Improvement of Retinal Vessel Segmentation Method Based on U-Net

Round 1

Reviewer 1 Report

Retinal vessel segmentation is quite a delicate task to be done while operating on a patient. The authors proposed and improved U-Net based method to do this task efficiently. Overall paper presentation is quite good. However, in the results section, the accuracy of the proposed method is less, though a small difference,  than the state of the art. 

I would like the authors to mention what improvement they see can be done in a method that may result in improvement in accuracy. They can propose this as future work to guide the research community who may like to work on a similar problem.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The following revisions are required.

1. In literature review, add 3 to five more relevant and latest techniques.
2. Add Comparison table at the end of section 2 and compare with at least 8 to 10 techniques.

3. Please make sure your paper has necessary language proof-reading.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The author adapted the Unet model and added RESNET block and DFB to improve the model performance in the retinal vessel segmentation task. Please add more details on the model.

1. How the data was prepared for training, validation, and test set. Add numbers.  

2. Plot the training and validation curve against epochs. Mention the hyperparameters used in the training model-decay, learning rate, optimizer, iterations used to train, last layer activation function etc.

3. Add the Hausdorff dimension, True positive rates,  True negative rate parameters in metric Tables 2 &3.

4. What was the input image size, any processing (cropping, normalization, rescaling) done on images.?

5. Figure 6, DRIVE and CHASE_DB1 model prediction shows some fragments in the vessel and that the vessel is not continuous. Does the author try to apply any post-processing method for tackling? 

6. The methods used in table 3, present those model prediction images comparison in figure 6. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Traditional U-net criticisms:

U-Net is prone to "feature loss" due to the operation of the encoder convolution layer. 

Also U-Net causes the problem of mismatch in the processing of contextual information features caused by the skip connection part.

Proposed solutions: 

We propose an improvement of the retinal vessel segmentation method based on U-Net to segment retinal vessels accurately.

1. "to extract more features from encoder features": we replace the "convolutional layer with ResNest network structure" in feature extraction, which aims to enhance image feature extraction.

2. "to deal with mismatched processing of local contextual features by skip connections" a Depthwise FCA Block (DFB) module is proposed.

Comments:

1. The paper is well written.

2. However, the proposed modifications of the traditional U-net exist in the current literature and therefore do not contribute to the existing knowledge in the field.

3. The pre-processing step (see 2.5 Image pre-processing) as well as the mentioned sentence on data augmentation make subtle contributions in this direction. 

   In addition, effects of preprocessing/data augmentation were not quantified separately to determine their effectiveness. 

4. In Table 2 and 3 (i.e., on both datasets), the possessive pronoun "ours" could be replaced by the name of the proposed method. 

   In addition, as can be seen, the proposed method does not make significant segmentation accuracy difference when compared to the traditional U-net to justify the need for proposing the mentioned modifications of U-net.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Finding metrics (true positive, False positive, false negative) do not need to perform new experiments but analyzing the model inference with ground truth. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Thank you for updating the manuscript. Your paper is well written. However, there was probably a kind of misunderstanding in my previous comments:

In my comment (2) I wrote this: However, the proposed modifications of the traditional U-net exist in the current literature and therefore do not contribute to the existing knowledge in the field.

In the authors replies, authors wrote this: Thanks for the references, which are now included in the revised manuscript. We have added three more references to demonstrate the contribution in this area. 

As can be understood, I did not share or suggest any references for the authors to thank me for their inclusions in their manuscript. In addition, effects of preprocessing/data augmentation were not quantified separately (in the new version) to determine their effectiveness, etc.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 4 Report

Thank you for your efforts to improve retinal vessel segmentation method based on U-Net. Note that your proposed modifications of the traditional U-net exist in the current literature and therefore do not contribute to the existing knowledge in the field. I suggest either to (1) develop a new CNN specific to retina segmentation or (2) introduce a new pre-processing technique and evaluate its effects on the traditional U-net training outcomes.