Segmentation of Retinal Blood Vessels Using U-Net++ Architecture and Disease Prediction
Round 1
Reviewer 1 Report (Previous Reviewer 2)
The authors have improved the manuscript. However, a few minor issues still remain.
Comments:
1. Clearly state the novelty of the proposed methodology. On page 4, the contributions are listed, but any of the step such as using the Gaussian filter, Gabor filter, the HOG or LBP features, etc. has been done by other authors before.
2. Table 2: recognizing any disease should be easier than recognizing a specific disease. Why the specificity for “Existence of disease” is lower than the specificity for “Diagnosis of Retinopathy” and “Diagnosis of Macular”?
3. Figure 8: provide the units of measurement (%).
4. The IOU metric is defined in Equation 6, but the value is not given in any of the tables.
5. Some of the discussed references are not related to the topic of the paper, for example, refs. [8,11].
Author Response
Thank you for the valuable comment and constructive suggestion.
Author Response File: 
Author Response.pdf
Reviewer 2 Report (Previous Reviewer 3)
The manuscript is revised well and carefully and now I am fully satisfied with the revised version. In my opinion now the manuscript is ready for publication and so I recommend it to be accepted for publication in its current form.
Author Response
Thank you for your positive comments.
Reviewer 3 Report (New Reviewer)
1. The writing of the introduction section should be improved. First, the contributions of the paper should be summarized and highlighted. Second, I would suggest to start a new "Related Work" section to present the relevant works. Now the introduction is too dense and hard to follow.
2. In the related work section, more relevant research on medical image segmentation or general image segmentation should be discussed. Some works, e.g., Volumetric Memory Network for Interactive Medical Image Segmentation, Rethinking Semantic Segmentation: A Prototype View, should be discussed.
3. I am concerning whether the pre-processing modules in Fig.4 is essential. No experimental results are provided to confirm this. Can we just train the U-Net in an end-to-end manner?
Author Response
Thank you for your constructive and helpful comments.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report (New Reviewer)
The revision has addressed my concerns.
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
1- What is the new contribution you presented in this paper?
2-The presentation of some figures are not good. Improve them.
3- You have to deeply discuss the computations cost and the complexity of your method.
4- I think the authors have to add more comparisons, from recent years 2022, 2021.
5- What about the applications of other deep learning methods?
6- Improve the conclusion by adding the challenges and limitations.
7- Refresh your intro with recent segmentation methods, such as: Automatic superpixel-based clustering for color image segmentation using q-generalized Pareto distribution under linear normalization and hunger games search; Automatic clustering method to segment COVID-19 CT images; Detection of Diabetic Retinopathy in Retinal Fundus Images Using CNN Classification Models.
Author Response
Point1: What is the new contribution you presented in this paper?
Response 1: Thank you for the valuable comment and constructive suggestion. This manuscript used the novel segmentation technique for the Retinal Blood Vessels. We applied U-Net++ architecture and disease prediction using retinal fundus images. We add new information for more clarification added to page 4.
Point2: The presentation of some figures is not good. Improve them.
Response 2: The authors entirely agree with this comment. So, the figures’ quality has improved in the revised version. We also removed figure 1 as one of the reviewers recommended.
Point3: You have to deeply discuss the computational cost and the complexity of your method.
Response 3: We agree, and because of this method's high computational cost and complexity, our team is working on some modifications and improvements. We hope we can publish the outcome of our progress in future. We mentioned this issue as a limitation of this technique in the article.
Point4: I think the authors have to add more comparisons, from recent years 2022, and 2021.
Response 4: Thank you for the comment. We have updated Table 3 using the most newly published studies in the revised version.
point5: What about the applications of other deep learning methods?
Response 5: This is a good point and the idea of using other deep learning methods and the applications can be published as an extended review paper in future.
Point6: Improve the conclusion by adding the challenges and limitations.
Response 6: In the revised version, challenges and limitations were added to the conclusion and discussion section.
point7: Refresh your intro with recent segmentation methods, such as Automatic super pixel-based clustering for color image segmentation using q-generalized Pareto distribution under linear normalization and hunger games search; Automatic clustering method to segment COVID-19 CT images; Detection of Diabetic Retinopathy in Retinal Fundus Images Using CNN Classification Models.
Response 7: Some new articles were added to the introduction section on page 3 based on the reviewer's suggestion.
Reviewer 2 Report
The paper presents a deep learning based methodology for blood vessel segmentation and disease detection from retinal images. The results of experiments on two benchmark retinal image datasets are presented. The paper needs to be revised and improved according to the comments presented below before it could be considered for publication:
1. The novelty of the study is not clear. The U-Net architecture and its various modifications and optimizations have already been used before in various studies including for recognition of retinal diseases. Other parts of the methodology (Gabor filter, HOG and LBF feature extraction) are also commonly used. The authors must explicitly state their innovation in the research field and differences from previous works
2. The discussion on related studies is based on some older and outdated works in this rapidly developing research field, which no longer represent the state-of-the-art. In fact, only a single reference is newer than 2020. The selection of works to discuss seems to be ad hoc and the discussion is presented without any particular order. More recent state-of-the-art papers should be discussed. The authors are encouraged to consider Kadry, et al. (2021). Retinal vessel segmentation with slime-mould-optimization based multi-scale-matched-filter. Maqsood, et al. (2021). Detection of macula and recognition of aged-related macular degeneration in retinal fundus images. Rajinikanth, et al. (2021). Machine-learning-scheme to detect choroidal-neovascularization in retinal OCT image. Finalize by discussing the limitations of existing methods as a motivation of your proposed solution.
3. Add a workflow diagram to present and explain a sequence of steps in your methodology visually.
4. Figure 1 is trivial and textbook-level. Please remove.
5. Gabor filter, HOG and LBP are well-known. These methods have been described and explained many times before in other studies. There is no need to provide a detailed explanation. You can provide the supporting references and shorten subsections 2.2 and 2.4, while focusing on your own innovation instead.
6. Explain the selection of hyperparameter values for training, Did you do optimisation/tuning? Is it based on an ablation study?
7. Why was training discontinued after 200 steps? How do you avoid/prevent overfitting?
8. Table 2: more precise values must be given.
9. Figure 10: the value of “105%” makes no sense. Please remove.
10. Did you cross-validate? Explain in detail.
11. Statistical reliability of the results should be evaluated. You can use confidence limits or standard deviation, for example, based on values obtained from different runs or cross-validation folds.
12. Add the discussion section and discuss the limitations of your methodology.
13. The conclusions should be supported by main numerical findings.
Author Response
- The novelty of the study is not clear. The U-Net architecture and its various modifications and optimizations have already been used before in various studies including for the recognition of retinal diseases. Other parts of the methodology (Gabor filter, HOG, and LBF feature extraction) are also commonly used. The authors must explicitly state their innovation in the research field and their differences from previous works.
Response 1: Thank you for the valuable comment. As you mentioned, the U-Net++ technique is an extension of the popular U-Net architecture. The main aim of designing this technique is to improve the segmentation accuracy compared to the U-Net architecture. Therefore, we used this technique as well as a combination of feature recognition algorithms to increase accuracy and sensitivity. The idea of this combination and achieving improved results could be somehow considerable and novel. For more clarification, some explanation was added to Section 2, page 4 of the revised version.
- The discussion on related studies is based on some older and outdated works in this rapidly developing research field, which no longer represent the state-of-the-art. In fact, only a single reference is newer than 2020. The selection of works to discuss seems to be ad hoc and the discussion is presented without any particular order. More recent state-of-the-art papers should be discussed. The authors are encouraged to consider Kadry, et al. (2021). Retinal vessel segmentation with slime-mould-optimization based multi-scale-matched-filter. Maqsood, et al. (2021). Detection of macula and recognition of aged-related macular degeneration in retinal fundus images. Rajinikanth, et al. (2021). Machine-learning-scheme to detect choroidal-neovascularization in retinal OCT image. Finalize by discussing the limitations of existing methods as a motivation for your proposed solution.
Response 2: We agree, and based on the reviewer's suggestion, some newly published literatures were added to the revised version of the manuscript.
- Add a workflow diagram to present and explain a sequence of steps in your methodology visually.
Response 3: The workflow was added as Figure 4 to the revised version.
- Figure 1 is trivial and textbook-level. Please remove.
Response 4: Thank you for highlighting this issue. We removed Figure 1 from the manuscript. Please check out the revised manuscript.
- Gabor filter, HOG, and LBP are well-known. These methods have been described and explained many times before in other studies. There is no need to provide a detailed explanation. You can provide supporting references and shorten subsections 2.2 and 2.4 while focusing on your own innovation instead.
Response 5: We agree, the changes mentioned in sections 2.2 and 2.4 were applied in the revised version.
- Explain the selection of hyperparameter values for training, did you do optimization/tuning? Is it based on an ablation study?
Response 6: As mentioned on page 14 of the article, Adam's optimization method was used in our study.
- Why was training discontinued after 200 steps? How do you avoid/prevent overfitting?
Response 7: As the reviewer knows, small data as well as the presence of noise or unprocessed data, will cause overfitting. To avoid this issue, we performed both pre-processing and the Data Augmentation method. Therefore, we could remove noise and increase the amount of data.
- Figure 10: the value of “105%” makes no sense. Please remove.
Response 9: The reviewer is correct. We modified Figure 10 in the revised manuscript..
- Did you cross-validate? Explain in detail.
Response 10: We used separate training and test images to train and test the network. The cross-validation method was not applied in our study as this method is computationally costly. Additionally, the time complexity of our work is almost high, so we did not use it.
- Statistical reliability of the results should be evaluated. You can use confidence limits or standard deviation, for example, based on values obtained from different runs or cross-validation folds.
Response 11: Thank you very much for the valuable suggestion. We will consider the mentioned point in our future study.
- Add the discussion section and discuss the limitations of your methodology.
Response 12: The discussion section and the method's limitations were added to the revised version.
- The conclusions should be supported by main numerical findings.
Response 13: The conclusion was modified based on the suggestion of the reviewer.
Reviewer 3 Report
The manuscript contains novel idea, written well and is correct scientifically and mathematically. There are some minor issues regarding language errors. I suggest that the manuscript should be accepted for publication after a revision of language issues.
Author Response
Thank you for the valuable comment.
Round 2
Reviewer 1 Report
The authors addressed all comments in last round, so, this new version can be accepted for publication.
Reviewer 2 Report
The manuscript was well revised and can be accepted for publication.
