The U-Net Family for Epicardial Adipose Tissue Segmentation and Quantification in Low-Dose CT
, Runlei Ma 2,3, Peter M. A. van Ooijen 2, Matthijs Oudkerk 4, Rozemarijn Vliegenthart 2, Raymond N. J. Veldhuis 1,5 and Christoph Brune 1,*Round 1
Reviewer 1 Report
This paper presents a study to compare the U-net variants in the epicardial adipose tissue segmentation and quantification in low-dose CT. There are some concerns about this study. This paper needs to be refined before it can be accepted.
- The motivation of the tissue/lesion segmentation in low-dose CT should be further enhanced. The current introduction tends to separate the epicardial adipose tissue segmentation and the low-dose CT into two independent issues.
- The explanation and discussion of the claimed contributions should be enriched.
- The implementation details of the methods should be further enriched.
- Some methods in the cardiac image analysis should be cited to enrich the literature review, e.g. Causal knowledge fusion for 3d cross-modality cardiac image segmentation; Vessel contour detection in intracoronary images via bilateral cross-domain adaptation; Multi-level semantic adaptation for few-shot segmentation on cardiac image sequences. More studies could be listed by the authors themselves.
- In the data collection, is there any inclusion or exclusion criterion ?
- Please provide the software version of 3D-slicer when annotating the data.
- This study compares four methods. It is better to presents the network structure for all methods. In addition, it is better to remove the background color in Figure 2 to Figure 5.
- In the experiments, some evaluation indices (e.g. DSC) looks not very well (50%~90%), as well as the large variation. Please add some analysis and discussion about these results.
- Some grammatical errors.
Some grammatical errors.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This study evaluated some UNet based architecture for EAT segmentation. While the methods are well described, there are some concerns as follows:
Main comments:
1. Label type a is not pericardial fat but includes myocardium and blood pool. The reason behind using label type a should be discussed in the introduction/methods.
2. Were both types of labels provided to the networks during training? If so, please state in the methods.
3. Ground truth labels are not fully corrected by humans as they were processed automatically. So it is not the best ground truth to compare performance. At least the validation labels have to be manually corrected.
4. Please also compare inference times of the different UNet types used.
5. Are the images, labels and code shared? Please state in methods.
6. Was informed consent obtained under an IRB for this study? Please state.
7. Please provide a breakdown of the type of subjects in this study: age, gender, disease etc.
8. There is no comparison to external datasets although it was criticized by authors in studies by others. Please evaluate on an external dataset.
9. There is no comparison to EAT segmentation methods by others. Please explain in the discussion.
Minor:
Figure 2 legend: Explain Conv, ReLU and BN
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
No further question.
minor refinement
