A Supervised System Integrating Image Processing and Machine Learning for the Staging of Chronic Hepatic Diseases
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors17/03/2025
Review for paper titled, “A Supervised System Integrating Image Processing and Machine Learning for the Staging of Chronic Hepatic Diseases”
The authors develop a model to accurately discern different stages of fibrosis in the medical field. Given the small size of datasets, they create a simple method incorporating image processing and machine learning. The study has merits as it seeks to address the challenge of small dataset sizes. To improve the manuscript, the authors should address these comments.
The authors should specify which adaptive image processing operations they perform on the ultrasound images. Are these done pre-, in-, or post-processing with CNN training.
The authors should include a “Related Work” section to review past work on medical image analysis and applications of machine learning in detection of medical defects and other defects e.g. cracks in structures. These past works includes;
- Wong, K. C., Syeda-Mahmood, T., & Moradi, M. (2018). Building medical image classifiers with very limited data using segmentation networks. Medical image analysis, 49, 105-116.
All symbols used in the article should be defined, for example in the functional mapping I(x,y), what do symbols “Z” and “N” represent.
Define “a” and “b” in equation (1). Are they histogram threshold values? How are these values and slope parameters determined?
On page 3, is the total 803 images obtained after the image processing operations of adaptive contrasting and rotation?
Add a reference on the quantitative evaluation methods used in the article.
The proposed architecture in Fig 3. should be in the “Materials and Methods” or rather “Methodology” section and not “Results and Discussion” section. Also, the image processing step is missing in the architecture, please include it. Increase text size in Fig 3.
In Tables I and II, is Loss a metric? If so, is it important in determining the improvement of the medical defect detection model?
Page 5, Conclusions section, the claim that this is the first time contrast enhancement is used for data augmentation is not true, as shown by past literature. Exclude such statements and instead specify that you use image contrast enhancement for small fibrosis datasets.
What is the advantage of the shallow supervised model compared to transfer learning based models if their performance is comparable? Is it a reduction in processing cost, risk of importing biases from transferred models?
To enhance the discussion on image pre- and post-processing, I suggest the authors review the following references on integrating image processing and machine learning systems. These could be included in the review of previous research on Page 2 or in an entirely new section on related work.
- Obunguta, F., Hanpasith, S., Sasai, K., & Kaito, K. (2024, November). Segregation Method for Pothole and Manhole Features Segmented in Pavement Smartphone Images Through Deep Learning. In 2024 International Conference on Digital Image Computing: Techniques and Applications (DICTA) (pp. 538-544). IEEE.
The references could point out some ideas on dataset generation through image rotation, the tradeoff between detail and effort required for annotating images vs. accuracy level required at different stages of medical conditions, plus effects of varying image contrast and brightness etc. as image processing operations. Moreover, the literature could also highlight the importance of verifying machine learning detections against expert analysis.
The article should be written following the Electronics journal format.
Finally, the authors should correct spelling and grammatical errors throughout the article e.g. “reuslts” instead of “results”, paragraph 3 under subsection “B. Image Processing” on Page 2.
Comments on the Quality of English Language
The paper will benefit from a detailed proofreading to correct grammatical and spelling errors.
Author Response
We thank you for your review and include all responses to comments in the word file.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper proposes a classification system for liver fibrosis using B-mode ultrasound imaging.
The proposed methodology is not compared with the already established state-of-the-art method, and its description is neither motivated enough nor discussed to encourage medical practitioners to deploy it in practice. In my opinion, the paper is not mature enough to be published in a high-impact journal, Electronics.
Detailed comments:
1) Abstract: "This transformation creates" - which transformation? The abstract is not sufficiently informative about the method used in the paper.
2) The abstract should give a general idea of the accuracy of the proposed method.
3) The keywords have a marginal connection to the abstract - you should extend the abstract to make it more descriptive
4) Section II C. "The CNN model for staging hepatic disease based on the available data is presented here." There is no model presentation in this section.
4) The proposed methodology is not compared to any established method from the literature. Please discuss the results obtained using state-of-the-art methodology.
5) The biggest concern with machine learning, and especially the application of machine learning in medicine, is the reproducibility of results. Please publish both the dataset and source codes of your research in an open online repository (GitHub or any other). Without it, your results are virtually impossible to reproduce.
Author Response
We thank you for your review and include all responses to comments in the word file.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe draft introduces an innovative supervised system for classifying liver fibrosis stages using B-mode ultrasound imaging, integrating adaptive image processing with a shallow convolutional neural network (CNN). This approach aims to address the challenges of limited medical image datasets and the need for non-invasive, cost-effective fibrosis staging methods. By focusing on morphological features, particularly the Glissonian capsule surface, and employing a novel contrast enhancement technique, the system achieves an impressive accuracy of 0.94 for binary classification and 0.73 for multiclass classification. These results suggest that the method could serve as a viable alternative to invasive techniques like liver biopsy and costly imaging modalities such as MRI and CT, with potential applications in clinical practice and education. However, while the study demonstrates promise, several areas require further clarification, elaboration, and refinement to enhance its scientific rigor, reproducibility, and practical relevance.
- The dataset comprises 392 ultrasound images from 215 patients, which is relatively small for training a CNN, even with augmentation. The exact distribution of images across fibrosis stages (F0-F1: 92, F2: 40, F3: 42, F4: 41 patients) is provided, but the number of images per stage is not specified, nor is the balance between training (637), validation (71), and test (95) sets detailed across stages. This lack of transparency hinders assessment of potential class imbalance and biases, which could affect model performance and generalizability.
- The study uses a linear probe focused on the liver’s third segment to capture the Glissonian capsule, but no justification is given for this choice over other segments or probe types (e.g., convex probes commonly used in liver imaging). Without a clear rationale, readers cannot evaluate the appropriateness of this methodological decision or its potential influence on the results.
- The adaptive contrast stretching technique, a cornerstone of the method, uses parameters determined experimentally, but these values are not reported.
- The CNN is described as shallow with two convolutional layers, but critical details such as the number of filters, kernel sizes, pooling dimensions, and dropout rates are omitted.
- For the three-class classification, the test accuracy (0.701) is notably lower than the validation accuracy (0.8873), and the MAE without image processing (1.1933) seems unusually high for a three-class problem (range 0-2). This suggests potential overfitting or differences in data distribution between sets, undermining confidence in the model’s performance. The MAE’s interpretability is also questionable in this context.
- The model is compared to deeper networks (ResNet, VGG16, DenseNet) in Tables III and IV, but the conditions of these comparisons (e.g., same dataset, preprocessing) are unclear. The claim of “comparable performance”lacks specific benchmarks or references.
- The draft highlights potential clinical and educational applications but does not elaborate on how the model could be integrated into workflows, its limitations in real-world settings (e.g., operator variability), or its robustness across diverse patient populations.
- Terms like “surface neural system”(Abstract) and “circular” (Section III.C) appear to be typos or unclear (likely intended as “shallow neural system” and “cirrhotic”), and the writing occasionally lacks precision (e.g., “qualitative data augmentation” is ambiguous).
Author Response
We thank you for your review and include all responses to comments in the word file.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe manuscript can be made more interesting for the readers by incorporating the following comments in the revised version of the manuscript:
1. Explain more about the difficulties of the dataset, like possible biases, and suggest ways to fix these problems, like using other datasets or cross-validation methods.
2. To make it easier to understand and use again and again, give a more in-depth description of the focus-of-attention process, including how it was created and why it works so well for spotting liver fibrosis.
3. Please compare the suggested method to other cutting-edge deep learning models (like EfficientNet and MobileNet) to get a better idea of how well it works.
4. Give more details about how the suggested method could be used in clinical settings, such as any problems that might come up and how it could be added to current diagnostic processes.
5. Add a part or table that describes how the hyperparameters were tuned, including the learning rate, batch size, and optimiser settings, to make the process clearer and easier to repeat.
6. Talk about whether the model can be used with different groups of people or imaging devices, and suggest that more research be done to test the model on a variety of datasets to make sure it works well in all kinds of clinical situations.
Author Response
We thank you for your review and include all responses to comments in the word file.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for Authors7/04/2025
Review 2 for paper titled, “A Supervised System Integrating Image Processing and Machine Learning for the Staging of Chronic Hepatic Diseases”
The authors have addressed all comments.
They should carry out a final proof reading to address other grammatical and spelling errors e.g. “teh” instead of "the” in line 48.
Comments on the Quality of English LanguageThe authors should carry out a final proof reading to address other grammatical and spelling errors e.g. “teh” instead of "the” in line 48.
Author Response
Dear reviewer, thank you for taking the time to review the manuscript. The points that emerged from your comments were very helpful in better highlighting the content of our work.
Reviewer 2 Report
Comments and Suggestions for Authors The authors addressed all my remarks. In my opinion, paper can be accepted.Author Response
Dear reviewer, thank you for taking the time to review the manuscript. The points that emerged from your comments were very helpful in better highlighting the content of our work.
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper can be accepted.
Author Response
Dear reviewer, thank you for taking the time to review the manuscript. The points that emerged from your comments were very helpful in better highlighting the content of our work.