Acute Leukemia Diagnosis Based on Images of Lymphocytes and Monocytes Using Type-II Fuzzy Deep Network
Round 1
Reviewer 1 Report
This paper develops type II fuzzy deep neural networks to extract hierarchical features and improve the diagnostic accuracy of acute leukemia. The proposed method has a satisfactory generalization ability.
1. In the abstract section, the authors should present some numerical values to highlight the main findings of the study.
2. In the introduction, it is better to clearly state the main contributions of this paper.
3. Errors in Eq. (1) and Eq. (3) should be carefully checked.
4. The variables are missing in Line 227.
5. It is should be checked that all the notations have been fully defined and explained. For example, u and u(L-1) in Line 226.
6. “However, because the proposed model performs better in 340 practice with the SGD algorithm”, which contradicts the optimizer selected in Table 2.
7. For this classification problem, briefly describe the reasons for using the MSE instead of the cross-entropy loss function.
8. The ReLU not only suppresses the problem of gradient vanishing, but more importantly, it improves the training efficiency of the models, so the authors should add an experimental comparison of training time appropriately.
9. More comparison and discussion with related methods, such as Explainable multi-instance and multi-task learning for COVID-19 diagnosis and lesion segmentation in CT images, Knowledge-Based Systems.
Author Response
Original Manuscript ID: electronics-2046749
Original Article Title: “Acute Leukemia Diagnosis Based on Images of Lymphocytes and Monocytes Using Type-II Fuzzy Deep Network”
To: Guest Editors Dr. Sebelan Danishvar and Dr. Morad Danishvar
Re: Response to reviewers
Dear Respected Editors,
Please find below the response to the respective reviewers’ comments. We considered all of the comments in detail and did our best to modify the paper in the way they suggested. We believe that the comments have considerably increased the quality of the manuscript. We would be most grateful if you consider the revised manuscript entitled “Acute Leukemia Diagnosis Based on Images of Lymphocytes and Monocytes Using Type-II Fuzzy Deep Network” for possible publication in the Journal of Electronics. We have read and have abided by the statement of ethical standards for manuscripts submitted to the Journal of Electronics. We are uploading (a) our point-by-point response to the comments (below) (response to reviewers), (b) an updated manuscript with yellow highlighting indicating changes, and (c) a clean updated manuscript without highlights (PDF main document).
Also, Dr.. Sablan Daneshvar should be added as the responsible author.
Best regards,
Ahmad Habibizad Navin
Professor
Department of Computer Engineering
Tabriz Branch, Islamic Azad University, Tabriz, Iran
Office: ECE 215
Phone: +98 41 33393748
Fax: +98 41 33300819
E-mail: a.habibzad@srbiau.ac.ir
Reviewer#1:
Comments:
This paper develops type II fuzzy deep neural networks to extract hierarchical features and improve the diagnostic accuracy of acute leukemia. The proposed method has a satisfactory generalization ability.
- ⎫ Thanks to the esteemed reviewer, we believe that your comments have been very useful and effective in enhancing the scientific and writing framework of the manuscript. We have considered all the comments in their entirety and made every effort to correct the manuscript in the manner suggested by the honorable reviewer.
- 1. In the abstract section, the authors should present some numerical values to highlight the main findings of the study.
- ⎫ The manuscript is revised based on this comment. Yes, the opinion of the honorable reviewer is absolutely correct. Based on this, the quantitative results of the proposed model have been added to the abstract section which is highlighted in abstract section, page 1 and line 23.
- 2. In the introduction, it is better to clearly state the main contributions of this paper.
- ⎫ The manuscript is revised based on this comment. Based on the opinion of the respected reviewer, the contribution of the article was clearly added to the introduction section.
”The contribution of this research can be summarized as follows:
- a. Compilation of a standardized database of acute leukemia images.
- b. Presenting an automatic method based on the combination of type-II fuzzy networks with deep learning networks in order to solve the problem of uncertainties in model training.
- c. Providing an automatic end-to-end model with high speed and accuracy without the need for feature extraction/selection block.”
Which is highlighted in page 5 and lines 207-213.
- 3. Errors in Eq. (1) and Eq. (3) should be carefully checked.
- ⎫ The manuscript is revised based on this comment. According to the opinion of the second reviewer and consensus with the first reviewer, sections 2.1 and 2.2 have been summarized for better understanding of the readers and the relevant formulas have been removed which is highlighted in pages 5 and 6, and lines 222-260.
- 4. The variables are missing in Line 227.
- ⎫ The manuscript is revised based on this comment. According to the opinion of the second reviewer and consensus with the first reviewer, sections 2.1 and 2.2 have been summarized for better understanding of the readers and the relevant formulas have been removed which is highlighted in pages 5 and 6, and lines 222-260.
- 5. It is should be checked that all the notations have been fully defined and explained. For example, u and u(L-1) in Line 226.
- ⎫ The manuscript is revised based on this comment. According to the opinion of the second reviewer and consensus with the first reviewer, sections 2.1 and 2.2 have been summarized for better understanding of the readers and the relevant formulas have been removed which is highlighted in pages 5 and 6, and lines 222-260.
- 6. However, because the proposed model performs better in 340 practice with the SGD algorithm”, which contradicts the optimizer selected in Table 2.
- ⎫ The manuscript is revised based on this comment. Thanks to the careful opinion of the respected reviewer, the RMSProp optimizer has been used in the proposed network according to Table 2. In the network architecture description section, there was a typo and SGD optimizer was written instead of RMSProp, which was corrected.
- 7. For this classification problem, briefly describe the reasons for using the MSE instead of the cross-entropy loss function.
- ⎫ The manuscript is revised based on this comment. The use of cross-entropy is more useful in classification problems than MSE. Also, MSE is more useful in regression-related problems. However, MSE works faster than cross-entropy in classification problems. Considering that the accuracy of classification is above 95%, a compromise between speed and accuracy of classification should be established. Accordingly, in this study, MSE has been used as the loss function which is highlighted in pages 7 and 8, and lines 322-326.
- 8. The ReLU not only suppresses the problem of gradient vanishing, but more importantly, it improves the training efficiency of the models, so the authors should add an experimental comparison of training time appropriately.
- ⎫ According to the opinion of the respected reviewer, the performance of different functions in terms of computational efficiency has been presented.
“In order to clarify the computational efficiency of Type 2 Fuzzy, Leaky Relu and Relu functions, the performance of each is shown in Table 3. According to Table 3, as it is known, the computational efficiency of type II fuzzy function is lower compared to Relu and Leaky Relu, and it can be used in real-time applications.”
Table 3. Comparison of network training time with different functions.
Function Used |
Relu |
Leaky-Relu |
Type II Fuzzy |
Training Time |
5400 s |
5451 s |
5302 s |
Which is highlighted in page 13, and lines 410-413.
- 9. More comparison and discussion with related methods, such as Explainable multi-instance and multi-task learning for COVID-19 diagnosis and lesion segmentation in CT images, Knowledge-Based Systems.
- ⎫ According to the reviewer's opinion, the study presented in Moore's article was discussed and cited as reference 35 which is highlighted in page 13, and lines 428-436.
Author Response File: Author Response.pdf
Reviewer 2 Report
section 2.1, brief description of DCNN. Not sure if it is necessary, depending on the depth of knowledge of readers about deep learning.
section 2.2, the author can put more emphasis on introducing the difference between type-ii fuzzy network vs regular DCNN. Instead, the math, if not original by the author, can be put into appendix to help with understanding for the readers.
section 3.2, I assumed that images are taken on different microscopes with varying illumination. does image color play a role in affecting network outcomes? Before converting to grayscale, anything has been done to standardize colors to minimize difference?
ln 342, why use MSE instead of cross-entropy, etc. as loss function since it is a classification problem.
in Fig 5. the model almost achieved a perfect score. What are the 2 false cases look like? what are the reasons that caused them to fail?
Author Response
Reviewer#2:
Comments:
- 1. Section 2.1, brief description of DCNN. Not sure if it is necessary, depending on the depth of knowledge of readers about deep learning.
- ⎫ Thanks to the esteemed reviewer, we believe that your comments have been very useful and effective in enhancing the scientific and writing framework of the manuscript. We have considered all the comments in their entirety and made every effort to correct the manuscript in the manner suggested by the honorable reviewer.
- ⎫ The manuscript is revised based on this comment. According to the opinion of the respected reviewer, According to the opinion of the second reviewer and consensus with the first reviewer, sections 2.1 and 2.2 have been summarized for better understanding of the readers and the relevant formulas have been removed. The manuscript is revised based on this comment. Yes, the opinion of the honorable reviewer is absolutely correct. Based on this, the quantitative results of the proposed model have been added to the abstract section which is highlighted in abstract section, page 1 and line 23.
- 2. Section 2.2, the author can put more emphasis on introducing the difference between type-ii fuzzy network vs regular DCNN. Instead, the math, if not original by the author, can be put into appendix to help with understanding for the readers.
- ⎫ The manuscript is revised based on this comment. According to the opinion of the respected reviewer, According to the opinion of the second reviewer and consensus with the first reviewer, sections 2.1 and 2.2 have been summarized for better understanding of the readers and the relevant formulas have been removed. The manuscript is revised based on this comment. Yes, the opinion of the honorable reviewer is absolutely correct. Based on this, the quantitative results of the proposed model have been added to the abstract section which is highlighted in abstract section, page 1 and line 23.
- 3. Section 3.2, I assumed that images are taken on different microscopes with varying illumination. does image color play a role in affecting network outcomes? Before converting to grayscale, anything has been done to standardize colors to minimize difference?
- ⎫ With respect to the opinion of the respected reviewer, all the images are taken from one microscope based on 10 eyes, which is also shown in the figure below. Accordingly, there was no need to unify the color of the images. However, the honorable reviewer's opinion is absolutely correct. If different databases are used, it will be necessary to standardize the images.
- 4. ln 342, why use MSE instead of cross-entropy, etc. as loss function since it is a classification problem.
- ⎫ The manuscript is revised based on this comment. The use of cross-entropy is more useful in classification problems than MSE. Also, MSE is more useful in regression-related problems. However, MSE works faster than cross-entropy in classification problems. Considering that the accuracy of classification is above 95%, a compromise between speed and accuracy of classification should be established. Accordingly, in this study, MSE has been used as the loss function which is highlighted in pages 7 and 8, and lines 322-326.
- 5. In Fig 5. the model almost achieved a perfect score. What are the 2 false cases look like? what are the reasons that caused them to fail?
- ⎫ The manuscript is revised based on this comment. 2 other models, including ReLU and Leaky-ReLU activation functions in deep network, do not perform optimally in nonlinear mapping of input to output. Also, they cannot perform well in solving uncertainties and need more training time in order to reach the desired convergence of the network. The performance of different functions in terms of computational efficiency has been presented.
“In order to clarify the computational efficiency of Type 2 Fuzzy, Leaky ReLU and ReLU functions, the performance of each is shown in Table 3. According to Table 3, as it is known, the computational efficiency of type II fuzzy function is lower compared to ReLU and Leaky ReLU, and it can be used in real-time applications.”
Table 3. Comparison of network training time with different functions.
Function Used |
ReLU |
Leaky-ReLU |
Type II Fuzzy |
Training Time |
5400 s |
5451 s |
5302 s |
Which is highlighted in page 13, and lines 410-413.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The revised paper is improved.