Automatic Hepatic Vessels Segmentation Using RORPO Vessel Enhancement Filter and 3D V-Net with Variant Dice Loss Function

Round 1

Reviewer 1 Report

The authors introduced a workflow segmenting liver vessels from CT images, including vessel enhancement, image fusion, and segmentation. The proposed method was also validated on the improved 3D-IRCADb dataset. I have some comments below.

 

1. How is data encrypted during transmission between the local hospital and IT4I? Is privacy a concern? 

 

2. How is the dataset split? Is it random or manually chosen?

 

3. Could the author provide more details on how the additional labeling was done? 

 

4. For workflows in figures 5 and 6, have the authors considered using neural networks to do the fusion instead of summation?

 

5. In figure 6, are the four segmentation network the same or different?

 

6. In figures 5 and 6, why not use both Hessian and RORPO at the same time?

 

7. In figure 7, what is the motivation for using linear blending? Can't the network directly process both images? 

 

8. Could the authors add a section introducing the hardware used for training and inference?

 

9. Have the authors tried linear blending on other vessel enhancement methods besides RORPO?

 

10. The dataset size is tiny, is overfitting a concern? Could the authors also comment more on the generalization of the method?

 

11. Could the authors provide more details on how experiment 4 was performed? Specifically, how did the authors choose filters to be ablated? Was the ablation after the model training? Furthermore, what is the point of this experiment?

 

12. In table 7, are the metrics calculated on the same test dataset? If not, are the results between different methods comparable?

 

13. It might make the structure of the paper clearer to put some contents in sections 3 and 4 into methods.

Author Response

"Please see the attachment."

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript presents a new pipeline to automatically segment hepatic vessels. It gives a good contribution to the community by providing a plugin extension for 3D Slicer which can be used by clinicians and/or researchers to segment the liver vessels. A good additional contribution of this paper could be making the improved annotations publicly available.

The paper is well-written and clear. I have a few questions and suggestions I would like to see addressed to improve the manuscript.

Detailed comments:

-        Abstract

o   Please do not use acronyms in the abstract, e.g. RORPO.

-        Introduction

o   In formal English, abbreviations should not be used. Replace “doesn’t” by “does not”.

-   Related Work

o   Second paragraph: The following sentence is not clear “The extracted liver vessels were more complete and continuous when **applied to refined** than the original dataset.”

o   In sentence “They evaluated MultiRes U-Net architecture as the best for segmenting liver blood vessels on the 3D-IRCADb dataset.”, please specify what the best is?

o   In sentence “then the maximum was applied to extract the predicted segmentation.”, please specify what maximum are you referring to? Applied to what?

o   Please provide references to the datasets mentioned when they are mentioned for the first time.

o   Are 3D-IRCADb and 3Dircadb the same dataset? Please use the same name.

-        Datasets and Improved Annotations

o   Please define 3D-IRCADb acronym the first it is used.

o First paragraph: There are no “chapters” in this document, maybe you meant “section”?

o   What is the difference between “3D-IRCADb” and “3D-IRCADb-01”? In section 4.1, you mention it has 20 volumes, while in Table 1 it says 22 patients. Are these correct?

o   Section 4.1, 1st paragraph: what paper are you referring to in “the classes used in this paper”? Is it the reference [32]? Please clarify this in the text.

o   A new subsection 4.1.1 is not needed. This could be included in section 4.1.

o   As mentioned before, a good additional contribution of this paper could be to make the improved annotations publicly available. As stated in this section, this additional work could have been avoided if [5,35] provided their own corrections.

 

-        Methods

o   In Section 5.2, please rephrase the sentence “it is image data preprocessing, vessel enhancement, and image segmentation”. It is not clear what you meant by this sentence.

o   In Section 5.2, “Figure 5, 6, 7 shows” should read “Figures 5, 6, 7 show”

o   In all section 5.2 and subsections, it is not clear which methods are yours or are from other authors. In section 5.2.1, it seems you are only describing your own methodology. When presenting each pipeline or method, please state clearly if it is your method or reference the author.

o   In Section 5.2.1, what are “80HU” and “220HU”?

o   In Section 5.2.2, equations 15 and 16 do not seem to represent what is shown in Figure 6. In this case, you do not consider a sum at the end of the segmented results but a personalized intersection between pairs of segmented results of each filter.

o   In Section 5.2.3, how is the alpha coefficient determined? You explain this later but it should be stated in this section.

o   In Table 2, are alpha and beta different from the parameters of the linear blending? Please give a distinct parameter name.

-        Experiments and results

o   In Table 3 caption, “nad" should read “and”

o   The division in subsection 6.2.1 is not needed as both sections 6.2 and 6.2.1 are related to table 5.

o   Section 6.3.1, please explain clearly the relationship between “blending with 60% of the enhanced image” and values of alpha and beta.

o   The loss functions presented in 6.3.2 should have been defined in section 5.3.3. when you first present the loss functions. You can briefly describe them and cite their definition.

o   In Section 6.3.2, it is important to have defined beta with a different name to avoid confusion with the beta of the linear blending.

o   Figure 12 does not add new information to the paper.

o   Section 6.5. could show another case where the segmentation result is worse.

-        Discussion

o   The discussion should be more than comparing the performance of your methods with the previous ones. You should address these methods in a more critical manner, addressing also their advantages/disadvantages and limitations.

o  What is the usefulness of the fourth experiment? What did you modify in your pipeline according to these results? Its objective is not clear.

o   Limitations of your method should be discussed. For example, there are a few parameters that you still need to define: alpha of blending and beta of loss function.

Author Response

"Please see the attachment."

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Thank the authors for the revision. All my comments have been addressed properly. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript was significantly improved. However, there are still a few points that should be addressed before publication.

- One of my previous suggestions was to make the improved annotations publicly available as an additional contribution to the community. The authors answered "yes, we agree". Are you planning to make that available with the manuscript? 

- Please review again English abbreviations in the whole manuscript (e.g. line 221, line 828). They should not be used in formal English.

- Line 270: “Figure 5, 6, 7 show” should read “Figures 5, 6, 7 show”

- Figure 12: Please check if the colours and/or order of images are correct, specially in (c)

- Line 836: number of section is missing

Author Response

Please see the attachment.

Author Response File: Author Response.pdf