List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Tao Wang, Hui Yu, Yan Liu, Huaiqiang Sun, Yi Zhang
Abstract
Metal artifacts in computed tomography (CT) degrade the imaging quality, leading to a negative impact on the clinical diagnosis. Empowered by medical big data, many DL-based approaches have been proposed for metal artifact reduction (MAR). In supervised MAR methods, models are usually trained on simulated data and applied to the clinical data. However, inferior MAR performance on clinical data is usually observed due to the domain gap existing between simulated and clinical data. Existing unsupervised MAR methods usually use clinical unpaired data for training and validation, which often distort the anatomical structure due to the absence of supervising information. To address these problems, we propose a novel semi-supervised MAR frame work. We use the clean image as the bridge between the synthetic and clinical metal-affected image domains to close the domain gap. We also break the cycle-consistency loss, which is often utilized for domain adaptation, since the bijective assumption is too harsh and does not accurately respond to facts of real situations. To improve the MAR performance, we proposed Artifact Filtering Module (AFM) to eliminate features helpless in recovering clean images. Experiments demonstrate the performance of the proposed method is competitive with several state-of-the-art unsupervised and semi-supervised MAR methods in both qualitative and quantitative aspects.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43999-5_20
SharedIt: https://rdcu.be/dnwwy
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 author proposed a novel semi-supervised MAR framework to bridge the domain gap between synthetic and clinical data. At the encoding step, feature maps with limited influence will be eliminated, where the encoder functions as a bottleneck only allowing useful information to pass through. To push the generated image close to the clean image domain, they used conditional normalization layers and propose a metal-free spatially aware module (MFSAM). The network was compared to SOTA MAR algorithms, both qualitatively and quantitatively on one simulated dataset as well as two clinical 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.
- In some domain adaptation-based networks, they usually assume that there is a bijection between two domains, which is implemented by cycle consistency loss. This bijection too harsh and does not accurately respond to facts of real situations. In this work, the author employed the clean image domain as the bridge, replacing the bijection with two straightforward mappings and loosening the strict assumption imposed by the cycle-consistency loss.
- To enhance the conversion of two images that have been corrupted by metal into ones that are free of metal, the author utilized a feature selection mechanism and introduced an Artifact Filtering Module. This Module acted as a filter, effectively eliminating any features that are not useful in restoring the images to their clean state.
- The authors conducted extensive experiments to evaluate the effectiveness of their proposed method. They tested the method on a synthesized dataset as well as two clinical datasets, and compared it with existing state-of-the-art metal artifact reduction methods. The results demonstrate that the proposed method outperforms other methods in terms of both quantitative metrics and qualitative visual assessment.
Overall, in this work, the author explicitly reduced the domain gap between synthetic and clinical metal-corrupted CT images for improved clinical MAR performance.
- 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 paper proposes a potentially effective approach for reducing metal artifacts in CT imaging, I still have some concerns:
- What are the C, K, S, and P in the Fig. 1. (b) and Fig. 1. (d)? What does the+- symbol in the Fig. 1. (a) mean?
- The authors utilized conditional normalization layers and introduced a metal-free spatially aware module (MFSAM) to minimize the distance between the generated image and the clean image domain. The MFSAM adjusts the mean and variance of the features to align with the style of the metal-free image, enabling the generated image to appear more natural and metal-free. However, I did not see the mean and variance in Fig. 1. (c). In addition, what is beta and gamma in Fig. 1. (c)? Are they the mean and variance?
- At the encoding step, the author proposed an artifact filtering module (AFM) for feature selection, where feature maps that contribute little to the reconstruction of clean images were considered to be related to noise. Therefore, the encoder functioned as a filter and only allowed useful information to pass through. This selection was based on two criteria. However, the author did not introduce how the two criteria were calculated. Generally, the weaknesses identified in the paper do not significantly diminish the main contributions of the proposed method. However, addressing these limitations can help enhance the credibility and influence of the approach and its results.
- 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
The complete network structure is presented. The data used in this article comes from the public dataset, and the training details and parameter settings are explained in the article. This article is with good reproducibility.
- 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
- In the Fig.1, the author should explain the meaning of C, K, S, P and the+- symbols.
- The author should briefly introduce AFM and MFSAM in the caption of Fig. 1, or indicate where detailed explanations can be seen in the manuscript.
- The meaning of beta and gamma in Fig. 1. (c) is missing. Please add them. And it also needs to be explained in the manuscript to facilitate readability.
- In artifact filtering module, the criteria for feature map selection are important. The author needs to introduce how the two criteria were calculated in detail.
- 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?
This paper explicitly bridged the domain gap between synthetic and clinical metal-corrupted CT images. At the encoding step, the encoder functions as a bottleneck only allowing useful information to pass through. Experiments demonstrate the performance of the proposed method is competitive with several SOTA MAR methods in both qualitative and quantitative aspects. In particular, the method exhibits good generalization ability on clinical data.
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
8
- [Post rebuttal] Please justify your decision
This is an interesting work, addressing an important problem of domain gap in the MAR. The authors responded adequately to the reviewers’ comments and have revise the paper accordingly. Overall, it is a solid piece of work. I recommend acceptance of the paper as it makes a contribution to the CT imaging field.
Review #2
- Please describe the contribution of the paper
This work proposed a semi-supervised MAR image translation framework using clean image to reduce the domain gap between the synthetic and clinical metal-affected 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.
During the image translation, mixing the synthetic artifacts and clinical artifacts with predicted clean image for domain gap reduce, which seems to be 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.
- What’s the remaining problems existing semi-supervised MAR methods cannot solve, what’s the key contributions in this work? Please give more detailed analysis in the introduction.
- Not all of the domain adaptation based networks are based on one-to-one mapping, what’s the novelty of this work over current domain adaptation based networks?
- The “overview” description and the corresponding figure (Fig. 1(a)) cannot well match. The overview is not clear to show the translations I_free to I_syn and I_free to I_clinical. For I_free to I_syn, I_c2f is used to represent I_free in Eq. 7, which is not accurate.
- For Fig. 2, the caption and the symbols in the Figure are not correct.
- For Sec. 3.2, statistical analysis results on clinical data (CL1 and CL2) should be given to show the effectiveness, besides the simulated data, like how many images out of the total dataset that the proposed method outperforms the second best result.
- 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
can be reproduced according to 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
This paper needs to be reorganized to better clarify the motivation.
- 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?
Please see the weakness.
- 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
4
- [Post rebuttal] Please justify your decision
After reading the author’s feedback, I still think that the motivation of using clean image as the bridge is not highlighted to me. For example, how to model the clean image as the bridge, what does that kind of bridge can provide for the domain adaptation process? The introduction part needs to be reorganized for better understanding.
Review #3
- Please describe the contribution of the paper
This paper proposes a novel semi-supervised MAR frame work for the metal artifact reduction (MAR) problem. And an Artifact Filtering Module (AFM) is proposed to remove features that do not help to restore a clean image. The method outperforms the state-of-the-art semi-supervised and unsupervised MAR 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.
- The method proposed in this paper is easy to understand and has a clear structure
- This paper effectively combines semi-supervised learning and GAN
- Experimental settings can reflect the generalization of the model on different domains
- 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.
-
Domain adaptation is a method that specifically deals with the problem of domain differences. From the perspective of the paper, it is aimed at the processing of artifact objects in medical images, and the comparison method also uses the Generative Adversarial Network. Therefore, is it reasonable to define the problem as domain adaptation in the introduction? It can be understood as artifact removal and denoising rather than domain gap
-
Figure 1 makes me confused due to some typos, i.e., I_syn and I_sny
- The layout of Figure 1 is not intuitive. For example, the input location is scattered and difficult to find. And what does C128K3S1P1 mean? I gauss it represents channel number=48, kernel size = 3, etc. The authors need to improve the presentation in this paper.
-
The experiment is not convincing. ADN achieved 33.6dB PSNR on SYN dataset in the unsupervised manner in their publication. However, this paper achieve much lower results even with supervison.
- Insufficient innovation in methods. There are also many methods for multiple generators. in ADN, networks are used to learn the relationship between noisy images and the original images. And this article uses addition and subtraction operations. Is metal artifact noise an additive noise?
-
- 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
This paper analyzes the shortcomings of unsupervised and fully supervised methods on the metal artifact reduction (MAR) problem raised. The structure of the article is reasonable, but there is a mismatch between the figure and the text.
- 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
We suggest adjusting the layout of the framework figure and explaining the differences between the comparative experiment and the original 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
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The model structure diagram is not intuitive and the layout is messy; there are some typos.
In the comparison experiment, some methods are relatively early methods, and there is no comparison of the domain adaptation method mentioned in the introduction. Is this method better than the current sota method?
- 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.
In the paper the authors present a metal artifact reduction method for CT imaging based on a semi-supervised learning strategy. The paper got diverse ratings in the reviews. R1 strongly recommend the paper for accept while R2 and R3 pointed out several issues of the paper and recommend for reject. Those issues agreed by R2 and R3 mainly focus on the novelty and contributions, motivations, experimental design, assessment method, results, and written issues. Although I lean to agree with R2 and R3, I would like to invite the authors to address these concerns from all reviewers in the rebuttal phase.
Author Feedback
We sincerely thank reviewers for their valuable comments. We have carefully checked the figures and the captions and revised the mistakes. We mainly address major concerns below.
#R2 1.What’s the remaining problems existing semi-supervised MAR methods and key contributions in this work?
Existing semi-supervised MAR methods usually use the same network to reduce artifacts on both simulated and clinical data and attempt to find a balance between the two datasets. However, the domain gap still exists and has led to a compromise solution. In this work, we explicitly reduce the domain gap by employing clean image domain act as the bridge. It is our goal to be able to convert simulated and clinical metal-corrupted data back and forth. As an intermediate product, clean images are our ultimate target. In addition, in the existing MAR methods, noise related features are involved in the entire process, which affects the restoration of clean images. To improve the MAR performance, we proposed a feature selection mechanism and eliminate noise-related features.
- What’s the novelty of this work over current domain adaptation-based networks?
Most of current domain adaptation-based networks attempt to bridge the gap between different domains and then use the unified model to solve the problems. In this work, we first assume that simulated and clinical metal-affected images share a common clean image domain and employ the clean image domain as a bridge to close the domain gap. Clean images are an intermediate result of domain adaptation and also our ultimate goal. To demonstrate the generalization, we also used clinical data collected with different scanning parameters to test the model.
#R3
- From the perspective of the paper, is it reasonable to define the problem as domain adaptation?
Our main purpose is to address the problem it is hard to obtain satisfactory results when we apply the model trained on the simulated data to the real clinical data. That’s the reason why we say this problem belongs to domain adaption.
- The experiment is not convincing. ADN achieved 33.6dB PSNR on SYN dataset in their publication.
The SYN datasets used in this paper and ADN are different. Clean images of ADN are from Deeplesion, while we used Spineweb for simulation, which also contains metal-affected image. The reason for choosing Spineweb is we assume simulated and clinical data share a common clean image domain. To demonstrate the generalization, we also used clinical data from other dataset with different scanning parameters to test the model. We used the released codes of ADN to train the model on our SYN dataset.
- Insufficient innovation in methods. In ADN, networks are used to learn the relationship between noisy images and the original images. Is metal artifact noise an additive noise?
We are not aim to learn the relationship between noisy and the original images. Here, we employ the clean image domain as the bridge to close domain gap between simulated and clinical data. It is our goal to be able to convert simulated and clinical metal-corrupted data back and forth. As an intermediate product, clean images are our ultimate target. In U-DuDoNet [2], metal artifacts have been proven to be additive.
- Some methods are relatively early methods. No comparison of the domain adaptation method.
Recently, IEEE TMI has accepted one paper [1] on addressing MAR domain gap problem. We compare this method and found that the scores on the simulated dataset are as follows: PSNR: 31.22 and SSIM: 0.8471, while our method achieved PSNR: 31.54 and SSIM: 0.8606. We will add the visual results with in the revision.
- The statistical analysis results on clinical data should be given.
The statistical analysis results will be given in the revision.
[1] Deep-learning-based Metal Artefact Reduction with Unsupervised Domain Adaptation Regularization for Practical CT Images [2] U-DuDoNet: unpaired dual-domain network for CT metal artifact reduction
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.
The paper received 1 accept and 2 reject in the initial review phase, and then was brought to the rebuttal phase. The authors provided a rebuttal to address the reviewers’ concerns. R1 strongly recommended to accept this paper in both initial and post-rebuttal evaluations, whereas R2 recommended to reject the paper after reading the rebuttal. R3 did not give any inputs in the post-rebuttal evaluations but left a reject in the initial recommendation. I looked at the rebuttal. I agreed with R2 and R3 on the novelty, motivation, insufficient comparisons to sota methods, and presentation. I do not think the paper is in a good shape at current phase. I recommend to reject the paper.
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 propose a novel semi-supervised Metal Artifact Reduction (MAR) framework to decrease the domain gap between synthetic and clinical data. There are contrasting comments from the reviewers regarding the clarity of the paper, the approach’s novelty, and the use of domain adaptation. Reviewer #1 appreciates the approach and its results, whereas Reviewer #2 and #3 highlight the need for better clarity and broader comparison with existing methods. The authors responded to the reviewers’ concerns, providing explanations for the conceptual novelty, experimental design, and the selection of the model’s features. They clarified their choice of datasets and provided additional comparative data with a recent paper. The paper’s novelty and promising results make it a suitable candidate for acceptance.
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.
Although one reviewer gave positive comments for this submission, two other experts gave negative reviews. After the rebuttal, the reviewers remain unconvinced about the significance of using clean images as a bridge in the domain adaptation process.