COVID-ResNet: COVID-19 Recognition Based on Improved Attention ResNet
Round 1
Reviewer 1 Report
In this paper, the authors proposed COVID-ResNet: COVID-19 Recognition Based on Improved Attention ResNet . Therefore, it is interesting and attractive. However, it should be major revised to enhance the quality, as follows:
1) In Section 1, the authors mentioned the introductory part of the ResNet framework, but still it requires some motivation, contributions and oraganisation subsection
2) At the final end of section 1, the authors should indicate the rest of this paper is organized how.
3) In lines 166-175, the equations should be derived more in details
4) A summary table of the mathematical notation should be provided for convenience for the readers.
5)Table 1 should be re-presented in details with valid references
6)Section 3.2 requires to rewrite with suitable examples
7)Conclusions and future scope sections are missing. Pl elaborate
Author Response
Please see the attachment about Reviewer 1's reply.
Author Response File: Author Response.pdf
Reviewer 2 Report
This paper provides a diagnosis methodology based on an improved attention Resnet model named as COVID-ResNet. COVID-ResNet is designed based on CT images to focus on the focal lesion area. Secondly, the SE-Res block is 20 constructed. The squeeze excitation mechanism with the residual connection is introduced into the ResNet. It 21 can enhance the correlation degree among different channels and improve the overall accuracy of the model. 22 Thirdly, MFCA (multi-layer feature converge attention) blocks are proposed, which extract multi-layer 23 features. In this model, it uses coordinated attention to focus on the direction information of the lesion area. 24 Different layer features are concatenated together so that the shallow and deep layer features are fused. The 25 experimental results show that the model can significantly improve the recognition accuracy of COVID-19. 26 Compared with similar models, this model has better performance. On the COVID-19 CT data set, the 27 accuracy, recall rate, F1 score, AUC value can reach 96.89%, 98.15%, 96.96%, 99.04%, separately. Compared 28 with ResNet model, the accuracy, recall rate, F1 score, AUC value are higher 3.1%, 1.53%, 2.52%, 0.79% 29 separately
-
Can the authors show the performance of the proposed methodology on more COVID and non-COVID CT images?
-
The authors compared the proposed methodology with other pre-trained model? Can they compare there the performance with other related works?
-
The confusion matrices’ figures are not clear and need to be improved in the quality?
-
What is the complexity of the proposed methodology?
-
What is the advantages and the disadvantages of the proposed methodology?
-
It is better to remove the equation measurements from the manuscript because they are already known or the authors can added them in the appendix section?
Author Response
Please see the attachment about Reviewer 2's reply.
Author Response File: Author Response.pdf
Reviewer 3 Report
After carefully going over the work, I found some important problems that need to be fixed before the paper's quality may be improved.
1. Rewrite abstract section and add finding properly to this section.
2. The literary quality of this entire study has to be raised.
3. To make the result section clearer, reference the ablation study. And be sure to state your reasoning for selecting a certain paradigm for this article.
4. A more thorough evaluation of model performance is necessary. The performance matrix should be expanded.
5. The system architecture has to be updated with an appropriate flowchart. Create a flowchart and give specific details for each area.
6. Cite a few recent publications, such as https://link.springer.com/article/10.1007/s11517-022-02543-x; https://www.mdpi.com/2306-5354/9/7/281; https://link.springer.com/article/10.1007/s00521-022-07424-w; and https://bmcmedinformdecismak.biomedcentral.com/articles/10.1186/s129
7. Add conclusion section in this paper properly.
Author Response
Please see the attachment about Reviewer 3's reply.
Author Response File: Author Response.pdf
Reviewer 4 Report
The paper is challenging to read, especially the introduction. Authors, jump from topic to topic, sentences are without the subject, suddenly are stopped, thoughts are interrupted, and there is no flow of the presentation.
Several abbreviations are without full meanings and many statements are without references to support them. The font in the confusion matrix is too small, the paper must be zoomed in to read. Also, there is some problem with references for other architectures used in the experiment.
Additionally, it is not clear what is the contribution of the proposed architecture, all network blocks are already used in some other publications. It seems that for comparison, the authors only selected networks for standard image classification, they didn't choose some architectures dedicated to the same problem they are trying to solve.
Overall, in its present form, with its existing presentation, unclear contributions, and current comparison method, the paper is not ready for publication and requires redesign and further improvements.
Author Response
Please see the attachment about Reviewer 4's reply.
Author Response File: Author Response.pdf
Reviewer 5 Report
The authors of the paper "COVID-ResNet: COVID-19 Recognition Based on Improved Attention ResNet" present an architecture based on SE-ResNet for the identification of COVID-19 in CT images.
The following are my comments on the paper:
1. The introduction needs to be improved. First, change some phrases, such as "ResNet[5] is a hot spot of deep learning at present", to more appropriate technical terms. Second, before the paragraphs mentioning ResNet for medical image processing, explain the importance of DL-based models in image classification and object recognition tasks. Third, include a paragraph that explicitly identifies the contributions of the article. Fourth, include a paragraph describing the content of the other sections of the article.
2. Include an introductory paragraph at the beginning of each section, for example, prior to 2.1, mention the summary of the contents of the material and methods section.
3. It is mandatory to correct some technical terms of the article. For example, it is not "convolute layer", but "convolutional layer" (Figure 1, figure 2 and others).
4. It is mandatory to complete some phrases of the article, such as “The input and output are defined as and respectively” (line 109).
5. The first time abbreviations are used, they should be explained. For example, the term SE-Res block is used many times in the document, but its meaning is presented only up to section 2.2.1.
6. The proposed architecture is not clear (Figure 3). It uses two types of blocks, the SE-REs block and the MFCA block, which are not explained in detail in the document (only with general diagrams, but not with concrete examples of block input-output images). Therefore, the reader cannot easily understand the purpose of each one in the architecture. The above, taking into account that both SE and MFCA are a channel attention mechanism, which use global average pooling.
7. Section 3.4 presents the results of nine external models along with the proposed model. However, the provenance of the external model values is not clear, since only the experiment conditions of the proposed model are mentioned in section 3.1.
8. The values presented in Table 4 do not coincide with those reported in the confusion matrices in Figure 12. For example, for SE-ResNet18 we have that acc=(303+309)/(303+17+16+309)=0.9488, but the corresponding value in Table 4 is 0.9519. This lack of consistency among the data detracts from the reliability of the document.
9. There are numerous errors such as "Reference source not found" in the document.
10. In general, the wording of the article and the use of technical terms must be improved.
Author Response
Please see the attachment about Reviewer 5's reply.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Authors must clarify several critical points, such as
1. Authors should provide bullet points to clarify their contribution to their work.
2. In this study, whether Loss function, Optimiser, or learning what is used? Correctly add to this paper.
3. The Reference section is in disarray. For example, in reference No. 3, is ORCID an author? What exactly is this? Check all properly.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
It seems that the authors adapted the paper according to the reviewers' comments. But still, it is challenging to follow the text. Sentences are sometimes suddenly interrupted. The manuscript should go through extensive editing of the English language and style.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 5 Report
The authors of the paper entitled "COVID-ResNet: COVID-19 Recognition Based on Improved Attention ResNet" have improved the document according to the reviewer's comments. It is only suggested to improve the wording and/or grammar in some sections of the document, as well as to unify the names of the models, for example DenseNet and not densenet.
The paper is now ready for publication.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 3 Report
Hello,
Thank you for completing all comments properly. This paper can be accepted now. Add some recent paper;
1. https://www.mdpi.com/2079-3197/10/10/177
2.https://www.sciencedirect.com/science/article/pii/S1876034121003774
3. https://www.mdpi.com/2227-7390/10/4/564
4. https://link.springer.com/article/10.1007/s11334-022-00523-w
Thank you
Author Response
Please see the attachment.
Author Response File: Author Response.pdf