Modular Neural Networks for Osteoporosis Detection in Mandibular Cone-Beam Computed Tomography Scans

Round 1

Reviewer 1 Report

I have read the manuscript that is aimed to segmented three-phase method to assess osteoporosis using convolutional networks (CNN) in CBCT scans.

I have following comments in this reference.
1. The literature survey is poor and the relevant algorithms are required to be cited in the paper.
2. The results from the relevant algorithms (segmentation) are required to be compared with the proposed method results.
3.THe Architecture was used in the manuscript is not detailed. Pelasedisclose the details of the deep learning architecture.
4. Paper is confusing that AI-based method is through the software must be explained in detail.
5. There are confusions are there regarding the statistics.
6. For discussions Literature is incomplete. There are many more articles to be discussed. Should have had a proper comparative table on literature and flow chart giving the flow of data analysis.
7. Results are not discussed with any available literature.
8. In general This paper uses a coarse-to-fine framework using an AI system, which aims to procure an easy, errorless and fast method. The two key points in this paper are dataset preparation for the evaluation and test the practicability of the system. The methods in this paper are reasonable, and it is meaningful to use AI system to classify and detection.
9. It is understandable that the classification for osteoporis  is an important research theme. However, there are soem issues to be clarified.

The English shoudl be improved.

Author Response

Dear Reviewer,

 

We would like to extend our sincere appreciation to the reviewer for their thoughtful and constructive comments on our manuscript. Your invaluable feedback has provided us with the opportunity to refine our work, both in terms of content and scholarly rigor. We are pleased to address your specific points as follows:

1. The literature survey is poor and the relevant algorithms are required to be cited in the paper.

In response to the constructive feedback provided by the reviewer, we have taken the opportunity to augment our manuscript's bibliography. We have carefully curated and incorporated an additional six articles that are highly pertinent to the subject matter at hand.

2. The results from the relevant algorithms (segmentation) are required to be compared with the proposed method results.

In response to the observations made by the reviewer, we wish to clarify that while our research does not employ semantic segmentation methods, we have taken steps to broaden the comparative framework within the manuscript. Specifically, we have incorporated an additional study that evaluates multiple algorithms and juxtaposes them against our proposed methodology based on ResNet-101 architecture. This inclusion not only offers a comprehensive performance benchmark but also highlights the novelty of our approach, which is infrequently covered in existing literature. We are deeply appreciative of the reviewer's keen insights, as they have been instrumental in enhancing the scope and rigor of our work.

3. The Architecture was used in the manuscript is not detailed. Please disclose the details of the deep learning architecture.

Within the framework of our research endeavor, we have judiciously employed a customized iteration of the Residual Network (ResNet) architecture. The intricacies of these adaptations, undertaken to better align the model with our specific research objectives, are comprehensively documented in the Materials and Methods section of our manuscript. To augment the scholarly ethos of transparency and reproducibility, we have taken the proactive step of publicly releasing the source code for all algorithms deployed in this study. This code repository is hosted on the GitLab platform, and a hyperlink directing readers to this resource is conveniently embedded within the manuscript. Furthermore, we have enhanced the academic rigor of our document by including a scholarly citation that delves into the utility of ResNet architecture in the nuanced field of medical data processing. We would like to extend our heartfelt gratitude to the reviewers and readers who have lent their expertise and constructive critique, as their input has been invaluable in refining the quality and depth of our work.

4. Paper is confusing that AI-based method is through the software must be explained in detail.

In response to the insightful feedback provided by the reviewer, we wish to elaborate on the architectural nuances of our machine learning model. In both the first and second stages of our study, we utilized the state-of-the-art ResNet-101 architecture as the foundational framework. To specialize this architecture for the qualitative analysis of Computed Tomography (CT) data, we executed targeted modifications on the classifier layer situated at the terminal portion of the architecture. The reconfigured classifier layer comprises a sequence of components: an initial linear layer with 2048 input neurons and 536 output neurons, a Rectified Linear Unit (ReLU) activation function, a dropout component for regularization, followed by a second linear layer. In the first stage, which is focused on classification, this second linear component accepts 536 input neurons and yields 2 output neurons corresponding to predefined classes. A softmax function is subsequently appended to convert raw scores into a probability distribution. For the second stage, which tackles a regression problem, the classifier layer's architecture remains largely identical, save for two pivotal distinctions: the second linear component features 14 output neurons rather than 2, and the softmax function is omitted, in line with the absence of a need for probability distributions in regression tasks. The modification of the classifier layer was imperative to achieve qualitative outcomes that align with the specific requirements of our CT data analysis.

We express our sincere gratitude to the reviewer for their perceptive comments, which have significantly aided in clarifying and enriching the technical exposition of our study.

5. There are confusions are there regarding the statistics.

In the current investigation, two types of machine learning tasks—classification and regression—are scrutinized through graphical representations that delineate the evolution of key performance metrics over training epochs. The overarching objective is to illuminate the differential nature of the learning processes and outcome interpretations for these two distinct tasks.

The initial set of graphical representations focuses on the classification task. In this context, the graphics delineate the trajectory of model accuracy across successive epochs for both the training and validation datasets. The metric of accuracy is particularly salient here due to its intuitive interpretability: it quantifies the proportion of instances that the model correctly classifies into either of the dichotomous categories—commonly denoted as 'yes' or 'no'. The graphical depiction serves as a visual assay, enabling researchers to pinpoint the epoch at which the model's accuracy plateaus, thereby indicating a likely convergence to an optimal state.

Contrastingly, the second set of graphical representations pertains to a regression task. Unlike classification, regression tasks aim to predict a continuous numerical output, making the concept of accuracy less straightforward. Here, the salient metric is the loss function, specifically the Mean Absolute Error (MAE). The graphical representation of the loss function across epochs provides invaluable insights into the model's progressive refinement in approximating the desired output values. A decremental trend in the MAE would signify that the model is successively minimizing the discrepancy between the predicted and actual values. From these graphical depictions, one seeks to identify the epoch at which the MAE reaches its nadir, thereby marking an optimal state for the model.

Lastly, an additional graphical representation plots the ground truth values against those predicted by the regression algorithm. This provides a visual quantification of the model's predictive fidelity. The degree to which these sets of values cluster along the line of identity serves as an empirical measure of how closely the algorithm's outputs approximate the ground truth.

6. For discussions Literature is incomplete. There are many more articles to be discussed. Should have had a proper comparative table on literature and flow chart giving the flow of data analysis.

In response to the valuable input provided by the reviewer, we would like to clarify the methodological constraints imposed by the specific article type under which our work falls. As a 'Technical Note,' our manuscript is designated for the presentation of novel methods or modifications to existing methodologies, rather than comprehensive data analyses typically seen in full-length articles. Accordingly, our analytical approach has been focused primarily on conventional validation techniques and performance metrics, such as accuracy levels, at various stages of our computational system.

To offer an accessible yet robust quantitative summary, we elected to include graphical representations that depict the mean error between the algorithmically derived values and the corresponding ground truth. This choice is aligned with the aim of succinctly yet effectively demonstrating the efficacy of our proposed method or modification, as is customary for the 'Technical Note' category. Thus, while we appreciate the suggestion for a more extensive data analysis, the article type's stipulations guided our decision to adopt a more streamlined analytical presentation.

7. Results are not discussed with any available literature.

We wish to express our gratitude to the reviewer for their insightful recommendation to augment the comparative scope of our manuscript. In alignment with this valuable suggestion, we have enriched our work by incorporating additional literature sources into the bibliography.

 

On behalf of all authors,

Faithfully Yours,

Edgars Edelmers

Robotics and Machine Vision Laboratory

 

Institute of Electronics and Computer Science

Latvia, Riga, Dzērbenes iela 14

4th floor, room 408

Reviewer 2 Report

When writing scientific paper it is more suitable to use passive forms rather than active.

Sone parts of the manuscript shoul be rephrased in order to be more easy to read and understand.

Pls, see the attached file.

Comments for author File: Comments.pdf

When writing scientific paper it is more suitable to use passive forms rather than active.

Sone parts of the manuscript shoul be rephrased in order to be more easy to read and understand.

Pls, see the attached file.

 

Author Response

Dear Reviewer,

I am writing on behalf of my co-authors to express our gratitude for your invaluable feedback on our manuscript. Your constructive comments have significantly improved the quality of our work.

We have also utilized MDPI's language editing service to address the English language issues you pointed out, further enhancing the manuscript’s readability.

We have submitted the revised manuscript and are hopeful it will meet your expectations. Thank you for your time and expertise, which have been instrumental in elevating our work.

On behalf of all authors,
Faithfully Yours,
Edgars Edelmers
Robotics and Machine Vision Laboratory

Institute of Electronics and Computer Science
Latvia, Riga, Dzērbenes iela 14
4th floor, room 408

Round 2

Reviewer 1 Report

The relevant changes were made, thus can be accepeted as an original article.

The relevant changes were made, thus can be accepeted as an original article.