Paper Info Reviews Meta-Review Author Feedback Post-rebuttal Meta-Reviews Back to top
List of Papers By topics Author List
Paper Info Reviews Meta-review Author Feedback Post-Rebuttal Meta-reviews

Authors

Haiyang Mao, Yanyang Wang, Hengyong Yu, Weiwen Wu, Jianjia Zhang

Abstract

Metal artifact reduction (MAR) is important to alleviate the impacts of metal implants on clinical diagnosis with CT images. However, enhancing the quality of metal-corrupted image remains a challenge. Although the deep learning-based MAR methods have achieved impressive success, their interpretability and generalizability need further improvement. It is found that metal artifacts mainly concentrate in high frequency, and their distributions in the wavelet domain are significantly different from those in the image domain. Decomposing metal artifacts into different frequency bands is conducive for us to characterize them. Based on these observations, a model is constructed with dual-domain constraints to encode artifacts by utilizing wavelet transform. To facilitate the optimization of the model and improve its interpretability, a novel multi-perspective adaptive iteration network (MAIN) is proposed. Our MAIN is constructed under the guidance of the proximal gradient technique. Moreover, with the usage of the adaptive wavelet module, the network gains better generalization performance. Compared with the representative state-of-the-art deep learning-based MAR methods, the results show that our MAIN significantly outperforms other methods on both of a synthetic and a clinical datasets.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43999-5_8

SharedIt: https://rdcu.be/dnwji

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #3

  • Please describe the contribution of the paper

    The paper proposes a novel Multi-resolution Adaptive Iteration Network (MAIN) for Metal Artifact Reduction (MAR) in computed tomography (CT) images. The proposed method integrates multi-domain, multi-frequency band, and multi-constraint into the scheme by exploiting wavelet transform and formulating such knowledge as a multi-level optimization model. The proposed method derives an iterative algorithm with the proximal gradient technique for solving the model, and an adaptive iteration network is constructed with the guidance of the algorithm, making it more interpretable. The proposed MAIN significantly outperforms existing methods in synthesized and clinical datasets according to the experimental 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 paper has several main strengths. Firstly, the proposed optimization model for Metal Artifact Reduction (MAR) in CT images is a novel formulation that integrates wavelet transform and a multiresolution adaptive iteration algorithm. This approach characterizes metal artifacts into different frequency bands and optimizes the model using a unique multi-level optimization model. Secondly, the paper generates synthesized data with a simulation procedure and uses clinical data collected from postoperative images of pelvic fractures, providing opportunities for controlled experiments and demonstrating the feasibility of the proposed method in real-world scenarios. Finally, the proposed MAIN method is compared with five existing methods, including traditional methods and deep learning-based methods, using RMSE/SSIM as the evaluation metric for synthesized data and visual comparison for clinical data. The experimental results show that the proposed method significantly outperforms existing methods in both synthesized and clinical datasets, indicating its potential for clinical applications.

  • 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 paper has a few potential weaknesses that could be addressed. Firstly, the paper does not provide detailed information about the computational complexity of the proposed method, which could be important for practical applications. Secondly, while the paper provides some implementation details, providing code or an online demo of the proposed method would enable researchers to reproduce the experiments and facilitate further investigations.

  • 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 authors have provided sufficient details on the proposed method’s implementation and evaluation, which should allow for reproducibility of the experiments. The authors plan to provide the code for the proposed method in the future. In the paper, they describe the datasets, experimental setup, and parameters used in detail. Additionally, they reported quantitative results and provided visual comparisons with existing methods. Overall, this information can aid other researchers in recreating and evaluating the proposed method on different datasets.

  • 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

    My first comment relates to the limited clinical dataset used in the evaluation of the proposed method. While the paper demonstrates promising results on the postoperative images of pelvic fractures, it would be beneficial to test the method on larger and more diverse clinical datasets to assess its generalizability and robustness. This could enhance the relevance and significance of your findings and increase the potential impact of your research. Secondly, I suggest providing more ablation studies to investigate the effect of each component of the proposed method on the overall performance. By analyzing the contribution of different parts of the model, you could provide valuable insights into the strengths and limitations of the approach and guide future research.

  • 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 authors present a novel method for metal artifact reduction in computed tomography images that integrates wavelet transform and a multi-resolution adaptive iteration algorithm. The proposed method shows promising results in reducing metal artifacts in both synthesized and clinical datasets and outperforms existing methods. The paper is well-organized and provides comprehensive details on the experimental design and evaluation.

    However, there are some potential weaknesses in the paper that could be addressed. Addressing these limitations could strengthen the relevance and significance of the proposed method and increase its impact.

    Overall, I think that the paper presents an innovative approach to a challenging problem in medical imaging and has the potential to inspire further research in this area.

  • 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

    Based on the author’s reply, I recommend receiving it.



Review #4

  • Please describe the contribution of the paper
    1. The authors explore the spatial distribution characteristics of metal artifacts under different domains and resolutions, and find that metal artifacts show different spatial distributions under different frequency components. The application of wavelet transform enables deep-learning network to extract more abundant features and help the network better learn the distributions of metal artifacts.
    2. To solve the multi-resolution optimization model, the authors derive an iterative algorithm with the proximal gradient technique. They also construct the adaptive iteration network with the guidance of the algorithm, making the network design more interpretable. Since many deep-learning networks are entirely a black-box mechanism, the network built under the guidance of algorithm make the process of reducing metal artifacts more specific, and help us better understand the internal relationships between variables in the process of metal artifact reduction.
  • 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. Metal artifact reduction model with wavelet transform: The authors propose a novel method for metal artifact reduction in CT imaging, which integrates wavelet transform and dual-domain constraints into a multi-resolution optimization model. The authors find that metal artifacts are mainly concentrated in high frequency, and their distribution in wavelet domain is significantly different from that in image domain. Decomposing metal artifacts into different frequency bands enables better characterization of these artifacts. This formulation allows the MAIN to explicitly characterize the artifacts at different spatial frequencies and domains, leading to better artifact suppression.

    2. Iterative optimization algorithm: The authors derive an iterative algorithm with the proximal gradient technique to solve the proposed multi-resolution optimization model. The deep learning methods is often entirely a black-box mechanism. Typically, metal-corrupted images are sent to the network, and the metal-free images are obtained directly. In contrast, the study constructs the deep learning network with the guidance of the optimization algorithm, which make the process of reducing metal artifacts more specific. This gives a deeper understanding of metal artifacts reduction. The network built with the algorithm is shown to achieve better performance than state-of-the-art deep learning-based MAR methods.

    3. Evaluation: Comprehensive experiments are carried out. The authors evaluate the MAIN on both synthesized and clinical datasets and compare it with state-of-the-art MAR methods. The results show that MAIN achieves better performance than other methods in terms of quantitative metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), as well as qualitative visual assessment such as box-plot, the noise power spectrum and the intensity profiles. Moreover, the effectiveness of each important module has been verified with ablative studies.

  • 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 compare the MAIN with state-of-the-art deep learning-based MAR methods, but little comparison is made with non-deep learning-based methods. This limits the ability to assess the novelty and effectiveness of the proposed method compared to other established techniques. So far as I know, linear interpolation (LI) is also a widely used non-deep learning-based method. Adding linear interpolation results in the experiment can enrich the results and make the experiment more convincing.

    2. There is a limited analysis of the algorithm’s convergence properties, such as its convergence rate or stopping criteria, which could be improved upon. Providing a proof of monotonic decrement would also be acceptable. A more rigorous analysis of the algorithm’s performance would provide a better understanding of its behavior and increase confidence in its use.

    3. The ablation study could be more comprehensive, particularly regarding the number of iterations in MAIN, which is an iterative network. Since the iterative structure is important for optimization algorithms and deep-learning networks, investigating the impact of varying the number of iterations on the results could further demonstrate the advantages of the theoretical model.

    4. Although the authors construct an adaptive iteration network that follows the guidance of the optimization algorithm, the proposed model still lacks clear interpretability in terms of how it actually suppresses metal artifacts. Providing more insight into the underlying mechanisms of the MAIN could help improve the understanding of how it works and how it can be further improved.

  • 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

    This paper’s methodology appears to be reproducible. While the authors did not include the code, they provided comprehensive details on the model, dataset, and evaluation. Therefore, the work seems to be replicable to a significant extent. Moreover, the authors plan to provide the code for the proposed method in the future.

  • 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

    1. Code Availability: While you have provided detailed descriptions of your methodology and software specifications, it would greatly enhance the reproducibility of your research if you make your code publicly available. This would allow others to replicate your results more easily or build upon your method. If possible, please consider releasing your code under an open-source license.

    2. Comparison with Non-deep Learning Methods: Little non-deep learning-based methods are compared with the method, which limits the ability to assess the novelty and effectiveness of your proposed method compared to other established techniques. I suggest adding a comparison with non-deep learning-based methods, such as sinogram correction techniques (Linear Interpolation).

    3. The ablative study to verify the effectiveness of network iteration number should be carried. Since I think the iterative structure is important for optimization algorithm and deep-learning network, the impact of verifying the number of iterations on the results can further demonstrate the advantages of the theoretical model.

    4. If the article layout permits, providing a proof of the monotonic descent of the optimization model would enhance our understanding of its behavior and increase confidence in its use.

    Overall, I think your work represents a great contribution to the field of metal artifact reduction in CT imaging. Addressing these comments could help strengthen the validity and impact of your proposed method and improve its ability to be replicated by other researchers.

  • 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 study proposes a novel method for metal artifact reduction in CT images, which outperforms existing methods on both synthesized and clinical datasets. The proposed Multi-resolution Adaptive Iteration Network (MAIN) improve the deep-learning approaches’ performance on MAR by introducing multi-resolution knowledge, iterative optimization algorithm, and adaptive wavelet transform module. The experiments are well-designed, using both synthesized and clinical data and comparing with five other methods. The results show that MAIN achieves better performance than the compared methods in terms of quantitative and visual evaluation metrics. Overall, the paper seems well-written and presents a promising method for metal artifact reduction in CT images.

  • 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

    no



Review #6

  • Please describe the contribution of the paper

    This paper proposes a multi-resolution adaptive iterative network (MAIN) for metal artifact reduction (MAR). The idea of this paper is to find an optimal solution to decompose the metal-corrupted CT image into a clean image and metal artifacts with constraints in the wavelet domain. The network framework is designed as an interpretable optimizer to minimize the regularized loss function. The experiments on synthesized and clinical datasets demonstrate that MAIN outperforms other deep learning-based 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 authors have integrated multi-domain, multi-frequency band, and multi-constraint by exploiting wavelet transform for metal artifacts reduction.

    The authors construct an interpretable adaptive iteration network to optimize the above multi-level model.

  • 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 improvement of the proposed metal artifacts reduction is very marginal with respect to either classical MAR methods or deep learning-based methods. No display windows were provided for any of the CT images, suggesting the clinicians can not interpret these images. More importantly, the authors are using clinically irrelevant metrics (SSIM) for quantitative assessment.

  • 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 have filled out the reproducibility form.

  • 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 Fig.1, what is the meaning of the yellow circle in the fifth subfigure?

    What is the physical meaning of the three constraints f_1 , f_2, and f_3? Are the three constraint functions corresponding to the modules 〖proxNet〗_A, 〖proxNet〗_X, and 〖proxNet〗_U? More interpretation of the constraint functions is better for readers to understand the role of the adaptive wavelet transform.

    The trainable parameters are crucial for the convergence of the network. So how are the trainable parameters initialized?

  • 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 method is innovative. However, the results are not well evaluated. The robustness of the method is also questionable. It is unlikely to be applied in routine practice.

  • 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 multi-resolution adaptive iterative network for metal artifact reduction in CT imaging. The paper received diverse ratings in the reviews - two strong accepts and a weak reject. Therefore, I invite the authors to clear these concerns raised by all the reviewers in the rebuttal phase.



Author Feedback

R3 Q1. The paper does not provide computational complexity. A1: Our network has 5.9M parameters, which are significantly fewer than the competing methods, e.g., 13.4M in DuDoNet and DuDoNet++. Moreover, our model shows more efficient memory usage during the testing stage.

Q2. More ablation studies. A2: We validated the impact of the iteration number (T) and observed an improvement in the MAR performance with increasing T. But we were unable to include these results in compliance with the length restrictions.

R4 Q1. Little comparison to non-deep learning-based methods. A1: DL methods become dominant in MAR, so we mainly compare with them. With the page limit, the traditional method of NMAR included in our paper is one of the most widely applied SOTA methods.

Q2. Proof of monotonic decrement. A2: Leveraging proximal gradient descent, our optimization algorithm inherits the monotonicity property, admitting a consistent decrease of the objective function.

R6 Q1. The improvement is marginal. A1: As highlighted by yellow arrows in Fig 3 and 5, the superior performance of our method over the competing methods has been well demonstrated. Especially, our method excels in reducing metal artifacts and restoring fine-grained structures, whereas the competing methods partially ignore these critical fine structures. Our method’s remarkable performance may not be fully reflected by the numerical metrics since they nearly reach the upper limit, but our significant improvement can be obviously perceived visually.

Q2. No display windows were provided. A2: The display windows of the synthesized and clinical images are [-224 634] HU and [-142 532] HU respectively. They were omitted due to page limit and will be provided in the final version.

Q3. Clinically irrelevant metrics (SSIM) is used. A3: SSIM is used to quantitatively assess the superior performance of our method since it is a widely used metric for assessing image quality in MAR (DuDoNet), sparse-view CT reconstruction (LEARN), and limited CT reconstruction (DIOR).

Q4. What is the meaning of the yellow circle in the fifth subfigure of Fig.1? A4: The yellow circles in Fig 1 highlights the distinct characteristics within different wavelet components. By examining metal artifacts in various wavelet components, the neural network can acquire a more comprehensive understanding of these artifacts and learn from a wider range of underlying image patterns, enhancing its capability in identifying and eliminating artifacts.

Q5. The physical meaning of the three constraints f1, f2, and f3? proxNetA, proxNetX, and proxNetU? A5: f1 and f2 are wavelet constraints for artifacts and images respectively, while f3 is image-domain constraint. f1 and f2 incorporate wavelet priors into model optimization. It helps the network obtain more comprehensive artifact and image information. f3 can improve the model’s sensitivity to image and generalization capabilities

proxNetA, proxNetX, and proxNetU are three network modules designed to optimize the variables A, X, and U, respectively. These modules correspond to the three constraint functions, enabling efficient optimization and enhancing the effectiveness of artifact reduction.

Q6. How are the trainable parameters initialized? A6: All these parameters are set to 0.01 empirically. The initialization of these parameters has little impact on the experimental results due to the remarkable learning capability of the proposed network. While appropriate initial values can expedite the network’s convergence in initial epochs.

Q7. The robustness of the method is also questionable. It is unlikely to be applied in routine practice. A7: Our method has been comprehensively evaluated on various clinical datasets, showing consistent effectiveness and generality. The results suggest that our method holds great potential for routine practice. These results were only partially reported due to space limit and more results will be presented in future version.



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 authors provided a good rebuttal to address the reviewers’ concerns, in particular to answer R6’s questions, who recommended reject in the initial review. In the final rating process, R6 did not give any inputs. I looked at the rebuttal, and combining the other two reviewer’s recommendations, I recommend to accept the paper. Also, the authors should make changes in their paper, as promised, according to the reviewers’ questions, e.g., the parameter setting, computational complexity, etc. In addition, probably a big issue of the paper, that is the title should be modified to show the imaging method, like CT imaging.



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 technical novelty of the proposed method is consistently recognized by all reviewers. Thus, I tend to side with Reviwers #3 & #4 to accept this paper. However, it is worth mentioning that the authors’ rebuttal did not fully address the key concerns regarding the method’s practical applicability (raised by Reviewer #5), i.e., general metrics (SSIM) and simulations (although based on clinical data) are far from clinical practice.



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.

    The paper proposes a novel Multi-resolution Adaptive Iteration Network (MAIN) for Metal Artifact Reduction (MAR) in computed tomography (CT) images. The proposed optimization model for Metal Artifact Reduction (MAR) in CT images is a good try, which integrates wavelet transform and a multiresolution adaptive iteration algorithm. In addition, extensive comparison experiments are included, the author evaluate the MAIN on both synthesized and clinical datasets and compare it with state-of-the-art MAR methods. Some weaknesses are pointed out, e.g., this paper compare the MAIN with state-of-the-art deep learning-based MAR methods, but little comparison is made with non-deep learning-based methods. The ablation study could be more comprehensive. Overall, combining the comments of the reviewer and myself, it is a strong paper with minor weakness.



back to top