Diagnosis of Breast Cancer with Strongly Supervised Deep Learning Neural Network
Round 1
Reviewer 1 Report
This paper studied the impact of slice-level precise ROI and lesion-level multi-slices diagnosis in using DCNN.
1- The section 4 should be extended to cover pros and cons of the proposed method in comparison to the state-of-the-art methods. Please spent enough time and energy to organize this section.
2- There is no proposal for future works in the manuscript which is odd for me. Referring to the comment #1, the authors will have proposed some future works to handle the limitations of the current work.
3- There is a lack of innovation in deep learning methods, and the key parameters in the model lack necessary explanations.
4- Methods proposed are not presented in the same order. The order should be SBBL, SL, and SBBS. (As stated in figure 3)
5- The term multi-slice is mentioned only in abstract and conclusion only. Authors should be consistent with the terms they use in their manuscript so that readers can follow easily.
Author Response
Dear Editors and Reviewers:
Thank you for your letter and the reviewers’ comments concerning our manuscript entitled “Diagnosis of Breast Cancer with Strongly Supervised Deep Learning Neural Network”. All the comments are valuable and very helpful for us to revise and improve our paper. We have studied all the comments carefully and have made corrections. The main revisions in the paper are as follows:
- We modified the first 3 sentences of the abstract to be more reasonable.
- We replaced breast cancer with the abbreviation BC except for the first one.
- We revised the description of the related works to be clear in paragraphs 4 and 5 of section 1 to be clearer.
- We have removed the other abbreviations of the SVM except the first one. We checked the same problem throughout the paper and corrected the ROI abbreviation too. In addition, we rewrote the “DCE MRI“ with “DCE-MRI” and deleted the “images” in “MRI images”.
- We have modified the "all slices " with "multi-slice" and unified the term.
- We added the parameters of the DCNNs description in section 2.2.2.
- We added an extra section to section 3 to describe the measure criteria of the performance, including TP, TN, FP, FN, ROC, AUC, True positive rate and False positive rate.
- We replaced the “*” with “x” in Equations 1 and 2.
- We almost rewrote section 4, and added more discussion of the ROI, DCNN and lesion level diagnosis criteria.
- We added our future work to address the limitations of our paper in the last paragraph of section 4.
- We revised section 5 to be clearer.
- We have checked the grammar throughout the paper and made corrections to some mistakes.
Response to Reviewer 1 comments
Comments and Suggestions for Authors
This paper studied the impact of slice-level precise ROI and lesion-level multi-slices diagnosis in using DCNN.
Point 1: Section 4 should be extended to cover prosthe and cons of the proposed method in comparison to the state-of-the-art methods. Please spent enough time and energy to organize this section.
Response 1: We almost rewrote section 4. The new section4 included the advantage of the SBBS ROI compared with the SBBL-ROI, SL-ROI, and the ROI of the whole MRI. We presented the advantage of DenseNet that we adopted at last in breast cancer diagnosis. We also presented the advantage of the multi slices weighting lesion level diagnosis compared with the slice-level diagnosis. In the last paragraph, we listed the limitation of our proposed approach and proposed feasible approaches in our future study.
Point 2: There is no proposal for future works in the manuscript which is odd for me. Referring to the comment #1, the authors will have proposed some future works to handle the limitations of the current work.
Response 2: In the last paragraph, we listed the limitation of our proposed approach and proposed feasible approaches in our future study as follows:
“There are several limitations to our study. First, all lesions in our data set were labeled manually by experts, which increased the difficulty of expanding the data set and the possibility of subjective prejudice. Second, the sagittal and coronal lesions are not distinguished in our study. Third, we didn’t use parameters like ADC and Time Intensity Curve (TIC) which could assist the radiologist in the diagnosis of lesions. Finally, there is room for improvement in accuracy and AUC. In our future work, to break the first limitation, we will adopt and improve the DCNN such as Faster RCNN to detect and diagnose the BC automatically and with high accuracy. As the second limitation, we will reorganize the dataset by case, as long as one of the sagittal lesions or coronal lesions is assessed as malignant, it is considered malignant, which may further improve the accuracy of the entire case. For the third limitation, we will combine the ADC and TIC with the features extracted from the ROI as the input of the DCNN classifier to improve the accuracy of the final diagnosis.”
Point 3: There is a lack of innovation in deep learning methods, and the key parameters in the model lack necessary explanations.
Response 3: We evaluated 5 DCNNs, including ResNet50, DenseNet, VGG16, GoogLeNet and AlexNet. These DCNNs are all classic, we kept all the parameters as their default value. And we added the sentence “All the DCNNs’ parameters such as the kernel size, stride, padding, depth, and so on, were kept as their default value.” to section 2.2.2 Slice-level training and classification to declare now. As for the training parameters, we have stated the parameters such as the epoch, learning rate and batch size.
The reason we adopted the classic DCNN was that there were no improvements in the performance when we modified the network’s parameters. We also added the attention branch to the DCNNs, and the results had no improvements too. While the improvement was very big when we replaced the usual SBBL ROI with the SBBS ROI. And when we replaced the slice-level diagnosis with the lesion-level diagnosis, the performance is better. So, we believe that the classic DCNNs are good enough when the ROI is accurate enough, and the ROI, the final diagnosis criteria are also very important for a good breast cancer diagnosis system with the DCNN. In our future work, we will improve the DCNNs to realize detection and diagnosis of the breast cancer at the same time.
Point 4: Methods proposed are not presented in the same order. The order should be SBBL, SL, and SBBS. (As stated in figure 3)
Response 4: The order of the ROI stated in figure 3 is incorrect. We have corrected the order as follows:
Figure 3. Original MRI and three kinds of ROI. (a) and (e) belong to the original. MRI images. (b) and (f) were SBBL-ROI. (c) and (g) were SBBS-ROI. (d) and (h) were SL-ROI. (a), (b), (c), (d) belonged to a malignant lesion, and (e), (f), (g), (h) belonged to a benign lesion.
Point 5: The term multi-slice is mentioned only in the abstract and conclusion only. Authors should be consistent with the terms they use in their manuscript so that readers can follow easily.
Response 5: we have modified the "all slices " with "multi-slice" and unified the term
Author Response File: Author Response.pdf
Reviewer 2 Report
In this work, the authors proposed a deep leaning network to diagnose breast cancer. The work has sufficient novelty. There are some points which needs to be resolved by the authors
1. Some of the abbreviations are repeated. For example- SVM is abbreviated many times.
2. In equation (1) and equation (2), replace * by x(\times command of latex) to make equations more attractive.
3. Justify the word- strongly(used in title) connected in the paper.
4. Check the grammar throughout the paper.
Author Response
Dear Editors and Reviewers:
Thank you for your letter and the reviewers’ comments concerning our manuscript entitled “Diagnosis of Breast Cancer with Strongly Supervised Deep Learning Neural Network”. All the comments are valuable and very helpful for us to revise and improve our paper. We have studied all the comments carefully and have made corrections. The main revisions in the paper are as follows:
- We modified the first 3 sentences of the abstract to be more reasonable.
- We replaced the “breast cancer”with the abbreviation “BC” except for the first one.
- We revised the description of the related works to be clear in paragraphs 4 and 5 of section 1 to be clearer.
- We have removed the other abbreviations of the SVM except the first one. We checked the same problem throughout the paper and corrected the ROI abbreviation too. In addition, we rewrote the “ DCE MRI “ with “DCE-MRI”, and deleted the “images” in “MRI images”.
- We have modified the "all slices " with "multi-slice" and unified the term.
- We added the parameters of the DCNNs description in section 2.2.2.
- We added an extra section to section 3 to describe the measure criteria of the performance, including TP, TN, FP, FN, ROC, AUC, True positive rate and False positive rate.
- We replaced the “*”with “x” in Equations 1 and 2.
- We almost rewrote section 4, and added more discussion of the ROI, DCNN and lesion level diagnosis criteria.
- We added our future work to address the limitations of our paper at the last paragraph of section 4.
- We revised section 5 to be clearer.
- We have checked the grammar throughout the paper and made corrections to some mistakes.
Response to Reviewer 2 comments
Comments and Suggestions for Authors
In this work, the authors proposed a deep learning network to diagnose breast cancer. The work has sufficient novelty. There are some points which needs to be resolved by the authors
Point 1: Some of the abbreviations are repeated. For example- SVM is abbreviated many times.
Response1 : We have removed the other abbreviations of the SVM except the first one. We checked the same problem throughout the paper and corrected the ROI abbreviation too. In addition, we rewrote the “ DCE MRI “ with “DCE-MRI”, and deleted the “images” in “MRI images”.
Point 2: In equation (1) and equation (2), replace * by x(\times command of latex) to make equations more attractive.
Response 2: have replaced the * with the x in equation (1) and (2) as follows:
(1) |
|
(2) |
Point 3: Justify the word- strongly(used in title) connected in the paper.
Response 3: We described the strongly supervised and weakly supervised DCNNs in the modified section 4. There are two kinds of DCNN breast cancer diagnosis systems according to the definition of the ROI. Normally, in strongly supervised DCNN, each sample has a label and most part of the sample is the lesion. In our study, Each SBBS ROI is a sample and has a label 0 or 1, and in ROI, basically, there is the lesion itself. In the weakly supervised DCNN breast cancer diagnosis, the whole MRI is a sample and has a label. But the lesion is only a small part of the sample. We used the word “strongly supervised” for comparison with the weakly supervised studies.
To connect the “strongly” with the content of the paper, we added the strongly supervised before the DCNN in the study.
Point 4: Check the grammar throughout the paper.
Response 3: We have checked the grammar throughout the paper and made corrections to some mistakes.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors addressed my concerns. Thanks for their efforts. The paper is suitable for publication now.