of UAM is cross-attention with confidence weight which calculated based on the UNet prediction. The confidence map is transmitted to each layer with downsample operation for size consistency. For the top layer, we used the fused features from AFM as the high-scale features. We have cleared up the details of UAM in Fig2
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 the rebuttal, the authors addressed well for the reviewers’ comments, including “joint training” “the improvement of RECM”. I would suggest to accept this paper.
- 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
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 critics of the reviewers were well addressed in the rebuttal, clearly showing the amount of work the authors put into this manuscript.
- 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
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.
Overall, the reviewers are supportive to this paper.
- 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).
1