COVID-19 Classification through Deep Learning Models with Three-Channel Grayscale CT Images

Round 1

Reviewer 1 Report

Review on BDCC-2114169

COVID-19 Classification through Deep Learning Models with a Three-channels Grayscale CT Images

 

Does the introduction provide sufficient background and include all relevant references?   - yes

 

In this work, the authors proposed a pre-processing methodology to convert the covid-19 input grayscale images into three channel grayscale datasets with two different image enhancement methods, namely  Histogram Equalization (HE) and Contrast Limited Adaptive Histogram Equalization (CLAHE). 

 

The proposed pre-processing methodology enables the feeding of input dataset to pre-trained Deep Neural Network architecture for the classification of covid19 affected and not affected cases

 

Are all the cited references relevant to the research?  yes 

                                   

Is the research design appropriate?  yes, the proposed design justifies the proposed aim

                                   

Are the methods adequately described? yes, but the motivation behind the chosen methods i.e. histogram equalization and Contrast Limited Adaptive Histogram Equalization needs to be included may be with motivation example

 

                                   

Are the results clearly presented?  yes

Figure 8 , Figure 9, Figure 10, Figure 11 may be removed as the values are given in Table 3, Table 4,  Table 5 and Table 6

 

                                   

Are the conclusions supported by the results?  yes

 

Detailed comments

 

Comment - 1

The authors are suggested to include why they have chosen HE and CLAHE methods for creating two more channels

(whether the preprocessing helps in assessment of images with respect to severity of the infection)

 

Comment - 2

Since the authors intercompared their results, with other similar work (as in Table 9), they can write about the experimental set up of the proposed work

 

Comment - 3

Figure 8 , Figure 9, Figure 10, Figure 11 may be removed as the values are given in Table 3, Table 4,  Table 5 and Table 6

 

Comment 4

The authors can present the tuned hyper parameters obtained with validation data set. 

 

Comment 5

The authors are encouraged to highlight the number of parameters used in similar related works

in Table 9.  Also, the information about hyper parameters may also be included, if available

 

Comment - 6

Line No - I 73,  the number should be 2481(from Table 1)

 

Comment - 7

type errors like the following needs to be checked

line number in line 25

ceptionV3(GRAY+E+CLAHE)

needs to be replaced with ceptionV3(GRAY+HE+CLAHE)

 

Comment – 8

 

The authors may think about image processing methods which help in assessment of severity of disease as it is of more practical importance, in addition to the detection of infection. If relevant, they add as future extension

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Summary:

The authors present a novel pre-processing method that improves COVID-19 detection accuracy. The authors compare the results using various neural networks. This study may be relevant to the research domain in providing a diverse comparative analysis that includes CNN-based architecture and vision-transformer models. It also attempts to provide insights into how image representations can be utilized to improve the recognition ability of the classification models. This manuscript falls within the scope of the journal. However, the manuscript still has many gaps in the results, literature review, and the novelty presented. The manuscript also lacks clarity in several places. Owing to these significant limitations, the reviewer cannot recommend accepting the manuscript at this time. Furthermore, extensive grammatical and syntactic changes are required to this manuscript. Please carefully modify and submit again after a thorough mutual check between coauthors.

comments to authors:

1) Line 13: The reviewer agrees to the fact that COVID -19 is a highly infectious disease. However, is CT the main source for detecting the virus that causes the disease? The CT contains harmful radiation, and it is unclear how often the technique is used for COVID - 19 screening.

2) In the abstract, the authors fail to show the sufficient background, significance, novelty, and impact of the proposed methods Shown in the study.

3) lines 43 - 47: CT may be the most effective way. However, the authors must include another few lines regarding the side - effects of CT on the human body. This may be a reason why the modality is not used for this particular case, and a sufficient dataset is not available.

4) lines 61 - 83: Sufficient literature survey is not conducted. The authors cite various articles showing the usage of various DNNs, but their contribution to the field of research and the results are not discussed.

5) In general, the entire section 1 (Introduction) is very poorly written and must be rewritten in such a way that the section introduces the readers to the problem at hand and why the authors designed the novel study shown in this article. Moreover, the paragraphs seem to be cut-pasted and lack any coherence.

6 ) Section 2 is redundant as it does not provide any significant insight. It could be added as part of section 1.

7) In line 180: what are all the parameters?

8) The reviewer fails to see any novelty in the methods and the results section. The data processing seems to be very complicated, and the proposed method fails to improve on the significant methods that are already in existence.

9) The reviewer would like to know if any noise was added to the images in the dataset. Usually, the generation of datasets using imaging modalities is added with noise to account for the estimation errors, movement artifacts, and blurring. The well-posedness and robustness of the DNNs cannot be checked for the algorithm without adding prior noise. Please verify the results.

10) In sections 4.1 and 4.2, where the authors discuss the results, The results are merely "shown" in a written format and are not "discussed." The authors should provide sufficient insight regarding their findings, especially how and why certain networks perform better than others.

11) Moreover, No training graphs have been included in the article. The training graphs form an integral part of the analysis, and the convergence of the loss function with increasing epochs should be shown for the training and testing.

12) Finally, there are far too many grammatical and syntactic errors in the article. The reviewer suggests that extensive English corrections be made to the article and that the article is rewritten. There is also not enough technical content discussed in the article. The contribution of the authors in enhancing this imaging method is very minimal and is explained poorly. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

 

Paper summary:

 

This paper introduces different pre-processing methods to convert one-channel grayscale CT to the three-channel grayscale image to match the input of widely used pre-trained models, so the pre-trained models can be used to improve the performance of COVID-19 classification after transfer learning.

 

 

Comments for author:

 

1. This paper is generally well-written. The logic is easy to follow.

 

2. The motivation is clear. Considering the limited (labeled) CT dataset available on COVID-19, it's necessary to develop more efficient data preprocessing solutions to apply pre-training models on ImageNet to medical images.

 

3. The main concern is the data-splitting strategy. Line 188, The selected dataset was randomly divided into training, validation, and testing sets with an approximate ratio of 70%, 20%, and 10%, respectively. But in fact, different CT images from the same patient are highly correlated. It should split the dataset by patient to make sure the train/validation/test split with no patient overlapping. Otherwise, the evaluation results tend to become much better and become worse when the trained model is used in the CT images of unseen patients. 

 

 

4. Another question is why ViT models not work well as CNN-based models. Is it easy overfitting? Or the data pre-processing method can be further improved for ViT models?

 

 

5. Once the three-channel grayscale images are obtained, is it still possible to apply the data augmentation solution for transfer learning?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

All the reviewer comments are answered very well. The reviewer has no further concerns. The manuscript is acceptable for publication.

Reviewer 3 Report

The paper is suitable for publication in its current form.