Skin Characterizations by Using Contact Capacitive Imaging and High-Resolution Ultrasound Imaging with Machine Learning Algorithms

Round 1

Reviewer 1 Report

The authors present a paper about "Skin Characterizations by Using Contact Capacitive Imaging and High Resolution Ultrasound Imaging with Machine
Learning Algorithms".

The topic is interesting anf the authors have demonstrated an in-depth knowledge of the subjet both in structuring the article and in the reference selection.

In the discussion section I would suggest that the authors get into deeper details regarding not only the technical aspects but also the possible clinical implications for example for discriminating neoplasitc lesions; in addition  the role of artificial intelligence has been recently highlighted also for treatment reasons. Some good references to consider about this point include PMID: 33299440 and PMID: 34182511

 

Author Response

Discussion section has been added to cover neoplasitc lesions and the role of artificial intelligence. Two references also added.

Reviewer 2 Report

The article entitled: “Skin Characterizations by Using Contact Capacitive Imaging and High Resolution Ultrasound Imaging with Machine Learning Algorithms” is in line with the Applied Sciences journal. The provided research has a practical dimension and covers up-to-date topic connected with skin diagnosis. Before the publication the article requires some changes or clarifications:

- Abstract: please specify scientific / research aim.

- Abstract: add the most important results

- Introduction  (first paragraph), please avoid further tense, the research usually have been done before publication.

- Introduction (line 39), please explain more detailed what is include in given references.

- Introduction – please add the paragraph with comparison used methods with other diagnostic methods; describe advantages as well as limitations.

- Introduction – please stress the novelty aspect of provided research.

- 2.1. please be focused more than method than equipment.

- Results and Discussion – first part should be in chapter 2 as an description of applied method (maybe point 2.3).

- Figure 2, please comment an approximate magnification for this pictures.

- Results and Discussion (lines 188-191) – please comment more detailed each issue.

- Results – more detailed comparison with the literature is required as a discussion.

Author Response

We have revised according to reviewer's comments.

1. Abstract revised, most import results added.
2. Introduction - further tense removed.
3. Introduction (line 39), more details added.
4.  Introduction – a paragraph of comparison added.
5.  Introduction – added the novelty aspect.
6.  2.1 revised and moved the first part of Results and Discussions to 2.3.
7.  Figure 2, added image pixel information.
8. Results and Discussion (lines 188-191) – more information added.
9. more detailed comparison with the literature added.

Reviewer 3 Report

This manuscript presents an interesting work. The authors used a modern high-resolution ultrasound imaging system and applied machine learning approaches to results classification. In my opinion, several comments below can improve the readability of their manuscript.
1. Introduction requires formulation of aim and main contribution of this paper.
2.  If the authors did not plan to give a related work I suggest this description can be extended in the introduction section especially for AI applied approaches.
3. The authors mentioned the use of different deep neural network architectures (lines 78-79). However, In my humble opinion, the use of the pre-trained networks only without any transfer learning or redesigning of output layers is a difficult problem. Could the authors describe these points more clearly about retraining existed architectures on the other datasets?
4. Table 1 presents some statistical characteristics, eg. mean and std deviation. But if the reader is not informed about normal statistics characteristics, consequently those estimations are not clear.
5. In section 3.1.2 if the authors give the exact time estimation then exact hardware and software description is required.
6. It is better to give a description of how the authors conducted the machine learning classification and obtained the results in Table 5. Which kind of data has been used for those tasks?
7. Fig. 8 and 9 require a description of obtained principal components.
8. Conclusion section requires discussion about the best deep learning network architecture suited for their research aims.

Overall estimation of this manuscript - it is worth to be considered for publishing after major revision.

Author Response

We have revised the manuscript according to the reviewer's comments. The following are the details of each comment point.

1. Introduction - revised.
2. Introduction - extended.

3. Transfer learning is a standard approach for texture analysis when there are not enough images for either full training or fine tuning, and this was just the case in our study. Consequently, we used pre-trained convolutional networks ‘off-the-shelf’ as feature extractors for exploratory data analysis in our experiments. Specifically, our image features were the L₁-normalised output of one intermediate layer of each network – the specific layers are specified in lines 280-281 of the original submission. This procedure has been validated in a number of previous works (see Refs [32–34] of the original submission) and is considered highly effective for texture analysis.

4. Table 1 is the results of normal statistic results from the normal healthy volunteers. 
5.  Hardware and software description added.
6.  The X training data used are the feature values, 6 factor values per channel and 18 factors per overall image. The Y training date are 1-7, representing 7 different skin sites. As illustrated in lines 212 -221.

7. Principal Component Analysis was carried out via Singular Value Decomposition (SVD) and LAPACK implementation (Python scikit-learn 0.23.2 / sklearn.decomposition.PCA). The input data was centred but not scaled for each feature before applying SVD.

8. Conclusions revised.

Round 2

Reviewer 2 Report

The authors corrected the most important issues, however the discussion is still quite generic and can be improved. The atricle requires some editing job. 

Author Response

Discussions updated and overall text is checked.

Reviewer 3 Report

The paper has been improved, so I can recommend it for publishing.

Author Response

We appreciate the reviewer's confirmation about the manuscript.