Automatic Lower-Limb Length Measurement Network (A3LMNet): A Hybrid Framework for Automated Lower-Limb Length Measurement in Orthopedic Diagnostics

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper, the authors address the task of measuring lower limb lengths from medical images. They propose a hybrid method combining neural network predictions and geometric strategies to identify anatomical landmarks, calculate pixel-based distances, and convert these into physical measurements. 

About paper title  

The current title, “Automatically Lower Limb Length Measure Network,” could be misleading, as it implies the model is entirely based on neural networks. In reality, the methodology combines neural network predictions with geometric strategies for measurement. A revised title that more accurately reflects this hybrid approach would better capture the essence of the work.  

Section 2.1

The collection of images and the creation of ground truth annotations represent a valuable contribution. However, the paper does not specify whether the dataset is publicly available, accessible upon request, or restricted. Providing this information is essential, as it directly impacts the reproducibility and potential impact of the study.  

Section 2.3.1

The authors should carefully review their mathematical notation, particularly in Section 2.3.1:  

- In Equation (1), the definitions of variables $j$ and $n$ are unclear. Please specify what these represent.  

Additionally, in this section:  

- Why is only the x-coordinate of the centroids from the femur and tibia segmentations used? Wouldn’t the method benefit from considering both $x$ and $y$ coordinates? A deeper explanation of this design choice would greatly enhance the clarity of the methodology.  

Section 2.3.2

When extracting the left and right tibia key points in Step B, the method selects a region of 40 pixels above the lowest point. However, this fixed pixel value could be problematic if image sizes vary. The authors might improve the generalizability of their approach by using a proportionate value (e.g., a percentage of the tibia’s size) instead of a fixed pixel measurement.  

Section 2.3.3

The explanation regarding the conversion from pixel distances to physical distances raises important questions. Specifically, the statement:  “To convert the pixel distance into the actual physical distance, the measured pixel value is divided by a ground truth value measured by an expert, resulting in the scaling factor $n$.” raises the following issues:  

- The variable $n$, already used in Section 2.3.1, is reused here as the scaling factor. This overlap in notation may cause confusion and should be avoided.  

- Is the ground truth value an average measurement from multiple patients, or does it come from a single individual?  

- What are the units of this ground truth measurement?  

Clarifying these points and ensuring unique variable names throughout the manuscript will improve readability and reduce ambiguity.  

Section 3

While the proposed model is interesting and the results are promising, the lack of comparison with state-of-the-art methods makes it difficult to assess its true contribution. Including such comparisons would strengthen the paper significantly and highlight the advantages of the proposed approach.  

 

Author Response

Reviewer #1

In this paper, the authors address the task of measuring lower limb lengths from medical images. They propose a hybrid method combining neural network predictions and geometric strategies to identify anatomical landmarks, calculate pixel-based distances, and convert these into physical measurements. 

 

About paper title  

The current title, “Automatically Lower Limb Length Measure Network,” could be misleading, as it implies the model is entirely based on neural networks. In reality, the methodology combines neural network predictions with geometric strategies for measurement. A revised title that more accurately reflects this hybrid approach would better capture the essence of the work. 

Response: Thank you for your valuable feedback. As per your suggestion, the title has been changed from the original to "A3LMNet: A Hybrid Framework for Automated Lower Limb Length Measurement in Orthopedic Diagnostics" to explicitly highlight the hybrid framework, emphasizing the combination of both AI and geometric approaches.

 

Section 2.1

The collection of images and the creation of ground truth annotations represent a valuable contribution. However, the paper does not specify whether the dataset is publicly available, accessible upon request, or restricted. Providing this information is essential, as it directly impacts the reproducibility and potential impact of the study. 

Response: I had already revised the manuscript following a separate request from the editor after submission. At the end of the manuscript, I added the following statement:

Data Availability Statement: Data sharing is not applicable (only appropriate if no new data is generated or the article describes entirely theoretical research).

Please understand that the clinical data cannot be made publicly available due to IRB restrictions.

 

Section 2.3.1

The authors should carefully review their mathematical notation, particularly in Section 2.3.1:

- In Equation (1), the definitions of variables $j$ and $n$ are unclear. Please specify what these represent.

Response: Thank you for your insightful comment. To address this, we have clarified the roles of n and j in the revised manuscript. Specifically, nj​ represents the total number of connected components within the segmented region for a specific structure, and j=1 or 2 specifies whether the connected components belong to the femur (j=1) or tibia (j=2). This explanation has been added to the sentence in Equation (1) for better clarity.

Additionally, in this section:

- Why is only the x-coordinate of the centroids from the femur and tibia segmentations used? Wouldn’t the method benefit from considering both $x$ and $y$ coordinates? A deeper explanation of this design choice would greatly enhance the clarity of the methodology.

Response: The decision to use only the x-coordinates of the centroids for femur and tibia segmentations is primarily based on the anatomical context and the goals of the methodology. In this specific application, the classification of left (L) and right (R) regions for femur and tibia relies on the horizontal spatial distribution of the segmented regions. The x-coordinate provides sufficient information for distinguishing left and right sides, as the division between the left and right anatomical structures is predominantly along the horizontal axis.

In contrast, the y-coordinate represents the vertical positioning, which does not contribute to the left-right classification. Incorporating the y-coordinate in this step would add unnecessary complexity without improving the classification accuracy, as it does not correlate with the anatomical differentiation of left and right regions.

We have added this explanation to the manuscript under Equation (5) for better clarity, and we hope this addresses your concern. If additional details are required, we are happy to elaborate further.

 

Section 2.3.2

When extracting the left and right tibia key points in Step B, the method selects a region of 40 pixels above the lowest point. However, this fixed pixel value could be problematic if image sizes vary. The authors might improve the generalizability of their approach by using a proportionate value (e.g., a percentage of the tibia’s size) instead of a fixed pixel measurement. 

Response: The reason we initially assigned a fixed pixel value was that we had resized the image to 1024 x 256. The 40 pixels roughly corresponded to 10% of the tibia's length. However, your suggestion seems more reasonable, so we modified the criterion from 40 pixels to 10% of the tibia's total pixel height. By establishing a more general standard, the value of the paper has been improved, and we deeply appreciate your insight.

 

Section 2.3.3

The explanation regarding the conversion from pixel distances to physical distances raises important questions. Specifically, the statement: “To convert the pixel distance into the actual physical distance, the measured pixel value is divided by a ground truth value measured by an expert, resulting in the scaling factor $n$.” raises the following issues:  

- The variable $n$, already used in Section 2.3.1, is reused here as the scaling factor. This overlap in notation may cause confusion and should be avoided.

Response: To avoid confusion, as per your suggestion, we have replaced the variable for the scaling factor from n to s.

- Is the ground truth value an average measurement from multiple patients, or does it come from a single individual?

Response: The ground truth value refers to the measurement of a single patient, including the lengths of both their left and right sides.

- What are the units of this ground truth measurement?

Response: As stated in Section 3.3, both the ground truth values and predicted values are measured in millimeters (mm).

Clarifying these points and ensuring unique variable names throughout the manuscript will improve readability and reduce ambiguity.

 

Section 3

While the proposed model is interesting and the results are promising, the lack of comparison with state-of-the-art methods makes it difficult to assess its true contribution. Including such comparisons would strengthen the paper significantly and highlight the advantages of the proposed approach.

Response: To compare our study, we cited the work of Moon et al., titled "A deep learning approach for fully automated measurements of lower extremity alignment in radiographic images," which was published in Scientific Reports in 2023. The comparative analysis has been incorporated into Section 4.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

 

The paper proposes a ML based method for lower limb length measurement in X-ray images. The topic is interesting and worthy of research. However, the work requires a number of clarifications before it is suitable for publication.

 

The discussion on the use of artificial intelligence, in particular machine learning models in supporting medical imaging diagnostics, should be expanded. It is recommended to include the following publications, including explainable AI and AI applications in orthopedics:

10.1002/jum.16524

10.3390/app14104124

10.3390/cancers16101870

10.1007/s43465-024-01189-1

 

Based on the literature review, it is necessary to identify knowledge gaps and justify to what extent the research conducted in the article fills these gaps and what is the contribution of this work to the field.

 

Are the obtained segmentation results presented in Table 1 average values obtained for many patients? If so, please also provide the standard deviation of these results.

 

Please compare the obtained results to the results of studies already published by other authors.

What is the computational complexity of the proposed method?

 

Please provide precise bibliographic data of the references [27]

 

The sentence: "In the first stage, semantic segmentation achieved a mAP 0.5 363 accuracy of xxx." should be corrected.

 

Author Response

Reviewer #2

The paper proposes a ML based method for lower limb length measurement in X-ray images. The topic is interesting and worthy of research. However, the work requires a number of clarifications before it is suitable for publication.

 

The discussion on the use of artificial intelligence, in particular machine learning models in supporting medical imaging diagnostics, should be expanded. It is recommended to include the following publications, including explainable AI and AI applications in orthopedics:

https://doi.org/10.1002/jum.16524

https://doi.org/10.3390/app14104124

https://doi.org/10.3390/cancers16101870

https://doi.org/10.1007/s43465-024-01189-1

Response: We have added the respective papers to the references. Thank you.

 

Based on the literature review, it is necessary to identify knowledge gaps and justify to what extent the research conducted in the article fills these gaps and what is the contribution of this work to the field.

 

Are the obtained segmentation results presented in Table 1 average values obtained for many patients? If so, please also provide the standard deviation of these results.

Response: Yes, that's correct. This represents the average information of 1,000 patients in the test set. However, metrics commonly used in semantic segmentation, such as Mean Accuracy and Mean IOU, typically do not provide deviations. Therefore, we performed K-fold validation to calculate the deviation for Mean IOU separately and included the results in Table 1 of the main text.

 

Please compare the obtained results to the results of studies already published by other authors.

Response: To compare our study, we cited the work of Moon et al., titled "A deep learning approach for fully automated measurements of lower extremity alignment in radiographic images," which was published in Scientific Reports in 2023. The comparative analysis has been incorporated into Section 4.

 

What is the computational complexity of the proposed method?

Response: “A3LMNet, was implemented using python on a computer with a GeForce RTX 3090, GPU 24 GB.” Also, we have developed software based on this. “It takes approximately 6 seconds from loading the file to generating the results for the X-ray image.” We added it to Sections 3 and Section 4.

 

Please provide precise bibliographic data of the references [27]

Response: Thank you for your confirmation. We have revised the citation by following the citation rules of the respective book chapter. Reference: https://doi.org/10.1007/978-3-031-25928-9_8

 

The sentence: "In the first stage, semantic segmentation achieved a mAP 0.5 363 accuracy of xxx." should be corrected.

Response: I made a significant mistake, and I sincerely appreciate your help in identifying it. I have revised the content as follows.

In the first stage, semantic segmentation of the femur and tibia achieved a mean accuracy of 0.958, 0.963 and Mean IOU of 0.963, 0.984 respectively.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

All the comments and suggestions raised during the previous review have been addressed by the authors. I am satisfied with the revisions made and believe the manuscript meets the necessary standards for publication.

Reviewer 2 Report

Comments and Suggestions for Authors

Thak you for proper considering all my comments. The paper is suitable for publication.