A Segmentation Method Based on PDNet for Chest X-rays with Targets in Different Positions and Directions
Round 1
Reviewer 1 Report
The paper proposes a Position and Direction Network (PDNet) for chest X-rays with different 3 angles and positions that provides more comprehensive information for cardiac image diagnosis, as well as as for surgery guidance.
The paper is well-written and well-presented. I have a few comments and suggestions :
1. The abstract is too general. It must explain the results achieved or the enhancement compared to other methods.
2. Please add a "Literature Review or Related Work" section to the paper.
3. Mention the novelty of the paper in the introduction section.
4. How many datasets do authors use? Is it only one (the JSRT dataset)? The JSRT dataset is an openly accessible dataset, and it is required to cite the source when using a dataset that is available to the public. Please cite the dataset's source.
4. The number of datasets the writers use. Is the JSRT collection the only one? The JSRT dataset is an openly accessible dataset, and it is required to cite the source when using a dataset that is available to the public. Please cite the dataset's source.
5. Since the dataset is openly accessible, it must be conceivable that it has been used in other related works (previous studies). If so, the authors should contrast the results of the proposed model with those of the previous studies.
Author Response
Thank you very much for your valuable opinions and suggestions. The following are responses and revisions according to your opinion.
- The abstract is too general. It must explain the results achieved or the enhancement compared to other methods.
We have made improvements based on the feedback provided regarding the summary. please refer to the highlighted section for the changes.
- Please add a "Literature Review or Related Work" section to the paper.
The article has been restructured.We have revised section 1.2 and added section 2.2 on to form a new chapter "related work". I've highlighted it in red.
- Mention the novelty of the paper in the introduction section.
The penultimate paragraph of Chapter 2 Related Work adds a description of the novelty of the article. I've highlighted it in red
- How many datasets do authors use? Is it only one (the JSRT dataset)? The JSRT dataset is an openly accessible dataset, and it is required to cite the source when using a dataset that is available to the public. Please cite the dataset's source. If the author only used the JSRT dataset, then it is not necessary to provide a citation for the dataset. However, if the JSRT dataset is used, the source should be cited appropriately. The JSRT dataset is freely accessible and can be cited as follows.The number of datasets the writers use. Is the JSRT collection the only one? The JSRT dataset is an openly accessible dataset, and it is required to cite the source when using a dataset that is available to the public. Please cite the dataset's source.
We really only used one dataset. There is indeed more than one dataset for chest x-rays, and the others are larger, but chest X-ray images have fewer data sets that are standard hand-segmented gold because physicians often label them incorrectly. JSRT is a publicly available dataset, and although the database is relatively old, each image in it has been confirmed by at least three physician and its correctness has been verified by CT. At present, it is the most valuable reference and is a complete and authoritative database. Its citation source has been added to the text and marked in red in the database description paragraphs.
- Since the dataset is openly accessible, it must be conceivable that it has been used in other related works (previous studies). If so, the authors should contrast the results of the proposed model with those of the previous studies.
There are indeed many studies based on this database. The article adds a list of other article segmentation results(The red on the table), and this article also reproduces some networks as pre-transformation networks(*-NO line). Due to the different designed parameters, the results are not exactly the same. In order to facilitate the comparison of the results before and after the transformation, all network parameters in this paper are unified. However, there are no articles studying at image segmentation after displacement and rotation (we guess they all assume that the network has linear invariant features), which is the focus of our study. In order to make this paper convincing, the algorithm of this paper is used to transform four networks and compare the results before and after the transformation.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
In the paper entitled „A Segmentation Method based on PDNet for Chest X-Rays with Targets in Different Positions and Directions” by Wu et al. a Position and Direction Network (PDNet) for chest X-rays with different angles and positions is proposed. The accuracy of segmenting targets was improved when a weight mask that identified the position and direction of the object in the middle layer was used.
1.The introduction should be restructured without subsection 1.2. Image Segmentation Method and 1.1. Chest X-Ray Image Segmentation.
2. The paper does not contain the Related work section
3. In section 3. Proposed Method please and an overview graph of your work
4. Figure 3. Convolution block. It should be improved and the red line from the box Concat removed
5. Please add the number of epochs for each image in Figure 4. Gradient-like ground-truth position maps.
6. Please compare your result with others from the scientific literature
7. Please add the limitations of the study
8. The active shape model (ASM) is the main element of the paper, but this is not described clearly, it is necessary for the mathematical approaches
9. Please add a table with all hyperparameters of used CNNs
Minor observations
In the text is used 8* operator 3*3, 5*5 2014*2014, please replace in the whole text with the corect operator 
It is unclear if figures 1 and 2 are created by authors, if not please add corresponding references
The letter k* is not in figure 3
Line 183, please add the correct argument for S(∗)and F(∗)
In Equation 4,5,6 the operator is 
not ≢
Line 243: what does mean Fpd?
Line 255: ds, please add s as subscript
Line 257: is used superpixels, but this notion is not taken into account previously
Lines 319, 331, and 325 contains capital letter inner of the sentence
Line 456: what mean s (*-DR and PD-*- DR), please check the entire manuscript
The public CXR database JSRT is not described in detail
4.3. Evaluation Indicators the accuracy metric is not described but is computed
Lines 382, 383 „The training parameters were MIoU = 0.707831, acc = 382 0.941764, sensitivity = 0.761343, specificity = 0.960188, and best_auc_sum = 1.58837” (are necessary only three decimals) the acc, sensitivity, specificity, best_auc_sum but these are not described
Figure 9. it is unclear, please improve it
Author Response
Since the article is a requirement for my doctoral graduation, I submit it in a hurry. If it had not been for your careful examination I would not have known so many mistakes in it. Thank you very much for pointing out the errors and suggestions on the revision of the manuscript. The following are responses and revisions to your comments.
1. The introduction should be restructured without subsection 1.2. Image Segmentation Method and 1.1. Chest X-Ray Image Segmentation.
2. The paper does not contain the Related work section.
The manuscript has been restructured.We have revised section 1.2 and added section 2.2 on to form a new chapter "related work". I've highlighted it in red.
3. In section 3. Proposed Method please and an overview graph of your work.
The work of this manuscript is presented in Chapter 4 (formerly Chapter 3) by listing the highlighted parts in red. It is difficult to use a graph to represent here, and the proposed new network architecture diagram is already shown in Figure 6.
4. Figure 3. Convolution block. It should be improved and the red line from the box Concat removed.
Figure 3. has been corrected and improved.
5. Please add the number of epochs for each image in Figure 4. Gradient-like ground-truth position maps.
Since there are too many intermediate layers to fully display, the more representative ones were selected to present. The five-column gradient-like feature maps are shown in Figure 5.
6. Please compare your result with others from the scientific literature.
There are indeed many studies based on this database. The manuscript adds a list of other article segmentation results(The red on the table), and this manuscript also reproduces some networks as pre-transformation networks(*-NO line). Due to the different designed parameters, the results are not exactly the same. In order to facilitate the comparison of the results before and after the transformation, all network parameters in this paper are unified. However, there are no articles studying at image segmentation after displacement and rotation (we guess they all assume that the network has linear invariant features), which is the focus of our study. In order to make this paper convincing, the algorithm of this paper is used to transform four networks and compare the results before and after the transformation.
7.Please add the limitations of the study.
Limitations of the study have been added in Chapter 6, the sentences highlighted in red.
8. The active shape model (ASM) is the main element of the paper, but this is not described clearly, it is necessary for the mathematical approaches.
A more detailed description of ASM has been added in section 3.2, highlighted in red. The manuscript only applied it and did not improve it. The initial contour was replaced during use.
9. Please add a table with all hyperparameters of used CNNsMinor observations.
Table 1 has added hyperparameters for various networks marked in red.
In the text is used 8* operator 3*3, 5*5 2014*2014, please replace in the whole text with the corect operator.
All * operators in the manuscript have been corrected.The operator that does not represent the same meaning has also been modified with other symbols and marked in red.
It is unclear if figures 1 and 2 are created by authors, if not please add corresponding references.
Originally,I thought that the reference of figure could be marked in the text, so it was only marked in the text. Figure 1 has now been corrected. Figure 2 is my own drawing, and the algorithm reference have been clearly marked.
The letter k* is not in figure 3.
The description of k * in the text has been corrected and highlighted in red.
Line 183, please add the correct argument for S(∗)and F(∗).
The argument for S (*) and F (*) have been added marked in red.
In Equation 4,5,6 the operator is not ≢.
If the network itself is linear invariant, that are,S(D(x)+dD(x))=S(D(x))+S(dD(x)), S(A*D(X))=A*S(D(X)) . However, S (*) cannot fully represent the relationship F (*) between pixels in the image. So its linear invariance may not necessarily hold true. The linear invariance is disconfirmed. We use ”not identities” to represent in the text , and the symbol used is' ≢'.
Line 243: what does mean Fpd?
Fpd in the figure 6 was corrected and marked in red.
Line 255: ds, please add s as subscript.
ds was corrected and marked in red
Line 257: is used superpixels, but this notion is not taken into account previously.
The notion of superpixels has been corrected in straightaway terms and marked in red .
Lines 319, 331, and 325 contains capital letter inner of the sentence.
The capital letters in section 4.3 have been corrected and marked in red.
Line 456: what mean s (*-DR and PD-*- DR), please check the entire manuscript.
The * - expression of the entire manuscript has been corrected and marked in red, as shown in the sections marked in red in the following paragraphs of the Table.
The public CXR database JSRT is not described in detail.
The JSRT database has been supplemented with description, as indicated in red in section 4.1.
4.3. Evaluation Indicators the accuracy metric is not described but is computed.
Lines 382, 383 „The training parameters were MIoU = 0.707831, acc = 382 0.941764, sensitivity = 0.761343, specificity = 0.960188, and best_auc_sum = 1.58837” (are necessary only three decimals) the acc, sensitivity, specificity, best_auc_sum but these are not described.
The description was corrected and was marked in red.
Figure 9. it is unclear, please improve it.
Figure 9 is magnified.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
All the comments and suggestions have been addressed
Reviewer 2 Report
The proposed manuscript was improved, and it can be published in its present form.
