OCT Image Restoration Using Non-Local Deep Image Prior
Round 1
Reviewer 1 Report
This is an interesting research paper. There are some suggestions for revision.
1. The pros and cons of existing solutions should be discussed.
2. The issues of existing solutions that can be solved by the proposed solution should be discussed in detail. Why do existing issues have these issues?
3. The contributions of the proposed solution should be highlighted and marked one by one in introduction.
4. Please discuss the following similarity-based sparse representation.
G. Qi, Q. Zhang, F. Zeng, J. Wang, Z. Zhu, 'Multi-focus image fusion via morphological similarity-based dictionary construction and sparse representation', CAAI Transactions on Intelligence Technology, 2018, 3(2), 83-94
5. The discussions of deep image prior model in section 2.1 are too general. The listed equations are existing ones. Please specify your contributions in section 2.1 and cite the related references.
6. In line 116-117, it mentions "in order to make the network parameters converge more quickly to the image details". Why does the proposed non-local deep image prior model converge quickly? It is not clear.
7. In section 2.2, it mentions the noise can be removed and the similarity can be calculated more accurately. Please specify the related reasons in detail.
8. As shown in section 2.3, please explain how to decide the number of layers and the size of each layer.
9. The proposed algorithm needs to be formalized.
10. In line 189, it sets alpha to 0.7. Where is 0.7 from.
11. More objective evaluation indicators should be added for comparison.
12. The time complexity of the propose algorithm should be discussed.
In comparative experiments, the processing time of the proposed solution should be compared with existing ones.
Author Response
Please kindly see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors propose a new method for OCT Image Restoration according to deep learning. Since the idea is quite interesting, there are major points that need to be addressed to meet the publication requirements:
1. There are a lot of grammar mistakes and typos in this manuscript. The authors should re-check and revise carefully.
2. How did the authors select the optimal parameters for their deep neural networks?
3. There are few data in their experiments (only testing on 18 data is not convincing the stability of their model).
4. The authors should explain more how they collected data as well as some statistics about the data.
5. Deep neural network has been used in a lot of works in biomedical data such as PMID: 28643394, PMID: 31277574, or PMID: 31362508. The authors should add more references related to this information.
6. What are the idea and reason of U-net choice? Since the authors have few data and it maybe do not need a network with a lot of layers and parameters. It is important to compare their performance with a baseline model.
7. Which evaluation method that the authors used in this study?
8. The authors should have more discussions on their findings.
9. It is important that the authors could test their method on an external dataset.
10. For comparison, please have some statistical tests to see the significant differences.
11. The authors should release source codes for reproducing their results.
Author Response
Please kindly see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
All my concerns have been addressed. This paper is ready for publication.
Reviewer 2 Report
My previous comments have been addressed satisfactorily.
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
This manuscript proposes a deep neural network of non-local deep image prior for the de-speckle of OCT images. The author claims the method can achieve the superior performance with both OCT signal recovering and speckle noise removal. Some selected works are reviewed while the review is not convincing. The non-local deep image prior model is described unclearly and the corresponding network architecture is absent. Experimental results are given without clarity, which makes the results not convincing. Details are given below:
1. The method is not well described. A brief description of the deep neural network, U-net has to be explained. The learning algorithm and learning parameter are also necessary to be given.
2. The mathematical descriptions of the method have many problems. For example, Equation (2) should be arg min optimization but not min. x0 is not defined. And Equation (5) comes out without clear explanation.
3. Figures 1 and 2 should be improved. Those images are too small. Some critical image patches should be enlarged to help the visualization. The descriptions of these images such as "resist" "descend" are weird. Some classical descriptions such as "oversmooth" "unclear tissue boundary" will be greatly better.
4. Experiments are not enough to demonstrate the method. Training protocol is never explained. How many images of the 18 OCT noisy images are used to training, and how many are used fro testing? What are the learning parameters such as learning rate and momentum?
Besides, the effects of important parameter on the denoising results are not discussed. From Figures 1 and 2, learning iterations are important to the denoising result, but no experiments to analyze and discuss this experimental factor. In addition to learning iteration, the patch size, neighborhood size, and impact factor are all critical to the denoising results, but this manuscript does not discuss them. More experiments are necessary to discuss those factors with image quality metrics and figures.
5. The review of deep learning for denoising can be improved. Some very related methods such as "Denoising Prior Driven Deep Neural Network for Image Restoration", "Fully Symmetric Convolutional Network for Effective Image Denoising", and "Deep Learning for Image Denoising: A Survey" should be included for discussion.
6.English writing needs to be improved. There are grammar errors and weird descriptions.
Reviewer 2 Report
This is an interesting research paper. There are some suggestions for revision.
1. The motivation is not clear. Please specify existing issues and the importance of the proposed solution.
2. Please highlight your contributions in introduction.
3. Please discuss and compare pros and cons of existing solutions.
4. Please explain why R(x) can be replaced by the implicit prior captured by the deep convolution network.
5. Section 2.1 is too long. Please only keep the contents that are used later. Please add the related references for the remaining ones.
6. Please add a figure of framework to give an overview of the proposed solution.
7. What is your contribution in Section 2.2? It is not clear.
8. The relationship between Eq.6 and 7 is not clear. Please add the related explanations.
9. Please explain why the parameters of the neural network can be trained by Eq. 10.
10. As shown in Fig.2, please explain how many iterations are needed for the proposed solution or how to find the optimal number of iterations.
11. As shown in line 164-165, please explain where l=10 and alpha=0.7 come from.
12. Please add the experiment environment and the related reference for experiment images.
13. Please add more objective evaluation metrics to evaluate the performance of the proposed solution.
