List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Jianning Chi, Zhiyi Sun, Tianli Zhao, Huan Wang, Xiaosheng Yu, Chengdong Wu
Abstract
Low-dose computer tomography (LDCT) has been widely used in medical diagnosis yet suffered from spatial resolution loss and artifacts. Numerous methods have been proposed to deal with those issues, but there still exists drawbacks: (1) convolution without guidance causes essential information not highlighted; (2) features with fixed-resolution lose the attention to multi-scale information; (3) single super-resolution module fails to balance details reconstruction and noise removal. Therefore, we propose an LDCT image super-resolution network consisting of a dual-guidance feature distillation backbone for elaborate visual feature extraction, and a dual-path content communication head for artifacts-free and details-clear CT reconstruction. Specifically, the dual-guidance feature distillation backbone is composed of a dual-guidance fusion module (DGFM) and a sampling attention block (SAB). The DGFM guides the network to concentrate the feature representation of the 3D inter-slice information in the region of interest (ROI) by introducing the average CT image and segmentation mask as complements of the original LDCT input. Meanwhile, the elaborate SAB utilizes the essential multi-scale features to capture visual information more relative to edges. The dual-path reconstruction architecture introduces the denoising head before and after the super-resolution (SR) head in each path to suppress residual artifacts, respectively. Furthermore, the heads with the same function share the parameters so as to efficiently improve the reconstruction performance by reducing the amount of parameters. The experiments compared with 6 state-of-the-art methods on 2 public datasets prove the superiority of our method. The code is made available at \url{https://github.com/neu-szy/dual-guidance_LDCT_SR
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43999-5_10
SharedIt: https://rdcu.be/dnwjk
Link to the code repository
https://github.com/neu-szy/dual-guidance_LDCT_SR
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a dual-guidance feature distillation network for the task of Low-dose CT image super-resolution. It is composed of a dual-guidance fusion module (DGFM) and a sampling attention block (SAB). By introducing the 3D inter-slice information and segmentation mask, and a parameter-shared head for both SR and denoise, the network achieves strong 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 experimental results are strong. It is the best compared with the counterparts.\
- The ablaton study is extensive and convincing. All the modules are verified by experimental results.
- The idea using dual guidance is interesting, though I am not sure if it has been used by others.
- 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 presentation is poor. The introduction reads like a literature, and there is only a short paragraph describing what the authors have done. This make the paper not well motivated.
- Some parts are not novel. For example, the idea of using multi-scale features and the cross-channel attention have been commonly used before in existing works. It reads like a combination of existing works.
- Some parts are not convincing. For example, the mask is used as a guidance so I will assume it is important. While it is obtained by a pre-trained segmentation network. What happens if the segmentation mask is not accurate (this will probably happen as there is no perfact system)?
- The visual results are somehow too smooth. I am not sure if there is any information missing with so strong smoothness.
- Details are missing. For example, detailed description of the loss function in sec. 2.2. In experiment, is it a standard protocal that “We choose 1663 CT images for training, 226 CT images for validation and 185 CT images for testing.”? And if so, is the splition the same as others?
- 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
Not sure about this aspect.
- 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, I suggest placing tables and figures at the top of page. 2, Each figure and table should be self-contained. 3, grammar and typo issues: “convolution without guidances causes essential information not highlighted”, “details-clear”, “(SAB). the DGFM”, “The aforementioned methods still exist drawbacks”
- 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 part and address them in the rebuttal.
- 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
5
- [Post rebuttal] Please justify your decision
The authors solved some of my concerns. The rating is consequently improved to be weak accept.
Review #2
- Please describe the contribution of the paper
The contributions of papers can be summarized in three main points:
- the authors design a dual-guidance fusion module (DGFM) model as a super-resolution network for LDCT.
- the authors propose a sampling attention block (SAB) to fuse the extracted features
- the authors design a multi-supervised mechanism based on shared task heads (denoising head and SR head).
- 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 DGFM fuses the 3D CT information and ROI guideance with mutual attention to make full use of CT features and reconstruct clearer textures and sharper edges
-
The SAB extracts the essential multi-scale features by up-sampling and down- sampling to leverage the features in CT images.
-
Shared task heads suppress more artifacts while decreasing the number of parameters
-
- 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 authors obtained a auxiliary pre-train model to generate the segmentation masks, which makes the comparision unfair to other methods. From the ablation studies (the second row), we can see the model without the seg mask is not comparable in SSIM with others. I suspect the improvement is more likely to come from auxiliary information rather than proposed modules.
-
I do not see the ablation studies regarding the effectiveness of Sampling attention block (SAB).
-
The detailed information of some process are not given, for example, what’s the name of the pre-trained segmentation network?
-
- 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
It is hard to re-implement without more details given (please see the weakness)
- 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
Please address my concenrs in the weakness parts.
- 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?
This paper proposes a interesting and effective methods (based on the illustration and main results). However, some details are missing while some experiment is not totally fair. Please address my concerns and I am willing to increase the rate.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
The manuscript proposes an LDCT image super-resolution network including a dual-guidance feature distillation backbone for elaborate visual feature extraction, and a dual-path content communication head for artifact-free and detail-clear CT reconstruction, addressing the fact that low-dose computed tomography (LDCT) suffers from spatial resolution loss and artifacts, which is designed to suppress more artifacts while reducing the number of parameters, and finally validates the superiority of the proposed method on two public 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 concepts of dual-guidance feature distillation and dual-path content communication are innovative.
- 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 method is still under the category of supervised learning and can’t address unpaired CT data training, which is an obvious shortcoming in clinical applications.
- 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
According to the source code provided by the authors, the paper can be reproduced perfectly.
- 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) It is suggested to enrich the caption of Fig.1 with a short description of each module in the figure for the reader’s convenience. 2) The format of Tables 1 and 2 needs to be further optimized.
- 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?
The research topic is highly significant, the presented concept based on dual-guidance feature distillation and dual-path content communication is innovative. The problem is clearly motivated and the paper is interesting to read. Therefore, I recommend the manuscript for acceptance.
- 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 a dual-guidance feature distillation network is presented for low-dose CT image super-resolution reconstruction. The paper got a weak reject and two weak accepts. Therefore, I invite the authors to address the reviewers’ concerns in the rebuttal stage.
Author Feedback
We sincerely thank all reviewers and ACs for their time and efforts. Below please find the responses to reviewers’ comments. 1 Responses to common comments: 1.1 For details and reproducibility issues, we offered codes, data, and weight files in the supplementary files which could perfectly reproduce our method and results. We will provide more details in the manuscript for reproduction. 1.2 For the grammar and composition issues, we will carefully check and reorganize following your comments. 1.3 The first row of Table 1(a) showed the results of using only CT image as input of our proposed network without any auxiliary information, while Table 2 showed those from different methods. The comparisons showed our method could reconstruct more convincing CT images than others even if we did not have the guidance at all. 2 Responses to specific comments from every reviewer 2.1 Responses to Reviewer 1: 1) We considered there existed numerous works for CT image super-resolution and it was necessary to summarize their drawbacks, so that the motivation of this paper could be highlighted. 2) Our network consisted of 3 novel parts: dual-guidance fusion module (DGFM), sampling attention block (SAB), and shared task heads mechanism (STHM). The DGFM introduced mutual attention mechanisms to fuse extra guidance. The SAB used sampling attention, channel attention, and anastomosis connections to exploit the fused features. The STHM introduced the denoising head into SR task to concentrate on the relation between the SR task and the denoising task. 3) We clarified in 1.3 that we could get better results without guidance. In practice, the very simple U-Net generated both convincing and inaccurate segmentation masks like an imperfect system. In this situation, we got better PSRN and SSIM with relatively low standard deviations, indicating the reconstruction results were not deviate from the average even if the masks were not so accurate. Otherwise, the standard deviation would be large in statistically. 4) According to the experts in medical field, our method could reconstruct the clearest structural edges. Meanwhile, our method outperformed others in terms of PSNR correlated with errors highly, which indicates the textures generated by other methods could not match the real distribution and caused the blurring of significant contours. 5) We divided the dataset in an 8:1:1 manner, and all methods used identical datasets. A detailed example could be found in the README of supplementary files. 2.2 Responses to Reviewer 2: 1) We clarified in 1.3 that we could get better results without guidance in a totally fair situation. Only introducing AVG CT may downgrade the performance as you said. Comparing the first and third rows of Table 1(a), we found only introducing mask could improve it a little. However, when both AVG CT and mask were introduced (comparing the third and fourth rows), the performance was further improved. Thus, the improvement not only come from auxiliary mask, but also from the combined effect of the AVG CT and mask, named as “dual-guidance”. 2) Due to page limit, the ablation results were abbreviated in terms of the number of SAB blocks and channels. We found the model with 10 blocks and 64 channels could achieve better PSNR/SSIM (30.4047/0.8974 for 2× and 27.7681/0.8579 for 4× SR) than all the other combinations of parameters. For example, the model with 80 channels provided 27.7681/0.8579 PSNR/SSIM for 4× SR, while the model with 12 SAB blocks provided 29.7571/0.8957 PSNR/SSIM for 2× SR. Those experiments could be reproduced by changing parameters in the train configuration files. 3) We used a pre-trained U-Net in data preprocessing due to its universality in the industry. Any segmentation network could be used here. 2.3 Responses to Reviewer 3 Our method was supervised learning and faced the problem of sparse paired data. Based on this study, we are exploring the semi-supervised method to alleviate the issue of data scarcity.
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 did a good job in their rebuttal. They convinced all reviewers and received consistent recommendations on acceptance. I am happy to recommend to accept 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.
A reviewer modified his/her score after the rebuttal and now the work has three positive reviews, which I tend to agree with.
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.
Three reviewers made some critical comments on the paper, including unclear descriptions, grammatical and typographical issues, and inadequate ablation experiments. In the rebuttal, the authors responded to these issues and promised to make changes in the final version. In the end, two reviewers maintained their previous scores and one reviewer raised the original score of 4 to 5. Three reviewers unanimously agreed that this paper should be accepted. Therefore, I recommend acceptance of this paper.