Review Reports

Round 1

Reviewer 1 Report

Nice and well described work.

It should be mentioned in the Title that the study is based on a VCT.

I believe the abstract should be made easier to read, if possible. I refer to the words/definitions 'Perlin noise', U-net, Dirichlet calibration, Dice? (associated with probability outcomes), Jaccard.

More efforts should be made here to make the abstract more reader friendly. For instance, to describe the essence rather than writing the words above, which understates that the reader of the paper is already familiar with them.

Is is possible to write 'realistic' rather than 'complex' in the following sentence:

Complex mammary parenchyma was 17 simulated using Perlin noise ; 

Overall, I think the limitations could be developed further, and the recent papers by Bosmans and Marshall (Performance evaluations of DBT systems) on VCTs and their limitations should be described further .

 

Goood quality!

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Authors:

This study aims at developing and validating U-net-based segmentation model to identify abnormalities using simulated NGT imaging data. The manuscript is well written overall and the proposed simulation and segmentation algorithms worked well in predicting major lesions, but I have some comments regarding the article as described below that I hope you find helpful.

 

Major comments:

1.        For potential readers who are not so familiar with breast tissue imaging, you could have described more about why you needed to fully simulate imaging data. It is not obvious to everyone that DBT includes multiple projections and that your NGT prototype proposes to initialize the process with a single projection. It would also be advisable to better explain that the NGT is not in clinical use and therefore not able to produce clinical images for assessment – hence the reason for the simulation. Please make this more clear in your introduction and provide the readers with a clear purpose and objectives at the end of your introduction.

 

2.        How did you validate the quality of simulated imaging data? Providing some metrics to see if the quality of simulated data appropriately described real NGT would be helpful for readers. Outcome measures used as well your assessment methodology (AUC) could be reported in a separate “Statistics” section at the end of your methods section.

 

3.        One of the goals in this paper is to identify suspicious regions found in NGT systems. However, you did simulate only four types of abnormalities. From this viewpoint, you could have justified why you did not utilize other approaches that can detect a wide range of suspicious regions (e.g., anomaly detection techniques)?

 

4.        You dealt with class-imbalanced data (line 156-158). You could have clarified more why you could not increase the number of small-sized classes by the simulation process. Obviously, simulation allows for generating infinite number of synthetic images by tuning random seed or other hyper-parameters.

 

5.        Dice (or Jaccard index) in Table 2 performed very well. However this may have resulted from the high similarity between training and testing datasets as both of them were generated from the identical simulation procedures. Did you do something to differentiate between training and testing datasets to show generalizability of the proposed algorithm?

 

6.        How did you determine the appropriate sample size?

 

Minor comments:

1.      What is a definition of cluster in line 188? Did you apply any post-processing (e.g., blob analysis) to identify clusters?

2.      Is it a reasonable assumption that lesion dimensions are all 7x7x7 mm^3? (Line 100) Is this assumption required to construct a model?

3.      In line 152 your split % does not add up to 100%

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Nice job addressing the reviewers' comments! I have nothing further.