A Systematic Literature Review on Applications of GAN-Synthesized Images for Brain MRI

Round 1

Reviewer 1 Report

This is a systematic review of the literature in the field of NGS for the generation of synthetic brain MRI. This review focuses on the last 6 years. Despite its quality, there are changes that I believe should be made:

- Respect the acronyms of everything that is defined. For example, Deep Learning is defined as (DL), but later the acronyms are not written, but Deep Learning is written again.

- Respect the capitalization of algorithms: deep learning.

- Respect spaces: C. Single network generation:

- There is a lot of talk about the benefits of MRI but very little about GAN. In addition, it should be added how to validate that the images are correct and it is not only at the discretion of the GANs.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper provides a systematic literature review on Applications of GAN-Synthesized Images for Brain MRI based on WoS and Scopus. It is an important contribution to the fast evolving field of GAN used in medical imaging. 

Major Comments:

- Writing needs to be improved significantly. Especially the use of the definite article 'the' as well as singular vs plural forms of nouns are often misused throughout the paper. I encourage the authors to carefully go through the whole manuscript regarding language and correct mistakes. E.g. in the last paragraph of the Introduction sometimes section and then Section is used; in 2.2. it says 'No SLR paper gives describes applications ... '; de-noising should be spelled without -; etc 

- The short descriptions of the different methods (Image Translation, Image Registration, etc) need to be carefully revised and backing up by references. 

- Table 8 needs to be connected to the revised manuscripts (which paper uses which loss function)

- Description of loss function should be more clear and contain formulas as well as references 

- Table 9 should also be connected to the revised papers. What kind of preprocessing was done? It would also be interesting to have charts to see what kind of preprocessing is most commonly used. Also needs references. 

- 3.4./Table 10 also lacks references to the evaluation metrics. Moreover, it is not clearly formulated that different kinds of images are compared here. E.g. dice metric compares segmented images, whereas SSIM or RMSE compare the full image. It is important to divide these categories since it tackles a completely different topic. Moreover, the descriptions of the measures need careful revision, they are not always correct or/and missing the most important information about the metric. 

Minor comments:

- Figure 1 is not clear. The title says 'The basic structure of a GAN' - this would be independent of a brain MRI. Also, preprocessing is not part of the structure of a GAN.

- WoS was not introduced as an abbreviation before Figure 2. 

- Figure 3 contains some missing part in 'Studies assessed for eligibility', the last column 'Studies included in the review' should contain the final number of considered studies

- Is there a reason why 'brain' was not considered as a keyword? I would expect since MRI already contains the word 'imaging' it would not necessarily have to include 'Brain Imaging' to be a paper on brain MRI 

- The terms Q1 and Q2 journals are not defined 

- Table 4 is not consistent in its language. E.g. 'Research studies have source..' and 'Research articles based on ..' Please use consistent grammar.

- Table 5: Divide more clearly the different topics (MRI-to-CT and MRI-to-PET) 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The author gives a review of current GAN-based image processing for Brain MRI. The overall contents are many.

But I have some suggestions and comments for this paper, as follows:

11.      As a review paper, I feel the author tries to involve as many works as possible, however, the contents are loosely structured. For example, image registration and motion correction are closely correlated, it’s better to merge them or find a way to make them closely organized.  Missing modality generation has some overlap with modality translation.

The author can refer the MRI image’s pro-processing steps from previous works, to better organize their work. Like MRI firstly motion correction, registration, reconstruction, super-resolution, enhancement, then many segmentation, modality translation work can be conducted.

22.       MRI scan is viral for brain navigation, detailed characteristics, and other cranial structures [3].?  Please pay attention to misnomer and mistypes.

33.      As a review paper, please give as complete reference works as possible. Like GAN variants like “ConditionalGAN, WasserteinGAN, and CycleGAN are more robust and used in image synthesis, translation, and super-resolution applications.”

44.      What is the 3.3. Brain MRI Preprocessing Software (RQ 3) for?  Is that also GAN-relevant? Please try to find some relevant work/software.

4.2. Transfer Learning is targeting to train a model under the condition that the training data is limited for the current task while taking advantage of previous relevant models. But the GAN is targeting to generate more simulated training data. I suppose transfer learning is not so relevant, or conflicted here. U-Net [133], 1119 ResNet [160], and DenseNet [161]. These architecture are employed in GAN generator or discriminator, which is common but looks not transfer learning requested. The author needs to better organize this part.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Nothing more to add.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Dear authors,

thanks a lot for incorporating most of my comments. 

- Medical descriptions in the Subsections of 3.1., such as in 3.1.4. "In MR imaging, different sequences (or modalities) can be received, providing valuable and distinct knowledge about brain disease. MR imaging typically acquires three contrasts in the image, T1, T2, and PD. The imaging process can highlight only one of them. " 
need some references. 

- I highly recommend to leave out the verbal descriptions of the loss functions (Table 8) and the performance measures (Table 10) or to completely overwork the summaries and add proper references. Since the outline of the paper is not the summary of loss functions or performance measures, and it is already very long, I would leave out the descriptions. In the present form it lowers the quality of the paper a lot.  For performance measures I then would recommend the following columns:
'Name', 'Fr/NR', 'Assessment', 'Range', 'Ref.No'
and for the loss function I would make an overview ordered by the number of papers using it, you could also think about sub-categories such as probability based etc

Author Response

Please see the attachment.

Author Response File: Author Response.docx