GCT-UNET: U-Net Image Segmentation Model for a Small Sample of Adherent Bone Marrow Cells Based on a Gated Channel Transform Module
Round 1
Reviewer 1 Report
The suggestions are as follows:
1. Overall English grammar and writing, and editing need to be checked. There are multiple faults/ mistakes identified.
2. The overall organization and writing need to be improved.
3. Equations (11, and 12) are not given.
4. The comparison between fig. 6a, 6b, and 7 need to be further elaborated.
Author Response
- The grammar and writing errors in the manuscript have been corrected, and the formatting errors in the manuscript have been corrected.
- Corrected the deficiencies in writing.
- Added equations 11 and 12.
- The comparison between Figure 6 and Figure 7 is further described.
Reviewer 2 Report
The document does not contain line numbers to indicate what the comments refer to.
· Page 1, ofadherent -> of adherent
· Page 1, which leads to the difficulty for cell (full stop at the end of a sentence)
· Page 1, When using the abbreviation GTC for the first time, write the full name of the term
· Pages 1-3, the text is not correctly formatted, it should be justified
· equations 1,2,3,4,5,6,7,8,9,10,11,12 are not written correctly
· Page 2, segmentationmethods[4] -> segmentation methods[4]
· Page 2, segmentation approach [6] (full stop at the end of a sentence)
· Page 2, convolutional layers [11].The main (add space after full stop)
· Page 3, The flow method we designed is shown in Figure 1.(rephrase the sentence)
· Page 4, Figure 1 (correct text in box) Data preprocessing
· Page 5, number of parameters.For (add space after full stop)
· Page 5, influenced by other regions [21]The Rectified (full stop at the end of a sentence)
· Page 7, Figure 3. (add axis names on the chart)
· Page 7, Figure 3 is not clear enough
· Page 7, the structure of GCT -> The structure of GCT
· Page 7, please provide more detailly description of Figure. 4.
· Page 7, channel. the GCT module -> channel. The GCT module
· Page 8, add space after variables ϒ ε
· Page 8, Gating deviation -> gating deviation
· Page 8, Then we have -> then we have
· Page 11, features by control-ling the
· Page 11, mechanism,since GC
· Page 12, From Figure 5we can
· Page 13, cellswhile for the detailed
· Page 13-14, Figure 6(a), Figure 6(a), Figure 7, expected and predicted values can be displayed on the same image in different colors
Author Response
Modify Description:
- Page 1. ‘ofadherent’is modified as ‘of adherent’.
- Page 1. Added full name with GCT for the first time.
- Page 1-3. Modified format errors and all formula errors.
- Page 2. ‘segmentationmethods’ is modified as ‘segmentation methods’.
- Page 2. Space added.
- Page 3. Rewrite ‘The flow method we designed is shown in Figure 1.’ into ‘The overall process we designed is shown in Figure 1.’
- Page 4. Modified the irregular box in Figure 1.
- Page 5. Space added.
- Page 5. Period added.
- Page 7. Added axis name for Figure 3.
- Page 7. Modified initial lowercase.
- Page 7. Modified initial lowercase.
- Page 7. Figure 4 is described in detail in the following sections.
- Page 8. Modified initial capitalization.
- Page 8. Modified initial capitalization.
- Page 11. ‘features by control-ling the’ is modified as ‘features by controlling the’.
- Page 12. Space added.
- Page 13. Space added.
- Page 13-14. In Figure 6 and Figure 7, the corresponding color cannot be modified because they are grayscale images.
Reviewer 3 Report
1. Most of the equations in the manuscript should be corrected – can not be read.
2. According to Table 1, the proposed method slightly outperforms other variants of UNET. Please provide more experiments on more datasets.
3. The contribution of this work should be better described and explain why these contributions are significant.
4. Provide some examples where the proposed method fails.
5. You may select an application to apply the proposed method (see the rest methods for example). In addition, you can include them in your bibliophagy.
[1] Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2018). Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4), 834-848.
[2] Panagiotakis, C., & Argyros, A. (2020). Region-based Fitting of Overlapping Ellipses and its application to cells segmentation. Image and Vision Computing, 93, 103810.
Author Response
Modify Description:
1.Corrected the wrong format of the equation in the manuscript.
2.Finally, the contribution of the work is further described in the manuscript.
3. Because there are no more data sets of adherent cells, experiments cannot be conducted on more data sets. If there are more data sets in the future, further supplements can be made.
4.Under the unsupervised method, our original data is too noisy to be segmented, and our dataset has a small amount of data, so it behaves generally when the model is complex.
5.We have included new papers in our bibliophagy.
Round 2
Reviewer 2 Report
Authors should thoroughly proofread the paper and correct obvious typographical errors before submitting for review.
There are too many typos in this paper, some of the typos that were highlighted for correction after the first review are still present in this version!
page 1, ofadherent -> of adherent
page 2, theaccuracy -> the accuracy
page 2, layers [11].The -> layers [11]. The
page 2, Many network structures with good segmentation performance have been born in the field of image
segmentation. - formulate the sentence differently
page 2, semantic segmentation, It includes -> semantic segmentation. It includes
page 3, GCT(Gated Channe -> GCT (Gated Channe
page 3, UNETbased -> UNET based
page 3, in Figure 1.In -> in Figure 1. In
page 3, expectations.Fi- -> expectations. Fi-
page 4, rameters.For -> rameters. For
page 4, image,it -> image, it
page 5, After completing the coding part in the network, It is not clear what coding part means?
page 5, ifx>0 -> if x > 0
page 5, ifx<0 -> if x < 0
page 5, remove folowing text:
ReLU has the following advantages. ReLU will not have the problem of gradient
disappearance. Because the gradient in the nonnegative interval is constant, it can main-
tain the final convergence rate of the network model in a stable state.
page 5, can betrained -> can be trained
page 8, Table 1 align the results in the last row
page 8, samples,and - samples, and
page 8, From Figure 3we -> From Figure 3 we
page 8, figure 3 and following text are not relevant for this paper
page 9, samples,especially
page 9, methods,it -> methods, it
page 9, separatedby -> separated by
page 10, effect.Figure 5 -> effect. Figure 5
page 11, model.We -> model. We
page 11, theHexi -> the Hexi
Author Response
Response to Reviewer 1 Comments
Point 1: page 1, ofadherent -> of adherent
Response 1: Thanks for your advice. We added spaces to it.
Point 2: page 1, theaccuracy -> the accuracy
Response 2: Thanks for your advice. We added spaces to it.
Point 3: page 2, layers [11].The -> layers [11]. The
Response 3: Thanks for your advice. We added spaces to it.
Point 4: page 2, Many network structures with good segmentation performance have been born in the field of image segmentation. - formulate the sentence differently
Response 4: Thanks for your advice. We rewrote this sentence in the resubmitted manuscript.
Point 5: page 2, semantic segmentation, It includes -> semantic segmentation. It includes
Response 5: Thanks for your advice. We modified the symbol.
Point 6: page 3, GCT(Gated Channe -> GCT (Gated Channe
Response 6: Thanks for your advice. We added spaces to it.
Point 7: page 3, UNETbased -> UNET based
Response 7: Thanks for your advice. We added spaces to it.
Point 8: page 3, in Figure 1.In -> in Figure 1. In
Response 8: Thanks for your advice. We added spaces to it.
Point 9: page 3, expectations.Fi- -> expectations. Fi-
Response 9: Thanks for your advice. We added spaces to it.
Point 10: page 4, rameters.For -> rameters. For
Response 10: Thanks for your advice. We added spaces to it.
Point 11: page 4, image,it -> image, it
Response 11: Thanks for your advice. We added spaces to it.
Point 12: page 5, After completing the coding part in the network, It is not clear what coding part means?
Response 12: Thanks for your advice. We gave a brief description in the resubmitted manuscript.
Point 13: page 5, ifx>0 -> if x > 0
page 5, ifx<0 -> if x < 0
Response 13: Thanks for your advice. We added spaces to it.
Point 14: remove folowing text:
ReLU has the following advantages. ReLU will not have the problem of gradient
disappearance. Because the gradient in the nonnegative interval is constant, it can main-
tain the final convergence rate of the network model in a stable state.
Response 14: Thanks for your advice. We deleted this paragraph from the resubmitted manuscript.
Point 15: page 5, can betrained -> can be trained
Response 15: Thanks for your advice. We added spaces to it.
Point 16: page 8, Table 1 align the results in the last row
Response 16: Thanks for your advice. We corrected the errors in Table 1 in the resubmitted manuscript.
Point 17: page 8, samples,and -> samples, and
Response 17: Thanks for your advice. We added spaces to it.
Point 18: page 8, From Figure 3we -> From Figure 3 we
Response 18: Thanks for your advice. We added spaces to it.
Point 19: page 8, figure 3 and following text are not relevant for this paper
Response 19: Thanks for your advice. Table 3 shows the change curve of loss function value of our network and other networks during training. We updated the style of Table 3 in the resubmitted manuscript to make it clearer.
Point 20: page 9, samples,especially -> samples, especially
Response 20: Thanks for your advice. We added spaces to it.
Point 21: page 9, methods,it -> methods, it
Response 21: Thanks for your advice. We added spaces to it.
Point 22: page 9, separatedby -> separated by
Response 22: Thanks for your advice. We added spaces to it.
Point 23: page 10, effect.Figure 5 -> effect. Figure 5
Response 23: Thanks for your advice. We added spaces to it.
Point 24: page 11, model.We -> model. We
Response 24: Thanks for your advice. We added spaces to it.
Point 25: page 11, theHexi -> the Hexi
Response 25: Thanks for your advice. We added spaces to it.
Author Response File: 
Author Response.pdf
Reviewer 3 Report
1. Some spaces are missing from Table 1.
2. Explain with details the post-processing step - the watershed algorithm and add a reference for this step.
3. Add some results that the proposed method fails.
Author Response
Major points:
Point 1: Some spaces are missing from Table 1.
Response 1: Thanks for your advice. We corrected the errors in Form 1 in the resubmitted manuscript.
Point 2: Explain with details the post-processing step - the watershed algorithm and add a reference for this step.
Response 2: Thanks for your advice. In the resubmitted manuscript, we added a simple description of the watershed algorithm and added a reference.
Point 3: Add some result that the proposed method fails.
Response 3: Thanks for your advice. In the resubmitted manuscript, we added the situation that the method we proposed was not very effective, and gave a brief description.
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
Corrections from the previous version of the review have been corrected.
