Explainable Liver Segmentation and Volume Assessment Using Parallel Cropping

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a GPU-based voxel-level parallel cropping method to enhance the efficiency and interpretability of liver image segmentation and volume assessment. The approach shows promising application value in the medical image processing domain. However, this paper lies mainly in engineering acceleration, not a fundamentally new algorithmic framework. needs further revision to enhance its innovation and improve its structure for better readability.

1. Introducing the liver medical image and image segmentation tasks in the introduction section for clarification.
2. Adding a diagram and a section to introduce the proposed method for clarification.
3. In the Performance Evaluation section, the dataset used for training and evaluating the model is not well illustrated.
4. No clear separation between training and testing, a lack of benchmarking against state-of-the-art deep learning models weakens the comparative depth.

 

 

Author Response

Response in pdf attached:

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I present my review in points to be addressed and, if possible, improve the article:

1. The introduction lacks general information about AI, and there is a lack of presentation of the technicalities. Therefore, please read and cite the following article in the introduction: DOI10.3390/diagnostics13152582.
2. Although the title and abstract declare the dual purpose of segmentation and volume assessment, most of the paper focuses on acceleration through “cropping,” while issues of the quality of the segmentation itself (especially the medical interpretability of the results) are dealt with in a cursory manner. A more thorough discussion of how volume accuracy was clinically validated and whether volume differences were relevant to treatment planning is needed.
3. It is unclear whether U-Net was re-trained on cropped data or whether a model trained on full images was used. Details of augmentation, loss metrics, number of epochs, cross-validation, etc. are missing. Without this, it is difficult to assess the reliability of comparisons between Chan-Vese and U-Net.
4. Despite the numerous Dice tables and indicators, there is a lack of analysis of cases of wrong segmentation (false positives/negatives), differences between models, systematic errors, e.g. in regions of low contrast or artifacts. It is necessary to illustrate not only typical, but also difficult cases and analyze where the method fails and why.
5. Work excessively focuses on technical details (e.g., “persistent thread blocks,” “grid-stride loops”) without linking them to clinical needs. Does 2× speedup matter in hospital practice? Does improving Dice from 0.938 to 0.960 translate into a change in therapeutic decision? The authors should place the paper more firmly in the context of practical application.
6. The text contains numerous repetitions (e.g., about the advantages of “cropping” in GPU) and duplicates arguments in different sections. Sections are inconsistent in places - for example, “Semiautomatic Cropping” and “Automatic Cropping” could be combined as variants of the same technique. The summaries are also too long - the conclusions stretch over 3 pages and repeat already known facts without new insights.
7. In addition, in the discussion, please refer to a relevant study and compare the results: DOI10.3390/jcm13133634

Author Response

Response in pdf attached:

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript is well written and well presented, However, it needs a lot more work to make it more scientifically sound and more relevant. Here are a few comments to improve its quality and standard; 

1. Despite the inclusion of U-Net, a more comprehensive comparative analysis with recent transformer-based or multi-attention DL models is lacking.  Explicitly define the performance trade-offs that extend beyond the Dice score and quickness.
2. The field of medical imaging is underrepresented in recent research on deep learning (DL) and machine learning (ML).  Kindly integrate and discuss on relevant studies, such as; ( Enhancing Recognition and Categorization of Skin Lesions with Tailored... ,  Optimized Deep Learning for Mammography: Augmentation...., etc ).
3. The dependence on internal CT data and LiTS is acknowledged.  It is recommended that the generalizability of the cropping method across modalities be explained, or that cross-dataset validation be addressed.
4. Although the proposed method is transparent and resource-efficient, the assertion of "explainability" is made without obvious visual or procedural evidence.  Provide examples or an explanation for interpretability that extends beyond the system-level description.
5. Certain figures, such as Figures 2–4, could be improved by better labeling and scale normalization.  Make the boundaries of the cropping process more clear.
6. The post-processing phase that enhances U-Net's Dice score is effective, but it is inadequately explained.  Provide a more concise explanation of the algorithmic steps or the logic behind them.
7. The manuscript contains minor formatting issues and grammatical inconsistencies.  It is advisable to conduct thorough corrections.

Author Response

Response in pdf attached:

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Тhis is an interesting paper related to medical imaging segmentation. More specifically, the authors analyze automatic and semiautomatic parallel cropping techniques used to segment and assess the volume of the liver in medical imaging. They apply automatic and semi-automatic parallel cropping techniques. This leads to improvement of the accuracy of liver segmentation as well as to significant reduction of unnecessary computations.

In the introductory section details concerning liver diseases and particularly liver cancer  together with related medical approaches in this field are given. The advantages and limitations of traditional methods and learning approaches are explained. Further, the relevant background and motivation to crop images and liver segmentation are presented.

The proposed approach is described in detail. The parallel graphics processing unit  approaches for automatic and semiautomatic cropping are explained. Corresponding algorithms are presented and described in detail.

The results of the application of parallel cropping to liver segmentation and volume assessment are presented. Further, performance results and the persistent and grid-stride loop-based parallel cropping comparison for the liver segmentation and volume assessment are analyzed.

A detailed discussion is provided. The properties of the proposed approach are compared with other methodologies. It is shown that the proposed methodology achieves a rapid parallel cropped image for liver segmentation and volume assessment while minimizing intermediate memory transfers between the central processing unit and graphics processing unit.

The presentation of the main results is clear and comprehensive. The results are valuable and worthy of being published considering their possible applications in medical diagnostics proposing a practical solution for real-time clinical applications.  

 

Minor revisions are suggested to improve the quality of the exposition:

p.4, lines 89: An opening bracket should be added before IGS.

p.4, lines 84, 123; p. 16, lines 381, 387 etc.: The meaning of abbreviation “(RoI) region of interest” is given several times, which is not necessary.

p.11-12: The contents of Figures 13-16 are not explained in the text. What is the difference between them and the previous Figures?

p.17, line 457: I suggest the name of the Conference for Ref. 5 to be added.

 

Author Response

Response in pdf attached:

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have complied with most of the reviewer's comments. They corrected the text of the article and thus contributed to its substantive value.

Reviewer 3 Report

Comments and Suggestions for Authors

It is fine now..