Quantitative Analysis of Benign and Malignant Tumors in Histopathology: Predicting Prostate Cancer Grading Using SVM

Round  1

Reviewer 1 Report

A classification method for classifying images in 4 Gleason score based prostate cancer classes is presented in this paper. Various interesting techniques are employed but all of them have been tested in MATLAB. It would be more useful if an autonomous low cost tool had been developed incorporating all the methodology in a unified environment.

From the referenced approaches it seems that one of the novelties is the classification in each grade separately while in the referenced approaches a two-step classification is often targeted: a) benign vs malignant and b) low grade vs high grade. The number of referenced approaches that show comparable results is small and should be exteneded (even if some compared approaches are not strictly for PC diagnosis). An additional reference that could also be used is:

[1] Rundo, L.; Militello, C.; Russo, G.; Garufi, A.; Vitabile, S.; Gilardi, M.C.; Mauri, G. Automated Prostate Gland Segmentation Based on an Unsupervised Fuzzy C-Means Clustering Technique Using Multispectral T1w and T2w MR Imaging. Information 2017, 8, 49

which is referenced in

[2] Petrellis, N. A Review of Image Processing Techniques Common in Human and Plant Disease Diagnosis. Symmetry 2018, 10, 270.

If there arent more references on PC to cite similar classification problems are listed in [2] to compare.

A classic K-means clustering algorithm was employed to segment the image into lumen, nucleus and stroma. The segmentation is performed in MATLAB. Oversegmentation is avoided by Watershed algorithm as described in section 3.3. SVM used 14 morphological and texture features for classification (area, perimeter, major axis length, minor axis length, circularity, diameter, nucleus to nucleus distance, nucleus to nucleus minimum distance, eccentricity and compactness). ANOVA was used to select the most important features. The methodology employed seems sound. 

SVM was applied in pairs to discriminate eg benign vs malignant, gleason score 3 vs 4,5 etc.

A more exhaustive comparison between various classification methods based on the same features could have been performed e.g. using the Weka tool.

240 out of 400 images were used for training. 

The description between Fig. 7 and Table 6 is somehow redundant since the information given there is already in the corresponding table and the flow is already described earlier. It would be more useful to describe better Tables 2-4. I cannot understand what is the difference between the two parts of each one of these tables.

Other Corrections:

25: "binary corrections was performed"

124, 126: "In this step," ???

Table 1: remove: "This is a table"

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Authors proposed a quantitative benign and malignant analysis in this paper. The proposed method makes full use of traditional machine learning methods of SVM to grade prostate cancer, and it is demonstrated to be efficient on the representative dataset. However, as we all know, deep learning methods make great breakthroughs in the field of medical image processing and analysis. Authors omit some related articles in the introduction section. Such as: 

(1) Jiao, Zhicheng, et al. "A deep feature based framework for breast masses classification." Neurocomputing 197 (2016): 221-231.

(2) Hu, Yang, et al. "Mammographic Mass Detection Based on Saliency with Deep Features." Proceedings of the International Conference on Internet Multimedia Computing and Service. ACM, 2016.

These papers designed combined deep learning and SVM methods for cancer prediction. Authors may also compare the proposed SVM-based grading method with combined deep learning and SVM models for further experiments.

Overall, I think this paper can be accepted after minor revision of related work review and optional comparison experiments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors proposed a pipeline, including K-means, distance transformation, Watershed and SVM, to score histopathology images. The overall idea is very interesting and the results are promising. However, i have some questions to be addressed by the authors

First, English needs to be extensively edited. I recommend the authors ask a native speaker or a proof reading service to go through the paper. Literature review is not enough. 

Second, the methodology present in the paper is very similar to the following papers, please consider to compare it with them 

Y Ding et al. Novel Methods for Microglia Segmentation, Feature Extraction and Classification IEEE/ACM Transactions on Computational Biology and Bioinformatics 

D Yang et al. .A portable image-based cytometer for rapid malaria detection and quantification  PloS ONE 12 (6), e0179161

Lastly, how the method compare with deep learning based methods ? Please discuss ?

I am happy to review the paper again. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round  2

Reviewer 1 Report

In this revised version the authors have added more references and improved the comparison with the referenced approaches. However my two major comments have not been addressed:

- I asked to compare the classification results they achieve with the specific features used with classification results achieved by classifiers different than SVM. As a convenient way to achieve this I suggested the use of Weka which is a tool that one can be familiarized with in a few hours! As long as the authors can save the features of their samples from MATLAB in a spreadsheet with each sample marked with the class it belongs, Weka tool can read it and repeat the classification in seconds with any popular classifier including neural networks (Multilayer Perceptron). Of course different classification tools can also be used. The authors refused to get involved in this process

- I believe that papers published in a journal with high impact factor such as Applied Sciences, should demonstrate either a new method proven by a strong theoretical background or new methods with a mature level of implementation. The method presented by the authors does not have a strong theoretical background and employs readily availlable tools of MATLAB which is OK as a starting point but not adequate on my opinion for presentation in a high reputation journal. So I insist that the contribution of this paper is limited.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Thanks for the authors revision of the paper as per my comments. 

However, I fail to see that the following paper has been reviewed in Literature Review. In the paper, the authors proposed similar method to score the pathology images. Please cite the reference in the final version. 

D Yang et al. .A portable image-based cytometer for rapid malaria detection and quantification  PloS ONE 12 (6), e0179161

Author Response

Please see the attachment.

Author Response File: Author Response.pdf