List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Xinchen Ye, Zheng Sun, Rui Xu, Zhihui Wang, Haojie Li
Abstract
Existing CNN-based low-dose CT reconstruction methods focus on restoring the degraded CT images by processing on the image domain or the raw data (sinogram) domain independently, or leveraging both domains by connecting them through some simple domain transform operators or matrices. However, both domains and their mutual benefits are not fully exploited, which impedes the performance to go step further. In addition, considering the subjective perceptual quality of the restored image, it is more necessary for doctors to adaptively control the denoising level for different regions or organs according to diagnosis convenience, which cannot be done using existing deterministic networks.
To tackle these difficulties, this paper breaks away the shackles of general paradigms and proposes a novel low-dose CT reconstruction framework via dual-domain learning and controllable modulation. Specifically, we propose a dual-domain base network to fully address the mutual dependencies between the image domain and sinogram domain. Upon this, we integrate a controllable modulation module to adjust the latent features of the base network, which allows to finely-grained control the reconstruction by considering the trade-off between artifacts reduction and detail preservation to assist doctors in diagnosis. Experiments results on Mayo clinic dataset and Osaka dataset demonstrate that our method achieves superior performance.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_52
SharedIt: https://rdcu.be/cVRTL
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 proposes a dual-domain end-to-end deep-learning network that predicts CT images from degraded CT images and sinogram data. To enhance the adjustability of the reconstruction process, the authors also propose integrating a controllable modulation module that gives the user room to do a trade-off between denoising and detail preservation. Experiments demonstrate that sinogram data work in the proposed network, and its controllable modulation module can control the degree of denoising.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The motivation is well clarified and meaningful.
- This paper proposed a novel method to learn from the image and sinogram domain, resulting in an improved final reconstruction. The experiments show that the introduced controllable modulation module can control the degree of denoising according to the needs of doctors’ diagnosis.
- 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 for the dual-domain network, the results of PSNR/SSIM outperform other methods in most datasets, but the inference time and model complexity are not considered. They are all important indicators to evaluate the performance of the model. Besides, as mentioned in this paper, “Current methods leveraging both domains by connecting them through some simple domain transform operators or matrices”. This work proposed a complex network but lacked an explanation about why and how these designs work.
-
As for the controllable design, the details of CB are not clear. As mentioned in this paper, the controller parameter f_{\alpha} is learned from fully connected layers, which is dependent on \alpha and learned parameters of fully connected layers. This means that doctors can not “manually” select features based on \alpha because the value of \alpha is different from the value of f_{\alpha}. Besides, the training is conducted in two steps, \alpha=0 for CB and \alpha=1 for MB, which represent two patterns of feature combinations. While the other combination patterns are not trained, it is supposed to show the overall evaluating results under other values of \alpha.
-
The dual-domain reconstruction methods show an overwhelming advantage over image domain reconstruction methods, but the introduction of sinogram data is not the innovation of this paper. So the authors need to consider comparing more dual-domain reconstruction methods to demonstrate the superiority of their method. The following papers can be considered for comparison: [1] Zhang, Zhicheng, et al. “TransCT: dual-path transformer for low dose computed tomography.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2021. [2] Adler, Jonas, and Ozan Öktem. “Learned primal-dual reconstruction.” IEEE transactions on medical imaging 37.6 (2018): 1322-1332.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The method is not clearly clarified, and it is supposed to provide the source code for reproduction. E.g., in Fig.1, the second block in FB takes four inputs, while the description in Section 2.2 shows that there are three inputs, which is not consistent.
For the relatively complex reconstruction network, the authors do not provide enough information on its settings, such as kernel size. Even the setting values of many important hyperparameters are not given, like gamma_1 and gamma_2 in loss functions.
One of the two datasets they used seems to be private too. Currently, the paper doesn’t have a strong 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/2022/en/REVIEWER-GUIDELINES.html
- Demonstrate the need for complex designs of dual-domain networks.
- Make a more comprehensive comparison, including inference time and complexity of reconstruction networks.
- P.2, “Besides, caused by mismatched resolution between the heterogeneous sinogram data and CT image…”: The authors raise the problem of mismatched resolution in dual-domain learning. It is interesting to see how the proposed method solves this problem. Maybe the author can explain it here?
- Experiments: The paper uses an iterative way to reconstruct CT images. How does its model size compare to other methods?
Some suggestions: (1) P.3, “Each DDB is composed of a PI block that transforms …, and a PS block …”: Abbreviations should state the full name on the first occurrence. (2) P.4, “Different colored rectangles are used to represent different branches and different operations in each branch.”: “branches” -> “sub-branches” (3) P.4, Figure 1: The colored rectangles which represent full connected layers and adaptive average pool layers look too similar. Please consider changing the color. (4) P.5, “the image details can be gradually preserved and simultaneously avoid the introduction of mottle noise and streak-like artifacts caused by FBP.” The argument will be more solid if the author can provide some examples of image degradation caused by FBP. (5) P.5, “Each controller block is composed of two convolutional layers with an ReLU in the middle.”: “an” -> “a”. (6) P.7, “We use real clinical dataset authorized…”: “real clinical dataset” -> “a real clinical dataset.” (7) P.7, “The dataset contains ten patients, in which…”: “in which” -> “of which”. (8) P.7, “the edge of the tissue reconstructed through the network is relatively blurred but showing the powerful denoising ability”: “showing” -> “shows”. (9) P.8, “Although the PSNR value is not the highest, but…”: remove “but” (10) P.8, “For fairly comparison, …”: “fairly” -> “fair”. (11) P.8, “which demonstrates the effectiveness of integrating both domain …”: “domain” -> “domains”.
- 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?
There are two obvious problems. First, the clinical value of the proposed strategy is not clear. Second, the proposed method is not reasonable, and experiments do not provide enough evidence that the fMRI feature has hints for predicting gaze patterns. And its improvement over DP-ResNet is not significant in some cases.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
6
- [Post rebuttal] Please justify your decision
The rebuttal answered my concerns, and I am inclined to accept the paper.
Review #2
- Please describe the contribution of the paper
The authors propose a neural network strategy to denoise low-CT images that considers a dual-domain approach (images + sinograms) and also allows to control the level of noise with an additional parameter.
- 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.
- It is an interesting research problem
- The analysis is performed in several databases
- The authors take into account the radiologist perception and not only image processing metrics
- 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 performance improvement with respect to state-of-the-art is not very significant
-
Neither the dual domain nor the controlable modulation are novel ideas
-
- 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
They authors do not share resources to reproduce
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
Perhaps a more thorough discussion explaining the results: why is your strategy a good choice even if it does not outperformed in certain databases; how did the controllable modulation helped in the visualization of certain features?
- 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?
Very nice approach that is novel. Good comparison with different dataset. Still the improvement with respect to state-of-the-art is not very significant
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
design an end-to-end dual-domain deep network to achieve accurate the adjusted low-dose CT recon.
- 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 a dual-domain network to fully explore the mutual dependency between the CT image and raw projection data. 2 design a controllable module to achieve fine tuning of low-dose CT recon process with different noise levels.
- 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 small dimension evaluation. 2 unclear description of the dual-domain module
- 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
The authors followed the rules of ethical and biological requirements for medical data processing.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
The authors may apply their algorithm to the full dimension dataset which has been scanned in CT or CBCT machines. The computational efficiency and memory requirement are exposed to the real clinical data. Dual domain network was proposed a few years ago. In previous publications, authors tried to link the info between image and projection domain. A more comprehensive evaluation of priors are expected in the introduction. Subtle improvements compared with existing algorithm. The real utility of the method is not clear.
- 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 authors are working on an old topic of low-dose CT imaging. Network designed in this manuscript is not new as the components within the network has been applied previously. Results are not impressive.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
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 presents a dual-domain (images + sinograms) end-to-end deep-learning network for low-dose CT image reconstruction, which also allows to control the level of noise with an additional parameter. As the reviewers agreed that the presented work is interesting, they raised several concerns, including
- The authors should have considered comparing their method with other dual-domain reconstruction methods to demonstrate the superiority of their method, like the ones pointed out by R1.
- The performance improvement with respect to state-of-the-art doesn’t seem to be significant.
- Neither the dual domain nor the controlable modulation are novel ideas.
- The reviewers have concerns about the reproducibility of the work.
- The description of the dual-domain module is unclear. (R1 & R3)
The authors are invited to provide a rebuttal.
- 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).
6
Author Feedback
To all reviewers: Thanks for all the positive feedbacks and constructive comments. The major concerns are listed as follows:
1) About our contribution. First, our key contribution is how to make full advantage of the dual-domain information. Current methods use hand-designed prior matrix to realize feature integration between both domains, which may introduce unnecessary artifacts by the over-idealized and fixed domain transformation matrix. In contrast, we realize the dual-domain interaction with a novel implicit learning architecture through adaptive pooling layer without any hand-designed modules, thus making the network better exploit the mutual benefits between both domains. Note that, due to the characteristics of adaptive pooling, it can automatically calculate the convolution kernels and strides to adjust the resolution mismatch between features of different domains, and map the data of one domain to the other domain without explicitly considering the domain transformation between both domains. The parameters in it can be totally learned through network training. Second, the controllable modulation we design is very necessary for doctors to control the denoising level for different regions or organs according to diagnosis convenience, which cannot be done by existing deterministic networks proposed by other works. We are the first to consider the interaction with doctors in real CT reconstruction scenarios. Both parts demonstrate our contributions against other SOTA methods.
2) About our performance. It should be noted that our purpose is to provide superior CT images to help the diagnoses, but objective indicators such as PSNR and SSIM may not represent the truly-wanted CT images. Therefore, the objective results we give in the manuscript aims to show our comparable (or even better) performance to other methods. The more important goal is to verify the effectiveness of the controllable modulation that allows the doctor to control the reconstruction results. Ablation study and Fig.3 in the manuscript have shown the overall evaluating results under different controlling values of \alpha, which adequately verifies the effectiveness of our method.
3) We will release our code and model for reproducibility.
Other questions in terms of each reviewer are listed as follows:
To Reviewer #1： 1) As for our dual-domain branch, we have six DDBlocks and each DDB mainly contains four convolutional blocks with some adaptive poolings and a simple attention operator, which is a relatively small backbone and running faster than state-of-the arts. Once adding the controllable modulation part, the inference time may increase somewhat. We will give the complexity performance in the final version.
2) Indeed，f_\alpha is learned directly from \alpha, and has consistent correspondence with \alpha. Thus, once choosing an \alpha, the f_\alpha is determined. Since f_\alpha is a high-dimensional vectors, the doctors can use a scalar \alpha to control the results instead of directly choosing features in the feature space. For training, we just need to train two extreme cases, i.e., \alpha=0 and \alpha=1 to obtain their corresponding results. The results of intermediate cases can be directly obtained in testing through the adaptive fusion from two branches. Fig.3 in the manuscript has shown the overall evaluating results under other values of \alpha (0.2-0.8).
3) We will add the objective performance of [1] and [2] in final version.
To Reviewer #2： We will give a more thorough analysis of the explanation of the objective results and the controllable modulation in the final version.
To Reviewer #3： Due to the computational efficiency and memory requirement, most methods still focus on singe-slice CT reconstruction (two-dimension), and our method also falls in this category. Dealing with full dimension (3D) data and designing corresponding lightweight architectures is out of our scope. We will consider this in future work.
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 rebuttal helps mitigate the reviewers’ concerns on novelty, performance and validation of the proposed method.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
2
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The strenghts of the work include an interesting formulation of the reconstruction problem with the the opportunity to trade off between denoising and detail preservation. It is my perception that the rebuttal successfully addressed the most severe weaknesses, prompting reviewers to upgrade their ratings. I am in agreement with this decision.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
1
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.
This paper works on low-dose CT reconstruction with controllable denoising via learning from both the image and the sinograms domains. Although dual-domain approaches are not novel in CT reconstruction, this work focuses on the controllable modulation that allows the doctor to control the reconstruction results, which is of interest. The authors’ responses to the concerns of novelty and performance comparison are, to some extent, convincing. The reviewers have consensus that the paper is above the acceptance bar.
- 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).
4