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

Rongjun Ge, Yuting He, Cong Xia, Hailong Sun, Yikun Zhang, Dianlin Hu, Sijie Chen, Yang Chen, Shuo Li, Daoqiang Zhang

Abstract

The low-dose computed tomography (CT) scan is clinically significant to reduce the radiation risk for radiologists and patients, especially in repeative examination. However, it inherently introduce more noise due to the radiation exposure. Nowadays, the existing LDCT reconstruction methods mainly focus on single domain of sinogram and image, or their cascade. But there still has limitations that the insufficient information in single domain, and the accumulation error in cascaded dual-domain. Though dual-domain can provide more information in reconstruction, how to effectively organize dual-domain and make complementary fusion still remain open challenges. Besides, the details inter-pixel in reconstructed LDCT is essential for structure maintenance. We propose a Dual-domain parallel network (DDPNet) for high-quality reconstruction from widely accessible LDCT, which is the first powerful work making parallel optimization between sinogram and image domains to eliminate the accumulation error, and fusing dual-domain reconstructions for complementary. DDPNet is constituted by three special designs: 1) a dual-domain parallel architecture to make joint mutual optimization with interactive information flow; 2) a unified fusion block to complement multi-results and further refine final reconstruction with triple-cross attention; 3) a pair of coupled patch-discriminators to drive the reconstruction towards both realistic anatomic content and accurate inner-details with image-based and inter-pixel gradient-based adversarial constraints, respectively. The extensive experiment validated on public available Mayo dataset show that our DDPNet achieves promising PSNR up to 45.29 dB, SSIM up to 98.24%, and MAE down to 13.54 HU. in quantitative evaluations, as well as gains high-quality readable visualizations in qualitative assessments. All of these findings reveal our method a great clinical potential in CT imaging.

Link to paper

DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_71

SharedIt: https://rdcu.be/cVRUf

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 paper provides a novel dual-domain parallel network for low-dose CT reconstruction. The network is composed of a dual-domain parallel architecture, a unified fusion block and a pair of coupled patch-discriminators. The proposed method achieved good noise reduction and structure maintenance.

  • 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.

    A novel dual-domain network, which is parallelly designed and not cascaded as the previous work.

  • 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 logic and readability of the article are relatively poor. a) The introduction to loss functions is ambiguous. The respective effects of the L1, SSIM, and adversarial losses should be properly explained. b) On page 6, the sentence “With the rich experience granted from the …” is lengthy and grammatically incorrect, making it difficult for the reader to understand. c) What does the “coherent constraint” refer to? How is it established in dual-domain cross-communication?
    2. In Fig. 2, the arrowhead should point at the Unified Fusion Block, but not the image ysin~img.
    3. There are many grammar and spelling mistakes. a) In the Abstract, “The extensive experiment” should be “The extensive experiments”. In the same sentence, the period “.” after “13.54 HU” should be dropped. b) In the Abstract, The sentence “All of these findings reveal our method a great clinical potential in CT imaging” can be replaced by “All of these findings suggest that our method has great clinical potential in CT imaging”. c) The word “readable” is repeated in the sentence “how to effectively make LDCT reconstruction in the more readable readable pattern…” on page 2. d) “to comprehensively integrates” should be “to comprehensively integrate” on page 4. e) “more realistic and details” should be “more realistic and detailed” on page 4. f) “Fig. 3: IIF connect” should be “Fig. 3: IIF connects” on page 5. g) What does the sentence “IIF extracts the intrinsic noise distribution, and guide it image domain …” mean on page 5? Does it mean “IIF extracts the intrinsic noise distribution and introduces it into the image-domain stream …”? h) “the sinogram domain extracted information” should be “the extracted information from the sinogram domain” on page 5.
    4. Lack of details, especially in section 3.1
  • Please rate the clarity and organization of this paper

    Poor

  • 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

    Average reproducibility. The description of the loss function and the coupled patch-discriminators is rather vague. The paper provided the details of materials and configurations, but did not tell the parameter in the loss function. Therefore, the reproducibility is not satisfactory.

  • 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. Please provide more details in section 3.1, such as parameters setting in the network.
    2. In Fig.5, the improvement of visual quality of the DDPNet is not obvious. Although the tiny structure is easier to observe compared with other methods, its shape has possibly changed. Please give more explanations. In addition, as seen in the enlarged ROI of DDPNet, the entire ROI is shifted a little bit to the right, which indicates the inaccuracy of DDPNet. Please check if it is due to wrongly selected ROI or the drawback of DDPNet.
    3. There are many typos in the paper. For example, ‘more readable readable pattern’ in Page 2, ‘is proposed to makes’ and ‘grade-based discriminator’ in Page 3, arrow direction from the fusion block to y_sin~img in Fig.2, ‘signogram-domain stream’ in Page 4, and so on. Please read the paper carefully and revise them.
  • 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?

    Good novelty, poor clarity, poor rigorousness.

  • 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

    6

  • [Post rebuttal] Please justify your decision

    The authors have carefully addressed our questions.



Review #2

  • Please describe the contribution of the paper

    The authors proposed a Dual-domain parallel network (DDPNet) for low-dose CT (LDCT) reconstruction. A parallel structure is proposed to make parallel optimization between sinogram and image domains streams to reduce the cumulative error caused by the process of the sinogram domain denoising. The dual-domain information is used complementarily by Interactive Information Flow (IIF) mechanism. Then the authors design a triple-cross attention block to fuse features from two domains.

  • 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 is written clearly, and the motivation is well justified.
    2. The proposed method utilizes both image and sinogram domain information through an IIF mechanism to reconstruct high-quality CT images from LDCT images. The cumulative error can be suppressed by parallel optimization between sinogram and image domains streams.
  • 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. This work is quite similar to the following work that explored the interactive dual-domain parallel network. Although slightly different tasks, the motivation is the same. Without citing this related work and discussing the differences, this paper seems to lack of novelty. Wang, Tao, et al. “IDOL-Net: An interactive dual-domain parallel network for CT metal artifact reduction.” arXiv preprint arXiv:2104.01405 (2021).
    2. Fig. 5 showed that the proposed method has HU shift compared to NDCT images; the obtained images are darker. Such HU shift certainly limit its potential in clinical practice.
    3. The authors use the downsampled image and sinogram in the dual-stream framework. But how to upsample the image is not mentioned in this article. The parameters of the FBP algorithm need to be scaled for the downsampled sinogram. If upsampling after reconstruction will bring errors. How can the authors eliminate this error?
    4. A triple-cross attention block is designed in this article to combine the features extracted by the two-stream structure and LDCT images. However, the advantages of the block are not clearly stated. If it is feasible to directly use a 1*1 convolution kernel or a fully connected layer for feature fusion? How much will the performance drop?
    5. In subsection 2.3, four operators are used to calculate the gradient image. Usually, only the first two operators are needed to calculate the gradient image. What are the functions of the last two operators, and how much does it improve the performance of the final model?
    6. How to set the weight parameter “alpha” in the loss function is not mentioned in this article. How to combine the adversarial loss of the gradient image and the original image is also not mentioned.
    7. The capitalization of symbols is not uniform. In “Overall performance” of subsection 3.2, “db” is lowercase. But in the following “dB” is capitalized.
  • 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 paper did not provide the detailed network architecture and key parameters such as alpha in the loss function, which make the reproducibility questionable.

  • 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. It would be better to clarify the difference between this work and most related works.
    2. It would be better to compare the most related works to show the effectiveness of the proposed method.
    3. It would be better to investigate the HU shift between the proposed one and the ground-truth as this limits the clinical value.
    4. Key details should be provided for reproducibility.
  • 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

    3

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

    The novelty is limited as the interactive dual domain has been investigated in literature; the authors did not cite and compare them. The detailed network architecture and hyperparameters were not provided. Most importantly, the results have HU shift compared to ground-truth one, which limits its clinical value.

  • Number of papers in your stack

    8

  • What is the ranking of this paper in your review stack?

    6

  • 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

    Thank the authors for their clarification in the rebuttal. It seems that the proposed method was built upon the IDOL-Net with some appropriate modifications for low-dose CT. The comparison with IDOL-Net (without mask) should be considered. Most importantly, IDOL-Net should be cited and discussed. Although the SSIM/PSNR are very high, the image quality in Fig. 5 gives a different impression (looks darker). Therefore, I increased my initial rating.



Review #3

  • Please describe the contribution of the paper

    The authors propose a novel DDPNet for low-dose CT denoising. They design special modules for DDPNet, including dual-domain parallel architecture, a unified fusion block using multi-head attention and coupled patch-discriminators. Extensive experiments prove the effectiveness of each module and the best performance of DDPNet.

  • 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 propose a dual-domain parallel solution with interactive information flow to fuse the dual-domain features and eliminate the accumulation error.
    • The unified fusion block use multi-head attention to better fuse dual-domain features.
    • Quantitative results and visual comparisons show the superior performance of DDPNet.
  • 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.
    • As authors claim DDPNet has a great potential in CT image, they should try the method on real clinical data.
    • The authors should perform ablation studies on different architectures of dual-domain frameworks to claim DDP can eliminate the accumulation error, like SD-ID, ID-SD.
  • 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 use public datsets but do not share the codes.

  • 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 arrowhead in Fig2 should be pointed from y_{sim~img} to Unified fusion block.
  • 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 performance of DDPNet is impressive since it is hard to improve PSNR from 44.90 to 45.29.

  • Number of papers in your stack

    5

  • What is the ranking of this paper in your review stack?

    1

  • 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

    6

  • [Post rebuttal] Please justify your decision

    The rebuttal resolves my concerns and I maintain the original rating.



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 proposes a novel dual-domain parallel network for low-dose CT reconstruction. The reviewers have provided and diverging scores and raised some concerns to be addressed in rebuttal. The key points to be addressed can be summarized as: 1) Similarity with the dual-domain parallel network and key methodological contributions of the paper; 2) Intuition behind the performance increase with triple-cross attention block 3) Lack of ablation studies and appropriateness of the experimental setup.

  • 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).

    9



Author Feedback

We thank R1 and R3 for accepting our paper as “Good novelty”, “good noise reduction and structure maintenance”, “best performance”, and thank R2 for appreciating our work as “written clearly”, “motivation is well justified”.

We thank AC for supporting our work as “novel” and summarizing 3 suggested rebuttal points.

1.{R1}Coherent constraint(CC) A1: CC refers to that the training of each domain stream gets constraint from the supervision of another domain, established by the information exchange in IIF. It enables comprehensive guidance for each domain stream, as sinogram- and image- domains are various forms generated from same CT.

2.{AC,R2}Discussion with interactive dual-domain parallel network(IDOL-Net) and our contribution A2:
Differences & contributions: 1)Although our method and IDOL-Net process image and sinogram in parallel, our parallel processing(PP) is specially designed for LDCT reconstruction with novel IIF to exchange information between parallel branches. Because LDCT reconstruction is generally converted as estimating the difference between LDCT and NDCT which can’t be directly transformed between domains with FP/FBP. The IIF makes space mapping and multiple calculations with dual-domain LDCTs to transform LDCT-NDCT difference got from one domain into another. So that vision difference information is exchanged to enhance visual interpration for sinogram-domain, and intrinsic projection difference is exchanged to complement CT essential information for image-domain. Besides, the IIF makes information compression and adjustment to avoid covering up the backbone and specific information of main domain, and the additional error caused by inter-domains sensitivity. However, IDOL-Net for metal artifact reduction adopts direct estimation of clean CT, and just uses simple FP/FBP and direct concatenation in PP. 2)We specially design Unified Fusion Block(UFB) to fuse the dual-domain results by cross reference among them and original input, so that makes refinement of reducing error and enhancing accurate reconstruction from each domain result, and getting the lost structure from the input. 3)Moreover, we design patch discriminator for image gradient in all directions to drive the correct change distribution between adjacent pixels for image details enhancement.

3.{AC,R2}Intuition behind the performance increase with triple-cross attention(TRA) block A3: TRA block,i.e. UFB, is actually designed to unify dual-domain results and make refinement for the final results. It makes cross reference among image-domain result, sinogram-domain result and LDCT input by integrating the triple and making self-attention with multi-head. Benefit from the adjustment by attention, the further refinements of reducing error and enhancing accurate reconstruction from each domain result, and getting the lost structure from the input, are made. To avoid noise from LDCT input, a residual connection linking dual-domain results is used at the end. All these thoughtful designs enable performance increasement.

4.{AC,R3}Ablation studies A4: Extensive ablation studies is made in Table1 to comprehensively evaluate each component. The study of different dual-domain architecture that R3 interested in, is actually made in Table2 comparing our method and CLEAR that is a SOTA dual-domain net in cascade form SD-ID. ID-SD is unacceptable as ID damages projection information, so next SD is meaningless. Our parallel structure DDPNet and its reduced versions all show superior results.

5.Experimental setup A5: Our experiments are set up for fair evaluation on public data. Evaluation metrics are MAE, SSIM and PSNR widely-used for LDCT. Extensive ablation studies for each component and comprehensive comparison with SOTA LDCT methods of all types including image-, sinogram-, and dual- domains, are also made. We used default weights 0.84 and 0.16 between L1 loss and SSIM loss, while the weight of adversarial loss is 0.005 based on experimental experience.



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 concerns on Similarity with the dual-domain parallel network and intuition behind the performance increase with triple-cross attention block are addressed in the rebuttal.I believe the work has value for the MICCAI community and I recommend acceptance.

  • 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).

    7



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.

    This paper presents a new dual-domain parallel network for low-dose CT reconstruction, consisting mainly of a dual-domain parallel architecture, a unified fusion block, and a pair of patch discriminators. The reviewers raised major questions about the methodological contribution and experimental details. After rebuttal, two reviewers improved their scores and stated that the feedback addressed the main issues. AC thinks that the feedback letter responded to the questions in detail. Although R3 expressed concerns about the experimental section, the overall strengths of the paper are more than the weaknesses and AC recommends acceptance.

  • 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).

    8



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 authors proposed a Dual-domain parallel network (DDPNet) to reconstruct low-dose CT (LDCT). The reviewers agreed on the superior quantitative results of the proposed method. The criticisms of R2 have been partly addressed in the rebuttal, lifting the paper over the acceptance threshold. The final version of the paper should include reviewers’ comments, e.g., discuss and compare IDOL-Net.

  • 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).

    7



back to top