Authors

Shixuan Chen, Boxuan Cao, Yinda Du, Yaoduo Zhang, Ji He, Zhaoying Bian, Dong Zeng, Jianhua Ma

Abstract

The harmful radiation dose associated with CT imaging is a major concern because it can cause genetic diseases. Acquiring CT data at low radiation doses has become a pressing goal. Deep learning (DL)-based methods have proven to suppress noise-induced artifacts and promote image quality in low-dose CT imaging. However, it should be noted that most of the DL-based methods are constructed based on the CT data from a specific condition, i.e., specific imaging geometry and specific dose level. Then these methods might generalize poorly to the other conditions, i.e., different imaging geometries and other radiation doses, due to the big data heterogeneity. In this study, to address this issue, we propose a condition generalization method under a federated learning framework (FedCG) to reconstruct CT images on two conditions: three different dose levels and different samplings at three different geometries. Specifically, the proposed FedCG method leverages a cross-domain learning approach: individual-site sinogram learning and cross-site image reconstruction for condition generalization. In each individual site, the sinogram at each condition is processed similarly to that in the iRadonMAP. Then the CT images at each site are learned via a condition generalization network in the server which considers latent common characteristics in the CT images at all conditions and preserves the site-specific characteristics in each condition. Experiments show that the proposed FedCG outperforms the other competing methods on two imaging conditions in terms of qualitative and quantitative assessments.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43898-1_5

SharedIt: https://rdcu.be/dnwAC

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #1

Please describe the contribution of the paper

This paper presents a deep learning reconstruction strategy in the federated learning framework to reconstruct CT images from different conditions, including different dose levels, different geometries and different sampling conditions. Extensive experimental results demonstrated that the proposed strategy outperforms the other competing methods in terms of qualitative and quantitative assessments.
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 proposed method can be treated as the extension of iRadonMAP to the federated learning framework. In each client, the sinogram at each individual condition is processed similarly to that in iRadonMAP, then the latent characteristics of reconstructed CT images at all conditions are processed through the framework via a condition generalization network in the server. This can address the issue of data privacy. The proposed method is novel and interesting, which can provide a new strategy for low-dose CT reconstruction.
2. The server in the proposed method holds a large amount of labeled data that is close to the real world, which is another contribution of the submission.
3. Extensive experiments on different conditions are conducted to demonstrate its reconstruction performance.
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 paper should review more recently developed methods in the introduction.
2. The low-dose simulation details are missing.
3. In the experiment, three clients with more noise levels should be added to demonstrate its robustness.
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 proposed method is reproducible with the details provided in the paper.
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. The paper should review more recently developed methods in the introduction.
2. The low-dose simulation details are missing.
3. In the experiment, three clients with more noise levels should be added to demonstrate its robustness.
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 FL framework in the CT imaging field can address the issues of data sharing and data privacy. With the help of the FL framework, the iRadonMap can process data from different geometries simultaneously, which is an interesting and attractive topic in the CT imaging field. Moreover, this paper provides a detailed description, which is easy to follow and reproduce.
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

This paper proposes a condition generalization method under a federal learning framework for CT reconstruction under different scanning conditions(low-mAs, sparse-view and limited-angle). This method shares only the network parameters of image domain, thus keep the sinogram domain network as a specificity preserving part. Unlike general federal learning frameworks, in this method, there is condition generated image domain network trained by large amount noisy image and participates in aggregation with an independent weight to enhance local model robustness. The results demonstrate that this method achieves robust CT reconstruction under multiple scanning conditions.
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.

This method proposes a new federal learning framework for CT reconstruction, which reduces model performance degradation due to perturbations from CT imaging geometries and scanning protocols, and provides performance improvement relative to models trained on local datasets. The method can be well applied in multi-site collaborations. The proposed method is logic and easy to follows. The strategy of partial parameter sharing preserves the local specificity and constructs a continuously trained conditional generation image domain network, which is attractive.
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.

To verify whether the local models are robust under different distributions data, more ablation experiments are needed to verify the effectiveness.
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 is basically reproducible, as it gives a clear procedure for the experiments and lists the settings of parameters, also, the network architecture is shown in the supplementary material. The researchers can reproduce the method easily.
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

A more detailed review of the literature in the research area would have suggested the quality of the manuscript.
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 CT imaging with FL framework in attractive and new in the CT field, and it can overcome the data privacy in the deep learning network. The proposed method provide a new way in FL-based CT imaging task.
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 #2

Please describe the contribution of the paper

This work proposes a condition generalization method by using a federal learning framework and the cross-domain strategy, in order to generally reconstruct CT images among different scanning conditions (low-mAs, sparse-view and limited-angle), which is termed as “FedCG”. All results demonstrate that this method can obtain a robust CT reconstruction performance under various scanning conditions.
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.

Each model in the client consists of a inaccessible sinogram-domain sub-network and a accessible image-domain sub-network for generally reconstruct CT images among vary scanning conditions. In addition, compared to the existing federal learning frameworks, the proposed FedCG method utilizes extra labeled data to finetune the model in the central server. Overall, the proposed method in the manuscript is interesting and 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 network parameters should provide in the manuscript.
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 author provides detailed information such as the model and datasets. Therefore, the reproducibility of the paper is great and sufficient.
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). The authors should add the corresponding details about training parameters. (1) More literature citations should be added in second paragraph in section Introduction to Increase the reliability of conclusions. (2) In section Abstract, the proposed FedCG method is applied on two conditions (“i.e., on two conditions: three different dose levels and different sampling shcemes at three different geometries”). In the last paragraph of the section Introduction, the proposed FedCG method is applied on three condition (“different conditions i.e., different dose levels, different geometries, and different sampling shcemes”). And the work exhibits the results on two conditions. Please unifying the expression of “two conditions” or “three condistions”. If it is “three condistions”, please add the corresponding results on condition # 3.
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?

This paper presents a condition generalization method by using a federal learning framework and the cross-domain strategy. This method is very easy to understand and follow. And the experimental results show that this method is very effective. Based on practical problems, the method proposed by the author has great application prospects.
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.

This paper proposes a novel and interesting condition generalization method under a federal learning framework for CT reconstruction under different scanning conditions(low-mAs, sparse-view and limited-angle). The idea is new, and the experimental results are convincing. I support acceptance of this paper.

Author Feedback

We sincerely thank all the reviewers for the valuable and constructive feedback, and we will improve on the deficiencies in the subsequent version of the paper. Furthermore, we would like to respond to the reviewers’ comments. Response to Reviewer #1: 1.As noted, due to limitations of pages, we are lacking in the review of recently developed methods, which we will have a more detailed review in the camera-ready paper.

2.The details of the low-dose simulation are mentioned in our previous work[1], we simulate the low-dose CT projection data by adding Poisson and Gaussian noise to the log-transformed projection data according to the intensity of X-rays, and other simulation parameters are given in the supplementary material. 3.We will validate the performance of our method under clients with larger noise levels in a subsequent study.

Response to Reviewer #2: 1.Due to limitations of pages, we do not provide network parameters in the manuscript, but a figure of detailed network structure was provided in the supplementary material. 2.The lack of literature citations is a shortcoming of our manuscript, therefore we will conduct a fuller literature review in the camera-ready paper. 3.In our manuscript, the two conditions (Condition #1, Condition #2) refer specifically to the experimental conditions.

For Condition #1, the three clients had different imaging geometries, anatomical regions and dose levels, and we mainly focused on the effect of different X-ray intensities and anatomical regions on the reconstruction performance in this Condition.

For Condition #2, the three clients had different imaging geometries, anatomical regions and scanning protocols, and we mainly focused on the effect of different low-dose scanning protocols on the reconstruction performance.

After the validation of the experiments under the above two conditions, the proposed FedCG has strong generalizability even if the imaging geometries, anatomical regions, dose levels, and scanning protocols of each client are varied.

Response to Reviewer #4:

We will conduct ablation experiments focusing on the training strategy and number of clients in the subsequent study, and a more detailed review will be mentioned in the camera-ready paper.

[1]Zeng, D., Huang, J., Bian, Z., Niu, S., Zhang, H., Feng, Q., Liang, Z., Ma, J.: A Simple Low-dose X-ray CT Simulation from High-dose Scan. IEEE Transactions on Nuclear Science 62(5), 2226–2233 (2015)

back to top

Federated Condition Generalization on Low-dose CT Reconstruction via Cross-domain Learning