Two-Stage Deep Learning Method for Breast Cancer Detection Using High-Resolution Mammogram Images

Round 1

Reviewer 1 Report

Overview and General comments: The paper proposes an end to end deep learning solution of breast cancer mammography. It is well written and well organized. The data and the methods are appropriate for the research conducted. One major concern is the lack of cross validation and its reuslts and instead only a 50-50 split, which is not recommended due to the possibility of overfitting. The topic is of great interest and the results are interesting.

 

Comments:

The authors state in the second objective that they would like to significantly increase accuracy. In this case, I really expect to see a statistical test to show that there are significant differences in accuracy between models.

Section 2.1 – line 134 and 137. Please describe how the radiologists annotated the bounding box. What were their instructions?

Section 2.1 – Does that mean that you only used BI-RADS > 3 for training and testing?

Section 3.3 – Line 225/226 – It is not clear to me the intention of this sentence. The classification model is a CNN that takes the detected bounding box tumour and uses pixels and the corresponding softmax scores as input or just the pixels? The model is trained in the X detected bounding boxes, but first you increase the bounding box by 1.7?

Line 249 – Please state clearly what is the additional number of conv and pooling layers of the larger network.

Section 3.2 – Please explain exactly which parts of the network were retrained and which ones remained the same, or was everything retrained, but the starting point was your model?

What would be the differences if you would not have used your pre-trained model? I asked this because this is a step that anyone can reproduce later with your paper, without having your private data.

Line 294 – Please comment on the amount of time required to train and do inference of your models.

Figure 6 – Inside the images, the boxes and notes are too small.

In Table 2, Number of FP and FN would be beneficial to have than just recall and Mean IoU. In general, the number of FP and FN can also be referred throughout the text.

Althought you show mAP results, I think confusion matrices are rather preferable to access really how do the BI-RADS are assigned to each class.

 

Small comments:

SSD – line 66 appears without explaining what it stands for

DDSM – line 95 appear without explaining what it stands for

In my opinion, formulas (2) are not necessary, but they certainly do not hurt.

Line 251 – typo: I guess you mean 0.95 or 95%

Author Response

Thank you for your comments. Please refer to the attached file.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript titled “Two-stage deep learning method for breast cancer detection using high-resolution mammogram images” describes a two-stage deep learning method to detect and diagnose breast cancer.  Overall, the manuscript is well written except a few mislabeling. I only have few minor marks below that the authors may want to consider.

1. Line 181: I would suggest the authors rephrase the description, for example, switch “is determined by (1)” to “is determined by the following equation (1)” or any other style you prefer to make it clearer. [See equation (2) description between line 204 and line 205.]
2. Line 231 Figure 3: I would suggest labeling the bounding box with a serial number for easier tracking.
3. Line 268 Figure 4 and line 325 Figure 6: Please label the image as a1, a2, and b1, b2 for easier tracking.
4. Line 353: Table 2 here should be Table 3. There is Table 2 on line 309.

Line 381: Table 3 here should be Table 4. Please go over the manuscript to match the Table citation well.

Author Response

Thank you for your positive review of our work. Please refer to the attached file.

Author Response File: Author Response.docx