List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yuchen Fei, Chen Zu, Zhengyang Jiao, Xi Wu, Jiliu Zhou, Dinggang Shen, Yan Wang
Abstract
Positron emission tomography (PET) is a pervasively adopted nuclear imaging technique, however, its inherent tracer radiation inevitably causes potential health hazards to patients. To obtain high-quality PET image while reducing radiation exposure, this paper proposes an algorithm for high-quality standard-dose PET (SPET) synthesis from low-dose PET (LPET) image. Specifically, considering that LPET images and SPET images come from the same subjects, we argue that there is abundant shared content and structural information between LPET and SPET domains, which are helpful for improving synthesis performance. To this end, we innovatively propose a bi-directional contrastive generative adversarial network (BiC-GAN), containing a master network and an auxiliary network. Both networks implement intra-domain reconstruction and inter-domain synthesis tasks, aiming to extract shared information from LPET and SPET domains, respectively. Meanwhile, the contrastive learning strategy is also introduced to two networks for enhancing feature representation capability and acquiring more domain-independent information. To maximize the shared information extracted from two domains, we further design a domain alignment module to constrain the consistency of the shared information extracted from the two domains. On the other hand, since synthesized PET images can be used to assist disease diagnosis, such as mild cognitive impairment (MCI) identification, the MCI classification task is incorporated into PET image synthesis to further improve clinical applicability of the synthesized PET image through direct feedback from the classification task. Evaluated on a Real Human Brain dataset, our proposed method is demonstrated to achieve state-of-the-art performance quantitatively and qualitatively.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_50
SharedIt: https://rdcu.be/cVRTJ
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 proposed a framework for low-dose PET reconstruction. The method consists of three components: a domain alignment module to regularize the consistency of the shared information between low-dose and standard-dose PET, a contrastive learning strategy to enhance domain-independent information, and a classifier to ensure the accurate diagnosis-related features.
- 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 is well-written and well-organized.
- The motivation of the method is very clear and well-explained.
- The method is relatively novel.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- The evaluation needs improvement.
- The dataset is too small.
- The performance improvement is subtle.
- 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
Code is not provided.
- 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 quality of the low-dose PET in Fig.2 is actually quite good. It leads to the question whether this application is meaningful. This can be proved by either the similar performance on a much higher dose reduction rate (visually worse image quality) or the comparison of the diagnosis (e.g. classification of NC vs MCI) on low-dose and each synthesized standard-dose PET.
- The model is quite big for a dataset of 16 subjects. Any overfitting problem?
- In table 1 and 2, the results show limited improvement. Statistical tests are needed.
- Since the classifier of NC vs MCI is already a part of the model, it would be natural to show the diagnosis of low-dose, synthesized, and ground-truth PET, and compare them.
- In Fig.2 and Fig.3, please add zoom-in image. It’s hard to differentiate the quality between methods.
- In introduction, it says current methods “do not take into account the applications of the synthetic images in analytical and diagnostic tasks”. This paper actually used the similar method.
Ouyang, Jiahong, Kevin T. Chen, Enhao Gong, John Pauly, and Greg Zaharchuk. “Ultra‐low‐dose PET reconstruction using generative adversarial network with feature matching and task‐specific perceptual loss.” Medical physics 46, no. 8 (2019): 3555-3564.
- 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 method is ok.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- 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
Review #3
- Please describe the contribution of the paper
In this paper, the authors hypothesized that the abundant shared content and structure information between LPET and SPET can help improve image synthesis performance. Based on this, the authors proposed a BiC-GAN framework that contains a master network and an auxiliary network to extract shared information from LPET and SPET. Both networks implement intra-domain reconstruction and inter-domain synthesis tasks, aiming to extract shared information from LPET and SPET domains, respectively. Additional contrastive learning and classification tasks were also used to boost the performance. The proposed method achieved results comparable to the state-of-the-art 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 idea to use shared semantic content and structure information between LPET and SPET in training is novel. Extensive evaluation with comparison with the state-of-the-art methods.
- 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 dataset in this study is small, with only 16 subjects in total. The authors also have to do 2D image processing to obtain more 2D slice samples, which is a limitation. The proposed method is likely not statistically different from the SOTA AR-GAN. It would be informative to discuss the pros and cons of the proposed method compared with the SOTA AR-GAN in different aspects, or how the proposed method can outperform AR-GAN in certain scenarios.
- 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 method description is clear.
- 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 proposed method obtained similar results compared with the current SOTA method AR-GAN, which is good, but it would be better if the authors can discuss the pros and cons of the proposed method compared with the SOTA AR-GAN in different aspects, or how the proposed method can outperform AR-GAN in certain scenarios.
- In Eq.(2), in the last term, I think it should be D_M(l, s_syn) instead of D_M(s, s_syn), according to Fig.1.
- In the “Contrastive Learning Module” section, the local features are processed by a 3x3 conv kernel. Since the local feature spatial dimension is 2x2, why use a 3x3 conv kernel?
- The symbol “C” is referred to both the classifier (in Section 2.4) and the positive constant in Eq. (6). I suggest the author to change one of 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Novel idea and strong evaluation.
- Number of papers in your stack
5
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #4
- Please describe the contribution of the paper
The paper studies how to generate high-quality standard-dose PET synthesis from low-dose PET images. The authors propose a bi-directional contrastive generative adversarial network (BiC-GAN), including a master network and an auxiliary network, for intra-domain reconstruction and inter-domain synthesis tasks. Moreover, a domain alignment module is designed to maximize the mutual information from two domains. Also, the mild cognitive impairment (MCI) classification task is incorporated into PET image synthesis. The authors demonstrate the robustness of the method compared with the state-of-the-art qualitatively and quantitatively.
- 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 is well organized, and the motivation behind low dose PET noise reduction is well justified.
- The work introduced a novel end-to-end bidirectional contrastive GAN (BiC-GAN), including a master network and an auxiliary network, for intra-domain reconstruction and inter-domain synthesis tasks.
- The authors proposed a domain alignment module to maximize the maximum mutual information between the two domains.
- The work achieves better results compared with the state-of-the-art methods quantitatively and visually.
- 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.
- This paper combines several existing technologies to form a new framework, thus the novelty is not significant. This is OK if the performance is extremely expressive, or the analysis is sufficient.
- Although the quantitative and qualitative performances of the proposed method described in the paper have shown better than state-of-the-art methods, the performance gains seem marginal, as shown in Table 2 and Figures 2,3. The magnitude of the improvement of the proposed method remains unclear.
- Although the overall architecture is novel, each individual components are largely inspired by previous works.
- To overcome marginal improvements, the authors should compare more recent works. Furthermore, the experiments on more benchmarks should be presented to verify robustness.
- Lack of enough ablation study analyzes the contributions of individual components to the final performance.
- 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
Will code be available? It is not mentioned that code is made available.
- 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
- This paper combines several existing technologies to form a new framework, thus the novelty is not significant. This is OK if the performance is extremely expressive, or the analysis is sufficient.
- Although the quantitative and qualitative performances of the proposed method described in the paper have shown better than state-of-the-art methods, the performance gains seem marginal, as shown in Table 2 and Figures 2,3. The magnitude of the improvement of the proposed method remains unclear.
- To overcome marginal improvements, the authors should compare more recent works. Furthermore, the experiments on more benchmarks should be presented to verify robustness.
- Although the overall architecture is novel, each individual components are largely inspired by previous works.
- Lack of enough ablation study analyzes the contributions of individual components to the final performance.
- The authors seem to have missed some relevant literature. Specifically they miss out on several relevant citations, e.g. “ CT Super-Resolution GAN Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE)”, “”Unpaired brain MR-to-CT synthesis using a structure-constrained CycleGAN”, and “CycleGAN denoising of extreme low-dose cardiac CT using wavelet-assisted noise disentanglement”.
- 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?
- Originality: Generative adversarial network and contrastive learning have been investigated in PET image synthesis. However, this paper combines them for PET analysis tasks. Novelty is limited.
- Quality: The paper introduces the model and provides few experimental evaluation. Especially, another contribution claimed by the authors is the loss function. However, the authors did not provide complete experimental results to analyze the impacts of each components of the loss function.
- Clarity: The paper is well-structured and mainly well-written.
- Significance: The ideas used here could be utilized in different medical image analysis tasks. The paper studies how to generate high-quality standard-dose PET synthesis from low-dose PET images with a bi-directional contrastive generative adversarial network (BiC-GAN).
- Number of papers in your stack
7
- What is the ranking of this paper in your review stack?
3
- 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.
Reviewers unanimously agreed that the paper provides important contributions and should be accepted for presentation at the MICCAI meeting. The introduction of the classification module into the image reconstruction framework is very important as most image reconstruction methods do not provide such a module.
Reporting the classification results would further improve this work. The authors should also address the important comments raised by the reviewers in the final version of the work.
- 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).
3
Author Feedback
We thank all the reviewers for their acknowledgement about our methodological contribution, and their constructive comments for further clarification. We have carefully studied the comments, and will address these questions in the final paper.