A Novel Approach for Diabetic Retinopathy Screening Using Asymmetric Deep Learning Features

Round 1

Reviewer 1 Report

1) Readability is poor and obvious typos exist in the whole manuscirpt. For example, Line 334 - 335, What does the area under the curve represent? What is the relationship between AUC and other evaluation indices including sensitivity, precision, accuracy, specificity and f1-score? In Figure 2 the last column "Optical Dsic" was wrong!

2) Comparison with different algorithms (Table 3) is not clearly explained.For example, ref[4] utilised DIABET DB1 and Diaretdb0 v 1 1 database for diabetic retinopathy detection, while the manuscript employed APTOS and MESSIDOR dataset. A good suggestion is to test the performance of different algorithms using the same image sets. 

3) Four parameters have been used to evaluate the algorithm performance.Which one provides a more reliable outcome? Or in another word, how to keep a balance when the performance indices are not consistent with each other, e.g. Class-4 in Table 1 has lower precision, recall and f1-score but higher accuracy rate?

4) In the proposed deep learning structure, SVM was used as the classifier. However, there is a lack of information regarding SVM such as the kernel function. In addition, executing time is a critical index when comparing different DL algorithms. Please amend the time consumed for training and testing classification?  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this work, the authors presented a multi-modal model for Diabetic Retinopathy Screening. The proposed network relies on two U-Net for the segmentation of blood vessels and optical discs in the images. Images with blood vessel and optical disc removal were used as input for subsequent CNN+SVM network training and testing. The results indicated an improvement in model performance over prior arts. Overall, the manuscript is well-written and the conclusions are generally supported by the results. The quality of the manuscript can be improved by addressing the following:

1. Line 21: avoid using acronyms without spelling them out for the first occurrence. (SVM)

2. Line 31-32: similarly, define Diabetic retinopathy (DR) .

3. Line 104: include short comparison with prior arts (summarizing table 3 content) here.

4. Figure 5: the input image here should be green image with blood vessels and optical discs removed?  

5. Figure 5: remove the red wavy lines under Relu

6. Accompany class-wise performance table (e.g., table 1) with a confusion matrix.

7. The contribution of the work is mostly 1. clean the image (removing blood vessels and optical discs from image) using U-Net before training "to enhance the individual learning performance". 2. using CNN as a feature extractor that feeds an SVM classifier. Therefore it is necessary to:

a. perform the same set of training and testing WITHOUT using the U-Net for image pre-processing, and compare this result with the reported result to illustrate the "enhance the individual learning performance" aspect.

b. clarify whether the SVM helps the model performance at all? In other word, try to remove the SVM and use a 4-node output layer, train and test with this setup and compare its performance with the reported model.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper proposes a novel diabetic retinopathy screening technique using an asymmetric deep learning feature. My comments are as follows:

1- The limitations of this work should be discussed in the experimental result and discussion section.

2- Some new references should be added to improve the literature review—for example, https://doi.org/10.3390/app12115500; https://doi.org/10.3390/diagnostics12081975.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Better to have a native speaker to revise the manuscript and improve the readability. "=" is the same as "represent", so please remove duplicate words (line 224 & Line 227).  "Fig.2 show" should be "shows" (Line231). In addition, figure abbreviations should be consistent in the whole paper (line235 & line231)

Author Response

Please find the round 2 review attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Thanks for addressing the last round of comments.

Author Response

Please find the round 2 review attachment.

Author Response File: Author Response.pdf