Class Imbalanced Medical Image Classification Based on Semi-Supervised Federated Learning

Round 1

Reviewer 1 Report

Here are a few comments from my side:

• Many statements and explanations are without any citations. Procedures followed or improvised, Comparison with state-of-the-art method or the iterations etc., mostly do not quote references from existing / previous literature. (Also experiment and the progress in the process were also felt as if they are only based on process followed alone and have no citations.)
• Equations may be numbered, and citations may be provided at appropriate places.(110-135)
• Sentences like “We perform….We found ….We fixed…..We can …..” (which is found in more than 50 places) & “our method……Our proposed….” Etc., can be rewritten in a better way.
• A thorough grammar revision may be carried out.
• Discussion / Conclusion may be improvised. It does not give out a concrete meaning. The contribution of the study may be more explicitly sated
• Paper needs thorough revision

Author Response

Dear reviewers, thank you very much for pointing out the problems, we have replied to your comments item by item ''Please see the attachment'', and have reworked the article, please review it.

Author Response File: Author Response.docx

Reviewer 2 Report

- Please improve figure 1 and increase font size. It is hard to read.

- Half of the introduction is about the proposed method. Please focus on introducing the problem, related works, define the gap in the literature and briefly summarize your contribution. The rest of the methodology should be in materials and methods section.

- The paper can benefit from discussion of related works in the introduction on semi-supervised and weakly supervised learning, with prior constrains, as well as the impact of using several deep learning architectures in medical imaging such as:

1- Amyar, A., Modzelewski, R., Vera, P., Morard, V. and Ruan, S., 2022. Weakly Supervised Tumor Detection in PET Using Class Response for Treatment Outcome Prediction. Journal of Imaging, 8(5), p.130.

2- Amyar A, Modzelewski R, Vera P, Morard V, Ruan S. Multi-task multi-scale learning for outcome prediction in 3D PET images. Comput Biol Med. 2022 Dec; 151(Pt A):106208.

3- Amyar, A., Guo, R., Cai, X., Assana, S., Chow, K., Rodriguez, J., Yankama, T., Cirillo, J., Pierce, P., Goddu, B. and Ngo, L., 2022. Impact of deep learning architectures on accelerated cardiac T1 mapping using MyoMapNet. NMR in Biomedicine, 35(11), p.e4794.

- Method section should be expanded.

- Part of the method is in the result section. Please move everything related to method to the previous section, and focus only on reporting the results. There is also a part that should go to discussion. For example from line 318 to 326.

- Discussion should be improved.

-Please add a conclusion.

Author Response

Dear reviewers, thank you very much for pointing out the problems, we have replied to your comments item by item ''Please see the attachment'', and have reworked the article, please review it.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper provides a federation learning method using a combination of pseudo label construction and regularization constraints, where the federation learning framework consists of a central server and local clients containing only unlabeled data, and labeled data are passed from the central server to each local client to participate in semi-supervised training. Combining these two methods with the aim of selecting fewer classes with a higher probability to make an effective solution to the class imbalance problem and improve the sensitivity of the network to the unlabeled data. We perform experimental validation on a public medical image classification dataset consisting of 10,015 images, and our method improves AUC by 7.35% and average class sensitivity by 1.34%. The work is interesting, but there exist some minor issues that must be defined.

1. What is the kind of medical images that are used in this study for experimenting and validation?

2. What is the value of the AUC without using any federated learning?

3. Can the authors show the training and the validation curves of the DenseNet121 model on the federated enviroment?

4. What is the complexity of the proposed methodology?

5. What is the advantages and the disadvantages of the proposed methodolgy?

6. Can the authors compare with other related works on the same public medical image dataset used?

7. Is the same performance achieved using the fedrated and non fedrated learning?

Author Response

Dear reviewers, thank you very much for pointing out the problems, we have replied to your comments item by item ''Please see the attachment'', and have reworked the article, please review it.

Author Response File: Author Response.docx

Reviewer 4 Report

Limitations of proposed method are not clearly emphasize, only it is mentioned limitation for multi-category multi-classification tasks.

Better describe clinical benefits.

It looks that comparison with other methods gives significant improvement under the small batch, and the sensitivity (recall rate) of important medical indicators under the optimal batch. Other comparison results did not get benefits of proposed method. More comments are necessary for these issues.

 

 

 

Author Response

Dear reviewers, thank you very much for pointing out the problems, we have replied to your comments item by item ''Please see the attachment'', and have reworked the article, please review it.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

NA

Reviewer 3 Report

The manuscript in the current form is suitable for publication.