The Detection of COVID-19 in Chest X-rays Using Ensemble CNN Techniques

Round 1

Reviewer 1 Report

In their article on classification of chest-Xrays using CNN the authors present a well structured and well written article on an important topic. The also provide sufficient details to follow the approach they used which is particularl backed up by using a large image data set used for training and analysis.

Nevertheless, I think there are a couple of minor and a few major issues, which should be addressed.

First, there are a couple of abbreviations used in the manuscript, such as GLCM, which are used since the early beginning however only clarified later in the document. This makes it unecessary difficult in some instances.

Line 71: The Abbreviation glcm has not been introduced accordingly.

 

Line 143: The Abbreviation GLCM has still not properly introduced.

 

Line 197-205: Its only then that the abbreviation GLCM is explained for the used implementation. However, the text alone is comparably hard to understand on its own. Adapting the graphic comparably the GLCM diagram from mathworks might be helpful.

 

Figure 11a and b: The y axis stretches beyond 1 to1 105 and respectively 1.1.However Accuracy should not extend beyond 1?

 

 One major point. In the early phase of the manuscript the authors highlight that it is especially important in early triage to classify for COVID and non-COVID pneumonia. However, while the authors generally provide a good overall assessment on generalization, this question is not considered further in the analysis discussion. Given that the authors used a very large dataset this is in important additional consideration which is currently missing. Considering that this a hugely important question for generalization it would be important to know if accuracy and F-Scores increase with an onset of the disease towards more severe cases and over time. I think this issue should be considered in more detail particularly because some of the main differences such as textures and fluid content will progress with more severe cases in both classes, whereby it might also correlate with later imaging?

 

This also brings me to my last point. Apparently the images from the COVID-QU-Ex dataset appear to originate from different sources. Only three of the original papers on the web side seems to have been considered. 

What in my eyes needs a serious explanation by the authors are the ethical permission for the image acqusition and how consents of patients for the study was obtained.

 

 

 

 

 

Author Response

Thank you very much for your review of our paper. Please see the attached file for our detailed responses.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Authors,

I read your manuscript with pleasure and find it valuable. It addresses the difficult though important matter of interpretation of medical images. In my opinion, the manuscript is good enough to be considered for publication. However, I found some minor problems I would like you to correct/explain before the paper is published. I enclose the list of them below.

1. In lines 120 and 121, you state that some images were excluded due to too low or too high contrast. Though it is nothing wrong in eliminating some of the corrupted or not suitable ones, the criteria for that should be clearly and precisely presented. Unfortunately, I did not find any details concerning that process in the manuscript. It would be valuable for the reader if you give some more explanation of how you measured the contrast and what was considered as the threshold for inclusion/exclusion of the image.

2. In lines 185 and 186 you describe the idea of transfer learning. According to my knowledge, the most important than limiting the time of computation is the fact, that you have a complete CNN backbone allowing for getting the useful features extracted from the image. And indeed you make the use of it. Please extend the description to include this advantage of transfer learning.

3. In section 2.6.4 you describe the ANN used for the classification. How you have chosen the described architecture, I mean particularly the numbers of neurons in hidden layers and the number of hidden layers. Just one sentence with an explanation of it would be fine.

4. The tables 6 and 7 are somewhat unreadable. Why do you emphasize the first row with bold? I would rather expect an emphasis on the best value for each of the measures used. Please correct it: either do not use bold or mark the best value in each row (for each measure).

I found the language of the manuscript at a very satisfactory level, in my opinion, no corrections are necessary.

Author Response

Thank you very much for your review of our paper. Please see the attached file for our detailed responses.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors compared CNN network analysis to detect COVID 19 pneumonia. The study had been done and repeated several times. However, the author provided new data analysis method and discuss their difference. I really appreciate the work. However, I've several questions

1. How could the author detect pneumonia by other pathogen rather than covid-19, for example, in illustrated figure, it is very difficult to tell even if pneumonia is present in Covid 19 and pneumonia patients

2. The goal of the article is to detect pneumonia by caused by covid 19 or detect an abnormal CXR and test if the patient got covid 19?

 

The English is well written 

Author Response

Thank you very much for your review of our paper. Please see the attached file for our detailed responses.

Author Response File: Author Response.pdf