COVID-19 CXR Classification: Applying Domain Extension Transfer Learning and Deep Learning
Round 1
Reviewer 1 Report
This paper employs deep transfer learning utilizing the ResNet-50, trained on source and target domain data pertaining to MIMIC CXR and Mendeley (COVID-19 CXR images), respectively, to classify unseen target data pertaining to COVID-19 CXR images. Experimental results demonstrate the good performance results when compared to others. The paper is well-written ans is well-presented. However, the following comments need to be addressed:
-Need to report results in Table 8 on the whole dataset pertaining to the target data using 5-fold cross-validation. For fairness of the performance comparison against the baseline, need to assess the performance on the same test folds provided to all models. Moreover, include additional performance measures such as balanced accuracy and Matthews correlation coefficient Also, report the standard deviation
-In "2.3. Domain Extension Transfer Learning" section, need to define the problem formulation for transfer learning (i.e., source data, target data ...etc).
-Table 10 shows the statistical difference between the proposed model and the baseline. What about the statistical difference of others in Table 8?
Author Response
We appreciate your helpful and constructive comments and suggestions to improve the paper.
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The manuscript proposes an effective deep learning methodology for quarantine response through COVID-19 chest X-ray image classification based on domain extension transfer learning. This paper has been written well and is easy to read. However, the authors should be considered and addressed the following concerns.
(1) Consistency concept is very important to present your work in a high reputation journal. Thus, the authors are recommended to follow and double-check sentences in the paper. Some of these examples are as follows. Upper letter or smaller for the first letter in an abbreviation such as > Coronavirus Infectious Disease-19 (COVID-19)>> World Health Organization (WHO), computed tomography (CT)
(2) Please cite the references after mentioning the authors’ names such as Line 59>> Ozturk et al. proposed Dark-CovidNet to identify COVID-19 59 lesions in CXR[18] >>>Replace to >> Ozturk et al. [18] proposed Dark-CovidNet to identify COVID-19 59 lesions in CXR. And Line 60 >> Mahmud et al. [19] proposed…; etc..
(3) What is 341 in line 76 “CNN model using 341 COVID-19 X”?
(4) Please use Section term instead of Chapter in the outline of the paper part in the introduction.
(5) Authors are recommended to create a new section called, related work or literature review, after the introduction section. In this section, some studies have been mentioned in the introduction needed to move here.
(6) In the conclusion section, authors are recommended to add limitations and future work.
Author Response
We appreciate your helpful and constructive comments and suggestions to improve the paper.
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Authors have presented a deep learning model for Covid19 X-ray classification. The manuscript is well written however the following are my concerns to enhance the manuscript.
L30 - Covid19 cases are not increasing every day now. This statement should be in past tense or rephrased it.
The introduction section is a bit chaotic. No real analysis or research gap is discussed. Authors are expected to properly present the background, role of deep learning in covid19, and the research gap the manuscript addresses.
Section 2 is started as "In this chapter" it should be changed to "Section".
Discussion on what is transfer learning, why transfer learning is, what is ResNet model and its basic description are missing.
How data imbalance is treated in this work?
The need for the precision, sensitivity, F1-Score and so on are not justified.
What is the baseline model to which the results are compared with? Reference missing.
I think figure 7 and table 8 shows the same values. Anyone can be retained if both are using same values
What was the threshold set for early stopping. Report it.
In all the non-covid heat maps, the model is seems to be learning from irrelevant parts of the Xrays. Justify the results.
Discussion should be elaborated in detail. As of now it is very short. Discuss the results, highlight the advantages, limitations and relevance to the current state-of-the-art research.
Conclusion is not supported results also the future work is missing.
Author Response
We appreciate your helpful and constructive comments and suggestions to improve the paper.
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Authors addressed comments raised in a previous round of review .
Minor comment:
In Line 250, (Y|X) should be P(Y|X)?.
The reason is that P(Y|X) is as f(x) (i.e., model learning from machine learning). However, P(Y|X) is in terms of turning the probability option. Also, for consistency with the notation (i.e., P(Y_t|X_t)) in Line 252.
Author Response
We appreciate your helpful comments.
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Authors have addressed my concerns. I appreciate the author's effort.
Author Response
We appreciate your careful review of the manuscript.
