U-Net-Based Semi-Automatic Semantic Segmentation Using Adaptive Differential Evolution
Round 1
Reviewer 1 Report
The manuscript tackles the important task of accurately segmenting medical images, specifically in detecting tumors and identifying bone structures using CT and MRI scans. Although techniques for detecting tumors have become widely used, accurately identifying bone structures remains a difficult area of research. The study compares traditional image processing methods with newer deep learning-based approaches for semantic segmentation. It also brings attention to the limitations of deep learning models, which are often needed by the lack of labeled medical image data at the pixel level.
The U-net, a successful CNN-based medical image segmentation model, gives the study's rationale because it requires significant ground truth data, which can be costly to create. The authors propose a novel semi-automatic semantic segmentation method for the U-net to address this.
The contributions are a semi-automatic GT generator and an optimal CT image generator that leverage clustering and combinatorial optimization techniques to produce complete boundary lines between bones and other structures in CT images. This approach should overcome the need for complete GT images for training the U-net, making it more practical and enabling automatic modeling for each patient.
The manuscript is sufficiently well written (I suggest English-language revision) but is well organized. Some suggestions are as follows:
1) The U-Net should be described in detail in Sec. 2 and not briefly in Sec 3.1 (which regards experimental results)
2) In 3.1, more details are needed regarding the training. For example, are the split performed considering the individuals or scan by scan? Please clarify.
3) The results in 3.2.1 are interesting, and the figure looks great. However, I think that numerical results are needed to :
- emphasize that the proposed semi-automatic GT generation works effectively from handwriting lines
- demonstrate it can reduce time-consuming GT generation tasks (for example, indicating the time measured for a GT generation)
4) Sec. 3.2.2 needs to clarify how the evaluation in terms of IoU was performed. First, classic segmentation metrics should be presented in Sec. 2 and then used for the evaluation (e.g., Dice, pixel-wise evaluation, Jaccard, and so forth).
5) The results shown in Figures 4,5,6 are great. However, even in this case, a comparison with the state-of-the-art method is needed (for example, using FCN, SegNet, and DeepLab, as presented in Sec. 1).
6) As an extension of point 5), a comparison (even if not direct) with existing literature methods (cited in Section 2) is needed, both for GT generation and segmentation.
7) The conclusions should deepen the obtained results from a general point of view and discuss the limitations of the current approach, even in relation to possible future works.
Minor:
- abstract: please revise all the acronyms (also throughout the entire manuscript)
- citation 14 contains a "?" at page 2
- Sec 3.1: 1times1 should be 1x1
- English-language revision is needed throughout the paper.
The manuscript requires a revision of the English language to enhance the writing or rectify the overall organization. I gave some suggestions in my review.
Author Response
Thank you for your helpful comments. The detailed responses to your comments are listed in an attached file, where your comments are written in bold with our reply in plain font. The main changes in the paper are shown in red, and the changes by English proofreading are shown in blue.
Author Response File: Author Response.pdf
Reviewer 2 Report
1:The abstract should provide a clear and concise summary of the paper's main contributions, methods, and results. Consider revising it to include details about the methodology and quantitative performance improvements.
2:The introduction should clearly articulate the problem and its importance. Consider expanding on why traditional methods of bone segmentation are inadequate and how the proposed method addresses these challenges. The first line of the introduction should be modified as well"Medical image segmentation has recently been paid much attention to as its performance has improved to the level used in medical diagnostics including retinal disorder [1-5], cancer [6-7],Finger veins recognition[8], etc.”
1. AI-Based Automatic Detection and Classification of Diabetic Retinopathy Using U-Net and Deep Learning; 2. A Transfer Learning and U-Net-based automatic detection of diabetic retinopathy from fundus images; 3. Improved Grey Wolf Optimization-Based Feature Selection and Classification Using CNN for Diabetic Retinopathy Detection; 4. Diabetic Retinopathy detection using Weighted Filters and Classification using CNN; 5. Diabetic retinopathy detection and classification using mixed models for a disease grading database; 6. Lung Nodules Detection using Grey Wolf Optimization by Weighted Filters and Classification using CNN; 7. IGWO-IVNet3: DL-Based Automatic Diagnosis of Lung Nodules Using an Improved Gray Wolf Optimization and InceptionNet-V3 ; 8: Finger-vein recognition using a novel enhancement method with convolutional neural network."
3:Provide an in-depth discussion of the results, highlighting why the proposed method outperforms the U-net model and any limitations or potential areas for further improvement.
4: Ensure the paper cites all relevant prior work, following a consistent citation style, to acknowledge existing research and contextualize the proposed approach within the field.
5: Review the paper for grammatical and typographical errors, and consider adding relevant illustrations to enhance clarity and readability, making the paper more accessible to a broader audience.
Moderate editing of English language required
Author Response
Thank you for your helpful comments. The detailed responses to your comments are listed in an attached file, where your comments are written in bold with our reply in plain font. The main changes in the paper are shown in red, and the changes by English proofreading are shown in blue.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
My concerns have been largely addressed. However, I believe that the manuscript still needs to be improved in the following aspects:
1. In section 3.1, it needs to be better clarified why the datasets used were considered merged. Therefore, it would be appropriate to check them separately as well.
2. In Section 3.2.2 and Table 2, additional metrics for evaluating segmentation should be included (examples: pixel-wise accuracy, Jaccard index, etc.).
3. In section 3.2.2, it would be appropriate to discuss and justify why the results obtained on Femour do not show significant improvement between U-Net and U-Net + jDE (proposal).
4. To validate the proposed method, it would be useful to compare it with other state-of-the-art methods besides DeepLabv3 (e.g., using FCN, SegNet, and DeepLab, as presented in Section 1).
5. Conclusions should elaborate on the results obtained from a general point of view and discuss the limitations of the current approach, also in relation to possible future work.
Minors edit required.
Author Response
Thank you for your helpful comments. The detailed responses to your comments are listed as an attached file, where your comments are written in bold with our reply in plain font. The main changes in the paper are shown in green.
Author Response File: Author Response.pdf
Reviewer 2 Report
The author addresses all the comments.
Author Response
We kindly thank you for your cooperation.
Round 3
Reviewer 1 Report
The changes provided by the authors are satisfactory, and the limitations present have been justified as best as possible. Evaluating the methods also using (at least) the Jaccard index is necessary to provide a complete overview of the proposed method.
English-language revision is needed throughout the paper.
Author Response
Thank you for your comments. We understand that the Jaccard index is the same definition of IoU according to Wiki(https://en.wikipedia.org/wiki/Jaccard_index) and https://www.tasq.ai/glossary/jaccard-index-iou/. If we make a mistake, we would appreciate it if you could point it out to us. Therefore, we didn't change our paper for this round.