AsymUNet: An Efficient Multi-Layer Perceptron Model Based on Asymmetric U-Net for Medical Image Noise Removal
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsTitle: AsymUNet: an efficient Multi-Layer Perceptron model based on asymmetric U-Net for medical image noise removal
The work presented is part of an area of high interest in the scientific community, as well as having a very relevant applicability.
In my opinion, there are many points where it needs to be improved.
In the abstract, when you type "... where medical resources..." (line 4), I think you want to refer to hardware/computational resources and not medical resources.
There are cases in which authors refer to "a characteristic", which is not clear from the proposed model. Examples are found in lines 203 and 204. Do the authors actually intend to refer to "a feature" or a vector of features?
Some figures need to be improved, as they are difficult to analyze/read, as is, for example, the case in figure 1.
Figure captions must be reviewed. For example, in the case of figure 1, the caption should be corrected to something like "AsymUNet General Architecture". Figures 5, 6 and 7 do not correspond to diagrams, but to images.
The description of the proposed model must be more rigorous and detailed.
Validation of the performance of the proposed model must be supported by a greater number of results. The results explained in figures 5, 6 and 7 do not seem sufficient to me since in the last two examples the differences compared to the results of other models are not visually noticeable.
Conclusions must consider a contextualization of the work presented. The improvement in performance of the proposed model in relation to existing ones must be quantified.
The text must be carefully proofread, correcting typos, grammar and some sentence constructions.
Comments on the Quality of English LanguageThe text must be carefully proofread, correcting typos, grammar and some sentence constructions.
Author Response
Point 1: In the abstract, when you type "... where medical resources..." (line 4), I think you want to refer to hardware/computational resources and not medical resources.
Response 1: Thanks for indicating the problem. We have changed "medical resources" to "computational resources" in the Abstract section (Line 4).
Point 2: There are cases in which authors refer to "a characteristic", which is not clear from the proposed model. Examples are found in lines 203 and 204. Do the authors actually intend to refer to "a feature" or a vector of features?
Response 2: Agree. To make the meaning of the article clearer, We have changed "feature" to "feature map" in the revised manuscript. You can see these changes in lines 207 and 208. In addition, we have revised similar issues throughout the revised manuscript.
Point 3: Some figures need to be improved, as they are difficult to analyze/read, as is, for example, the case in figure 1.
Response 3: Thanks for indicating the problem. We have replaced the figure 1 with clearer version (Page 4).
Point 4: Figure captions must be reviewed. For example, in the case of figure 1, the caption should be corrected to something like "AsymUNet General Architecture". Figures 5, 6 and 7 do not correspond to diagrams, but to images.
Response 4: Thanks for indicating the problem. We have already modified the captions for these images (Pages 6, 7, 10, 11 and 12).
Point 5: The description of the proposed model must be more rigorous and detailed.
Response 5: Thanks for indicating the problem. We have now incorporated detailed descriptions of the model (Lines189-192) and have included annotations for specific components in Figure 1.
Point 6: Validation of the performance of the proposed model must be supported by a greater number of results. The results explained in figures 5, 6 and 7 do not seem sufficient to me since in the last two examples the differences compared to the results of other models are not visually noticeable.
Response 6: We agree with your opinion. We have replaced the images in Figures 5 and 6 to better demonstrate the effectiveness of our method. (Page 10).
Point 7: Conclusions must consider a contextualization of the work presented. The improvement in performance of the proposed model in relation to existing ones must be quantified.
Response 7: Thank you for pointing this out. We have added the performance metrics of our proposed AsymUNet on the AAPM dataset in the conclusion. Additionally, we have emphasized the advantages of our algorithm compared to the suboptimal algorithm EFormer. (Page 13, Lines 410-413).
Point 8: The text must be carefully proofread, correcting typos, grammar and some sentence constructions.
Response 8: Thanks for indicating the problem. We have proof-read the manuscript carefully. The typos, grammar, and some sentence construction problems in the revised manuscript have been greatly improved.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsIn this paper, we propose AsymUNet, an efficient multi-layer perceptron model based on an asymmetric U-Net for medical image noise removal. AsymUNet utilizes an asymmetric U-Net structure to reduce computational burden and enhances feature interaction between the encoder and decoder through a multiscale feature fusion module. Additionally, spatial rearrangement MLP blocks are used to effectively extract local and global features of the image. Experimental results demonstrate that AsymUNet shows superior performance metrics and visual results compared to other state-of-the-art methods. This paper demonstrates the excellent performance of AsymUNet on Gaussian grayscale/color medical images and low-dose CT images.
1. Equation 4 (L1(IO, IR) = 1/T ∑Tt=1 ∥ItO − ItR∥): While the definition of the L1 loss function is clear, the rationale and benefits of using a multiscale loss function need to be explained more clearly.
2. The presented PSNR and SSIM values: The presented PSNR and SSIM values indicate that AsymUNet outperforms other methods. However, it is necessary to add statistical significance tests (e.g., t-tests) for each metric to enhance the reliability of the results.
Author Response
Point 1: Equation 4 (L1(IO, IR) = 1/T ∑Tt=1 ∥ItO − ItR∥): While the definition of the L1 loss function is clear, the rationale and benefits of using a multiscale loss function need to be explained more clearly.
Response 1: Thank you for pointing out the problem. The multi-scale loss function calculates the loss between the input and output feature maps at each level and accumulates these losses, thereby obtaining the overall loss of the neural network, which leads to better denoising performance. (Lines 283-284).
Point 2: The presented PSNR and SSIM values: The presented PSNR and SSIM values indicate that AsymUNet outperforms other methods. However, it is necessary to add statistical significance tests (e.g., t-tests) for each metric to enhance the reliability of the results.
Response 2: Thank you for pointing out this issue. In the field of image denoising, the mean and standard deviations are typically used to indicate the reliability of the results [18] [22]. In the original manuscript, we initially presented only the mean values for the measurement metrics of different algorithms. We have now added the standard deviations to better illustrate the results in the revised manuscript. More details can be found in Table 1 and Table 2. (Page 9, Page 11)
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsDear Authors,
Thank you for your answers to my previous questions/comments and for reviewing the initial manuscript. I recommend enlarging figure 1 so that the reader can analyze it.
Author Response
Point 1: I recommend enlarging figure 1 so that the reader can analyze it.
Response 1: Thanks for your suggestion. In the manuscript, we have enlarged the figure 1(Page4).
Author Response File: Author Response.pdf
