Machine Learning Methods in Skin Disease Recognition: A Systematic Review

Round 1

Reviewer 1 Report

line 52-56. Are the different dermatological datasets acquired under same conditions? What about the annotations for different lesions? I assumed that they came from different sources, therefore I am uncertain as to how they can be successfully combined without careful pre-processing.

line 62. needs citation to back up the claim that dermoscopic images achieved better performance.

line 81, "achieve "? 

line 91-100. what is the purpose of putting together this paragraph? 

line 126, larger dataset does not equate model overfitting.

line 142. grayscale images to reduce color effect? need a reference to back it up.

line 145-147. please provide references for these statements.

section 2.3. This is widely available information, which seemed to be redundant to be included here.

line 199. "better performance" does not convey much information. more details can reinforce the conclusions here.

line 210. more details about how the method achieved the end-result is beneficial. 

line 241. I wonder if expert knowledge and clinical experience is always required because some image processing algorithms can generate features useful for classification.

line 254. the color features can not be directly extracted without considering effects of things such illumination and sensor characteristics. more often , color correction is essential before utilizing color as a feature.

line 268. This paragraph seemed to serve no purpose.

line 307. reference please.

line 318. why only mentioned precision and accuracy?

line 319-325. supply more details to support these statements.

line 388-393. I don't find these statements helpful

Table 2. can such information be represented using a plot?

line 426. cite references

line 481. I would argue that researchers do not necessarily need clinical knowledge to implement DL models.

Author Response

Thank you for the constructive comments. The responses are attached in the file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Skin disease is a common  disease that affects the life quality of patients. Sometimes, there are certain difficulties in the diagnosis of skin diseases. This paper introduces the basic steps of skin lesion diagnosis, the existing publicly available skin lesion datasets, and the preprocessing, segmentation and classification methods of dermatological medical image data. On this basis, the current status of machine learning and deep learning in the diagnosis of skin diseases, the results achieved, the problems faced and the related research prospects are reviewed.

It is recommended to add more collation and analysis of related work, and review and introduce it according to different categories. This can help readers further understand the application of machine learning in dermatological diagnosis.

The manuscript is well structured and clearly expressed, and it is recommended to accept and publish.

Author Response

Thank you for the comments. 

As advised, more collation and analysis of related work are included and highlighted using italic font contents.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Editor

The study is well written and organized in a scientific manner.

-          Add the motivation to conduct the study.

-          Instead of las decade add specific duration.

-          The abstract missing a scientific conclusion.

-          Methodology section missing the following: study design, ethics, how they determine the data set, database search, etc.

-          What is the meaning of public dataset?

-          Add study limitation

 

Author Response

Thank you for the comments. The responses are attached in the file.

Author Response File: Author Response.pdf

Reviewer 4 Report

In this manuscript, the author provides information on research on skin cancer including 02 research directions: classification and segmentation. The manuscript shows an overview of the research situation from traditional methods to modern methods such as deep learning.

The essential contents of research in this field are introduced in detail, including dataset, pre-processing, and proposed CNN models.

Some suggestions and concerns to further clarify the issue are presented in the manuscript as follows:

#1. The dimensions of the images in the dataset should be provided.

#2. In the deep learning methods, which solutions have been used to handle the case that the images in the dataset have different sizes?

#3. There are many loss functions used in related research. Which functions will be suitable for the segmentation problem, and which will be suitable for the classification problem?

#4. The author provides metrics used to evaluate the quality of the proposed solutions. However, there has been no summary of the results obtained by the methods based on these metric values. A summary of the results will be essential to help the reader grasp the state-of-the-art for these tasks.

#5. In Table 2 are statistics of articles related to the classification task. How about for the segmentation task?

 

#6. As for figure 2, there's a bit of confusion here. The images after being preprocessed will be the segmentation step, then that will be diagnosed. In my opinion, this step is classification. The segmentation step here may not be appropriate. The author should clarify this figure.

 

Author Response

The response is attached in the file.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Thanks for addressing my concerns. I have no further comments

Author Response

Thank you.

Reviewer 4 Report

Thanks for the replies and the revisions of the authors. However, there is still one point I need to clarify. Table 2 does not yet show which method gives the best task results. I suggest that the author should provide more information.

Author Response

Thank you for the suggestion. The task performance is influenced not only by the applied methods, but also by the datasets. Thus, the direct comparison of the performance metrics may not be meaningful and valid.