Predicting Cell Cleavage Timings from Time-Lapse Videos of Human Embryos

Round 1

Reviewer 1 Report

This paper deals with interest topic. biomedical engineering stuff is well applied ti deep learning techniques. 

Please explain how the authors get datasets for the experiment and the work in this paper.

Regarding detection, have you tried other methods rather than YOLO and DETR?

What is the most important part among the sections 5.1, 5.2 and 5.3?

Please modify abstract so that the goal of this work is more clearly presented. 

Author Response

• Please explain how the authors get datasets for the experiment and the work in this paper?

Author's Reply: Thank you for the comment. We have now added further explanation on how we got the datasets for the experiments. The process by which embryologists monitor and record embryo development in the time-lapse videos is explained in Section 3.3. To increase the clarity, we have now added a table summarizing the creation and how the different datasets were used for different experiments.

 

• Regarding detection, have you tried other methods rather than YOLO and DETR?

Author's Reply: No, we did not try other methods, but we tried some other techniques within the YOLO and DETR methods. Since YOLO v5 performed the best in the embryo cell detection, and after we figured out that the non-maximum suppression (NMS) caused the YOLO to miss cells with overlapping cell boundaries, we decided to train and test two new variants of YOLO. The first one was YOLO v5 with soft NMS and the second one was YOLO v7. Unfortunately, the results did not improve. We have now revised the manuscript and added this information to Section 6.1.

 

What is the most important part among the sections 5.1, 5.2 and 5.3?

Author's Reply: In the section 5, we have described our methodology’s working using three inter-related subsections, 5.1, 5.2 and 5.3. Section 5.1 focuses on training YOLO v5 and DETR for cell detection, 5.2 on predicting the start of cell cleavage stages and 5.3 on reading out the hours after insemination associated to start of a cleavage stage. All three subsection are equally important and are all required for the methodology to be able to predict cell cleavage timings. However, Section 5.2 builds on the methods in 5.1 and 5.3 and represents the final part of the methodology. In this perspective 5.2 can be seen at the most important part. We have now rephrased the introduction to section 5 (Cell cleavage detection) to clarify this.

 

• Please modify abstract so that the goal of this work is more clearly presented.

Author's Reply:  In the revised manuscript we have rephrased the abstract to focus on why cell cleavage timings is relevant and how our proposed methodology intends to automate the time computation. We also focused on how our approach differs from traditional embryo cell detection techniques. Finally, we presented our main results and reported total execution time taken by the methodology in predicting the start time of cell cleavage stages for a video.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this article, the author describes a methodology for identifying the start and duration of human cell cleavage stages in time-lapse videos, which will be helpful for evaluation of the quality and viability of embryos chosen for uterine transfer during IVF process. The manuscript is in well shape and here are some concerns need to be clarified.

 

1. line 16, for the “in vitro fertilization”, “in vitro” should be shown as italics.

 

2. For the image from optical microscope (such as figure 1, figure 4 et al), the scale bar should be added.

 

3. Please check the two question marks in line 139.

 

4. Is this deep learning model available for public? More instructions should be addressed to share this model for the public readers.

 

5. The author should give a deeper discussion about the limitation for this research properly.

This manuscript is organized in good shape. I suggest that the English language should be modified to by a native speaker. 

Author Response

This manuscript is organized in good shape. I suggest that the English language should be modified to by a native speaker.

Author's Reply: Thank you for the suggestion. We have now carefully improved the quality of the English language in the paper.

 

• line 16, for the “in vitro fertilization”, “in vitro” should be shown as italics.

Author's Reply: Corrected.

 

• For the image from optical microscope (such as figure 1, figure 4 et all), the scale bar should be added.

Author's Reply: The time-lapse videos do not contain scale bar information, but we have clarified the scale of the images in the caption of the figures. In the revised manuscript, we have added the scale bars to all relevant figures (Figures 1, 4, 8, 10, 11 and 12).

 

• Please check the two question marks in line 139.

Author's Reply: Thank you for pointing it out. We have corrected it.

 

• Is this deep learning model available for public? More instructions should be addressed to share this model for the public readers.

Author's Reply: Yes, we have looked into this, and in according to ethical guidelines approved by Regional Committee for Medical and Health Research Ethics - Southeast Norway, we can share the AI models, but not the data, used in the paper. We used git repository to share the model. In Section 10, a link to a GitHub repository where the code and other necessary details for accessing the methodology is provided.

 

• The author should give a deeper discussion about the limitation for this research properly.

Author's Reply: Thank you for this comment. We have now extended the discussion section to dive deeper into the methodology’s performance, limitations, and future research directions. We have added a more detailed discussion for the cases where the methodology performed poorer. This could be due to limitations on the volume and variety of data used for evaluation, the methodology’s dependence on imaging modality and points that the functionalities and methodology’s parameters that should be validated before the methodology can used in clinical practices.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Overall the points that I raised have been addressed properly. I agree to publish this paper at the present form.