Region-Based Automated Localization of Colonoscopy and Wireless Capsule Endoscopy Polyps

Round 1

Reviewer 1 Report

The paper presents the analysis of the usefulness of region-based CNNs for polyp detection in medical images. It is well written and easy to follow, and in principle it can be accepted. Some additional effort should be made to proof-reading and style review (I'm including in the end only some examples). The methodology is correct and sound. Acceptable detection performance is reached with several of the proposed systems.

Authors can consider to further scrutinize the meaning of the internal layers they obtain in their best solutions, trying to identify some insights into their clinical interpretation.

- Page 4, footnotes better as references.

- Avoid contractions (hasn't) 

- line 40: physicians' -> physicians 

- line 56, haven been -> have been

- line 123, Fig. 3 Represents -> Figure 3 represents 

- Eqs (1) to (4), do not embed the text in the equation.

- Please carefully review the bibliography style and fields.

Author Response

to be added

Author Response File: Author Response.docx

Reviewer 2 Report

This manuscript presents a Deep Learning model for detection of polyps. The research is of great interest clinically. However, there are some major concerns which need to be addressed before considering it to be published. Please see below comments:

1. Although, there are some literature reviews in Section 1, still some of the recently published research are missed such as:

Wang, Pu, et al. "Development and validation of a deep-learning algorithm for the detection of polyps during colonoscopy." Nature Biomedical Engineering 2.10 (2018): 741.

Byrne, Michael F., et al. "Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model." Gut 68.1 (2019): 94-100.

Mori, Yuichi, and Shin-ei Kudo. "Detecting colorectal polyps via machine learning." Nature Biomedical Engineering 2.10 (2018): 713.

2. Section 2.4 and 2.5 are not very clear. Authors should explain their method in a better way.

3. In Section 2.6, it is described that how the weights of ResNet is fine tuned after 100 epochs of training by letting the network learn the new dataset for about 1000 epoch. However, there is no clear justification on how and why these limits have been chosen. Please elaborate on these.

4. In Section 3.1, it is claimed that "more feature information from the image data and balloon data closely resembles the shape of the polyps". Again there is no justification why deeper network performs better. Authors should prove that there is no overfitting problem with the deeper network.

5. Please include the computation time and complexity of the model.

6. Authors should include learning curves of the deep model. In addition, confusion matrices for both training and testing datasets must be shown. Please provide the training parameters of the model in detail, which helps others reproduce the results of your research. 

7. There is a lack of comparison with other state-of-the-art papers. Authors must compare their proposed their model with other published models/papers.

some minor comments:

8. Page 8, line 302, it should be misdetection not miss detection. 

9. The explanation of Table 1 in the text does not match with the table. For example, it is mentioned in the text that "CVC-ClinicDB 136 contains 612 standard definition still images of 388 x 284 resolution" but the resolution in the table is 384*288. Please double check these.

10. Page 5, line 155, it should be most of them not most them. Please edit all the typos in the manuscript. 

Author Response

to be added

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

I would like to thank Authors for considering my comments. The manuscript has been improved now and after a language check can be published.