Segmentation of Liver Tumor in CT Scan Using ResU-Net
, Naufal M. Saad 2, Danish M. Khan 2,3, Mahmoud Ahmad Al-Khasawneh 4, Kiran Perveen 5, Abdul Qayyum 6 and Syed Saad Azhar Ali 7,*
Xianchen Xu
Round 1
Reviewer 1 Report
The authors present a straight forward implementation of a UNET based algorithm used to segment livertumors in a public dataset. The authors claim to summarize recent knowledge. However they fail to include recent papers on the subject like. Further, the experiments are poorly conducted and apparently lack controls.
https://doi.org/10.1186/s42490-021-00050-y (using UNet)
https://doi.org/10.1016/j.rico.2021.100087 (using ResUNet)
https://doi.org/10.3389/fbioe.2020.605132 (using RA-UNet)
https://doi.org/10.3390/s20051516
https://doi.org/10.3390/diagnostics12040823
Materials and methods
MetricsWhile DICE is a great measure I would include other metrics like precision, recall and others.
Data: It appears that models are trained and evaluated on the same data. This is bad practice.
Discussion and conclusions: These sections are irrelevant since recent advances in the field are omitted in comparisons.
The language is understandable, but needs to be extensively copy-edited before the paper is publishable in a scientific journal.
Author Response
The authors of the article would like to thank the anonymous reviewer for his great effort and constructive comments which helped me in properly addressing the different issues in better presentation of Segmentation of Liver CT scan Using ResU-Net.
Point 1:
The authors present a straightforward implementation of a UNET-based algorithm used to segment liver tumors in a public dataset. The authors claim to summarize recent knowledge. However, they fail to include recent papers on the subject. Further, the experiments are poorly conducted and apparently lack controls.
https://doi.org/10.1186/s42490-021-00050-y (using U-Net)
https://doi.org/10.1016/j.rico.2021.100087 (using ResU-Net)
https://doi.org/10.3389/fbioe.2020.605132 (using RA-U-Net)
https://doi.org/10.3390/s20051516
https://doi.org/10.3390/diagnostics12040823
Response 1: Thanks for your response, we have summarized the recent knowledge suggested by the reviewer in line number [96-105] highlighted in red color in the literature, and have added recent work to the discussion part of the article highlighted in red color in line number [225-256].
Point 2:
Materials and methods
Metrics While DICE is a great measure I would include other metrics like precision, recall, and others.
Data: It appears that models are trained and evaluated on the same data. This is bad practice.
Response 2: Thanks for your response. We have added the metrics like accuracy, precision, specificity and RVD scores for measuring the segmentation of liver and liver tumors. We have divided the data into training and testing. 80% data was used for training and 20 % testing.
Point 3:
Discussion and conclusions:
These sections are irrelevant since recent advances in the field are omitted in comparisons.
Response 3: Thanks for your response. We have added the recent advances in the field of liver and liver tumor segmentation in the discussion part.
Point 4:
The language is understandable but needs to be extensively copyedited before the paper is publishable in a scientific journal.
Response 4: Thanks for your response. We have made necessary English correction and spelling mistakes.
Reviewer 2 Report
The claim of paper contributions in Line 86-92 should be moved forward into the Introduction section. Moreover, ResU-Net is not a new thing, thus it cannot be considered as a main contribution of the work. This needs some further discussions or reconsiderations.
The related work is not comprehensive. Some 3D medical segmentation algorithms should be included like Quality-Aware Memory Network for Interactive Volumetric Image Segmentation and Contrast-Attentive Thoracic Disease Recognition with Dual-Weighting Graph Reasoning.
In the experiments, there is no comparisons with a vanilla U-Net. This is necessary to better verify the techniques developed in this paper. In addition, the method uses extensive pre-processing and data augmentation techniques, however, it is not clear how they contribute to the performance.
Author Response
RESPONSE TO REVIEWER 2
The authors of the article would like to thank the anonymous reviewer for his great effort and constructive comments which helped me in properly addressing the different issues in better presentation of Segmentation of Liver CT scan Using ResU-Net.
Point 1:
The claim of paper contributions in Line 86-92 should be moved forward into the Introduction section. Moreover, ResU-Net is not a new thing, thus it cannot be considered a main contribution of the work. This needs some further discussions or reconsideration.
Response 1: Thank you for your response. We will add lines 86-92 into the introduction section. In this work, we have trained the ResU-Net framework using selected image pre-processing techniques, methods for increasing the number of images and their variation, and methods to tackle the issue of class imbalance. In particular, intensity normalization is applied to address the issues of inter-patient and inter-scanner variability as well as the issue of dominating background pixel over a liver pixel. Data augmentation is used to increase the variation in data. Class balancing and loss functions were used to cover the problem of class balancing.
Point 2:
The related work is not comprehensive. Some 3D medical segmentation algorithms should be included like Quality-Aware Memory Network for Interactive Volumetric Image Segmentation and Contrast-Attentive Thoracic Disease Recognition with Dual Weighting Graph Reasoning.
Response 2: Thank you for your response. We have added the Quality-Aware Memory Network for Interactive Volumetric Image Segmentation and Contrast-Attentive Thoracic Disease Recognition with Dual Weighting Graph Reasoning into the literature part highlighted in line [73-76].
Point 3:
In the experiments, there are no comparisons with a vanilla U-Net. This is necessary to better verify the techniques developed in this paper. In addition, the method uses extensive pre-processing and data augmentation techniques, however, it is not clear how they contribute to the performance.
Response 3: Thank you for your response. In the experiments, there are no comparisons with a vanilla U-Net. We have compared our results with state -of-art techniques highlighted in result and discussion section.
Reviewer 3 Report
This manuscript looks good, it should be published in applied science.
The author may need to consider using more train epochs until it is totally convergent.
Author Response
RESPONSE TO REVIEWER 3
The authors of the article would like to thank the anonymous reviewer for his great effort and constructive comments which helped me in properly addressing the different issues in better presentation of Segmentation of Liver CT scan Using ResU-Net.
Point 1:
This manuscript looks good, it should be published in applied science.
Response 1: Thanks for your acceptance of the manuscript. We have made necessary English correction and spelling mistakes.
Point 2:
The author may need to consider using more train epochs until it is totally convergent.
Response 2: Thank you for your response. We have trained on trained the network up to 500 epochs, it has improved the results and moreover, we have added more evaluation metrics for liver and liver tumor segmentation highlight in results in and discussion section of manuscript.
Reviewer 4 Report
Authors need to address the following corrections.
1. The gaps in the existing works related to gleason grading need to be consolidated and discussed at the end of related works section.’
2. Research contributions need to be given at the end of the related works section.
3. The novelty of the proposed work needs to be clearly highlighted
4. The related work section need to be rewritten based on the customizations made with respect to U-Net architecture. Please find below few works that have customized the U-Net architecture for effective segmentation. These works can be discussed in context to the proposed architecture.
U-isles: Ischemic stroke lesion segmentation using u-net
Delve into Multiple Sclerosis （MS） lesion exploration: A modified attention U-Net for MS lesion segmentation in Brain MRI
Delineation of ischemic lesion from brain MRI using attention gated fully convolutional network
Weak label based Bayesian U-Net for optic disc segmentation in fundus images
5. Fig -1 , Why data augmentation is mentioned twice in the diagram ? Also, what operation is performed with Hounsfield windowing unit as the dataset is directly downloaded in DICOM format ?
6. Table 2 and 3 can be represented using graph to interpret the convergence.
7. Performance comparison need to be strengthened.
Author Response
RESPONSE TO REVIEWER 4
The authors of the article would like to thank the anonymous reviewer for his great effort and constructive comments which helped me in properly addressing the different issues in better presentation of Segmentation of Liver CT scan Using ResU-Net.
Point 1:
The gaps in the existing works related to Gleason grading need to be consolidated and discussed at the end of the related works section.’
Response 1: Thank you for your response. We have added the existing works related to Gleason grading according to the suggestion of the reviewer in literature part highlighted in red color.
Point 2:
Research contributions need to be given at the end of the related works section.
Response 2: Thank you for your response. We have added the research contribution.
Point 3: The novelty of the proposed work needs to be clearly highlighted.
Response 3: In this work, we have trained U-Net Network along with residual blocks using selected image pre-processing techniques, methods for increasing the number of images and their variation, and methods to tackle the issue of class imbalance. In particular, intensity normalization is applied for addressing the issues of inter-patient and inter-scanner variability as well as the issue of dominating background pixel over a liver pixel. Data augmentation is used to increase the variation in data. Class balancing and loss functions were used to cover the problem of class balancing.
Point 4:
The related work section needs to be rewritten based on the customizations made with respect to U-Net architecture.
Please find below a few works that have customized the UNet architecture for effective segmentation. These works can be discussed in the context of the proposed architecture.
U-isles: Ischemic stroke lesion segmentation using u-net Delve into Multiple Sclerosis (MS) lesion exploration: A modified attention UNet for MS lesion segmentation in Brain
MRI
Delineation of the ischemic lesion from the brain MRI using attention gated fully convolutional network
Weak label-based Bayesian U-Net for optic disc segmentation in fundus images.
Response 4: Thank you for your response. We have re-written the literature according to the reviewer’s suggestion.
Point 5:
Fig -1, Why data augmentation is mentioned twice in the diagram? Also, what operation is performed with the Hounsfield windowing unit as the dataset is directly downloaded in DICOM format?
Response 5: Data augmentation is applied two times. In Hounsfield windowing, we cut zero values to set Hounsfield windowing unit and we can change it in simple numpy array. The image intensity values of all of the images were truncated to within the range of -384 to 384 HU to omit irrelevant information for the liver CT scan dataset.
Point 6:
Tables 2 and 3 can be represented using graphs to interpret the convergence
Response 6: Thank you for your response. We have highlighted the result in two ways: quantitative and qualitative way.
Point 7:
Performance comparison needs to be strengthened.
Response 7: Thank for your response. We have strengthened the performance comparison section according to the reviewer suggestion.
Round 2
Reviewer 2 Report
The revision has addressed my concerns.
